Vehicular Communication Network Enabled CAV Data Offloading: A Review

The connected and autonomous vehicles (CAV) applications and services-based traffic make an extra burden on the already congested cellular networks. Offloading is envisioned as a promising solution to tackle cellular networks’ traffic explosion problem. Notably, vehicular traffic offloading leveraging different vehicular communication network (VCN) modes is one of the potential techniques to address the data traffic problem in cellular networks. This paper surveys the state-of-the-art literature for vehicular data offloading under a communication perspective, i.e., vehicle to vehicle (V2V), vehicle to roadside infrastructure (V2I), and vehicle to everything (V2X). First, we pinpoint the significant classification of vehicular data/traffic offloading techniques, considering whether data is to download or upload. Next, for better intuition of each data offloading’s category, we sub-classify the existing schemes based on their objectives. Then, the existing literature on vehicular data/traffic is elaborated, compared, and analyzed based on approaches, objectives, merits, demerits, etc. Finally, we highlight the open research challenges in this field and predict future research trends.

communicating nodes and support for the various vertical industries, i.e., Industry 4.0, internet of things (IoT), and automotive network. Specifically, the support of the vehicular traffic puts an extra burden on cellular communication networks [1]. Furthermore, providing different vehicular internet services while on the move, anywhere and anytime, to support applications is undoubtedly a massive burden on existing overloaded mobile networks. These vehicular applications and services include location sharing applications, sensor data sharing applications, online games on the move, multimedia downloads, tourist/advertisement information, social media applications (i.e., Facebook, Twitter, WhatsApp, and WeChat), virtual reality (VR), and augmented reality (AR), mixed reality (MR), etc. It is pertinent to highlight here that the applications mentioned above are not only data-hungry but also computation-intensive and delay-sensitive [2], [3]. Therefore, to partially tackle the overloading problem in the cellular network, vehicular data offloading using different modes of the vehicular communication network (VCN) appears to be a promising solution.
In communication networks, offloading means transferring the data load from the overloaded cellular network to the WiFi network, small cells, and mobile devices/vehicles to reduce the cellular spectrum's use and accommodate more users with better quality of service (QoS) [4]. Here data can be related to vehicle safety, environment, entertainment, infotainment, or smart roadside/mobility. For instance, in some cases, vehicles fetch data for a better journey experience, and at other times vehicles generate data using sensors for infrastructure and vehicles in proximity [5]. VCN modes are employed in vehicular data to offload data primarily transmitted via the cellular network. The rudimentary idea of offloading involving vehicles is centered on VCN modes, i.e., vehicle to vehicle (V2V) [6], vehicle to roadside infrastructure (V2I) [7], and vehicle to everything (V2X) [8]. This technique helps to reduce the cellular network's traffic load partially. Smart vehicles are fitted with sensors, processing, and different wireless communication capabilities to enable data offloading in vehicular environments [9].
Specifically, the applications and services for connected and autonomous vehicle (CAV) based traffic make an extra burden on the cellular communication networks. It will drive the cellular traffic growth to a new level [10], [11]. For example, an autonomous vehicle will deliver more heterogeneous infotainment contents, i.e., 4K/8K movies, TV, music, and online games. However, such content needs to be fetched from remote cloud data centers coupled with varying channels  RELATED SURVEYS due to mobility, which can cause substantial transmission delays and, consequently, a drop in QoS. In such conditions, there are constantly designing and deploying challenges of the communication networks and computing ecosystem for efficient transport and processing of such a massive volume of data. In the data/traffic offloading scenario, an increase in the downloading and uploading of vehicular nodes and the existing mobile users' traffic load can cause cellular networks to be further congested in that region. In this situation, the vehicular traffic/data offloading leveraging VCN modes seems to be a promising solution for the mobile data explosion problem in cellular networks [12]. For offloading traffic (i.e., downlink and uplink data) in the vehicular environments can use other resources-rich (i.e., communication, computation, and storage) devices, vehicles, or roadside infrastructure. Therefore, the vehicular traffic offloading approach can minimize the cellular network's traffic and the cost of cellular connectivity apart from vehicular applications' QoS improvement.
There are some surveys tackling data offloading in vehicular environments [13], [14], [15], [16], [17] listed in Table I. Still, no one has deeply covered data offloading in vehicular environments considering the vehicular communication network's perspective. The authors in [13] presented opportunistic offloading techniques by considering both the traffic and computation offloading in 2018. However, this survey mainly focused on mobile data offloading with less consideration for vehicular data offloading. Another survey appeared in the year 2018 based on VANETs [14], in which the authors merely divided the data offloading approaches according to V2V, V2I, and V2X modes without further classification. However, data offloading schemes are reviewed in a limited fashion without lessons learned. In contrast, we covered all the advances in this area, presented a comprehensive survey, and deeply categorized and compared the schemes at the objective level to get better intuition. In 2019 [15], the authors focused on VEC-based computation offloading, content caching, and data delivery approaches. However, this survey mainly targets task offloading without focusing on the data. In [16], the authors discussed the current state of VANETs development, including critical technologies, resource management, safety applications, communications and data transmission protocols, and theoretical and environmental constructs. However, the primary emphasis was on addressing concerns related to VANETs, i.e., network security, reliability, and intelligence, with no attention paid to data offloading. In [17], the authors focused on using IP-based vehicular networking in smart road scenarios for V2V, V2I, and V2X communications. They overview the main technologies and discuss challenges and security considerations for secure and safe communication. The authors also offered directions for current and future IP-based vehicular networking and applications for humandriving, partially autonomous, and autonomous vehicles on smart roads. However, their focus was mainly on networking, specifically IP-based vehicular networks, but they did not consider the aspect of data offloading.
All the available surveys provide valuable contributions. However, none of the above provides comprehensive and deep coverage of vehicular data offloading based on vehicular communication modes (i.e., V2V, V2I, and V2X). Our proposed survey systematically categorized the state-of-theart techniques according to the V2V, V2I, and V2X communication modes as shown in Fig. 1. Then, the solutions under each category are divided further into downlink and uplink data offloading scenarios. Moreover, for deep-dwelling in the underlying technical area, all the schemes are then subcategorized based on their objectives.
Our survey's salient contributions are as follows: • We build upon the existing literature on CAV data offloading and present a comprehensive survey on recent advances. First, we introduce DSRC and cellular technologies' vehicular communication modes and their comparison. Next, we report the first up-to-date, complete research effort covering the data offloading protocols and Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. techniques in vehicular environments in the purview of VCN modes, i.e., V2V, V2I, and V2X.
• We elaborate, compare, and analyze the works in each mode (i.e., V2V, V2I, and V2X) according to the downlink and uplink data offloading sub-categories for delaytolerant and delay-sensitive data, respectively. Moreover, each category is supported with summary tables to enable the readers to catch the logical relations of the offloading mechanisms from various aspects such as objective, offloading scenario and data type, simulator and mobility trace type, merits, and demerits.
• Our survey brings fresh and novel summaries to motivate more in-depth studies in the area.
• Finally, we highlight the open research challenges and the future research directions in this promising area. The organization of the remaining survey is as follows. Section II provides preliminaries of vehicular communication technologies and their communication modes. Section III pertains to the overview of the vehicular data offloading domain and different categories, including V2V-based data/traffic offloading, V2I-based data/traffic offloading, and V2X-based data/traffic offloading, are discussed. In Section IV, we provide insights and some open issues-based research directions. Finally, the paper is concluded in Section V.

II. VEHICULAR COMMUNICATION TECHNOLOGIES
AND MODES The VCN comprises processing, communication, and several intelligent devices to provide communication and computing. Communication modules are installed at both the vehicle's on-board unit (OBU) and roadside infrastructure [18]. Generally, the roadside infrastructure here means RSUs, radio frequency identity (RFID) readers, traffic lights, road signs, lane markers, and parking meters, which are sometimes equipped with MEC [19]. Research in vehicular communication aims to provide an upgraded, reliable, and robust road safety, congestion control system, and other infotainment services for CAV. For the better implementation of these services, the vehicular communication technology provides V2V, V2I, and V2X communication modes [6]. Depending on different scenarios, these modes can be implemented using DSRC or cellular network technologies. However, the recent FCC report and order [20] specifies that 30 MHz of the 5.9 GHz ITS band is allotted for transportation and vehicle applications, 20 MHz of which is dedicated to cellular-V2X technology, and the remaining 10 MHz to either cellular-V2X or DSRC. While this current ruling may provide some technological clarity, the significant drop in ITS allotted spectrum from 75 MHz to 30 MHz will have a negative impact on the provision of high-end eV2X services.

A. DSRC Technology-Based Different Vehicular Communication Modes
The traditional transport system has evolved into an intelligent transport system (ITS) and now to cooperative-ITS [5]. The federal communication commission (FCC) announced the licensed ITS band and allocated 75 MHz bandwidth from 5.850 GHz to 5.925 GHz in 1999. DSRC uses OFDMbased 10 MHz channels in 5.9 GHz and common control  [21]. DSRC allows vehicles and roadside infrastructure to communicate information in vehicular networks. DSRC and WAVE use IEEE 802.11p for V2V and V2I communications [22]. IEEE 802.11p delivers 3-27 Mbps at 300 metres [23]. By enabling real-time traffic information and the synchronisation of traffic signals, DSRC, WAVE, and ASTM can assist in reducing congestion by enabling communication between vehicles and with traffic management systems to enhance traffic flow. following are the DSRC modes, which include DSRC-V2V, V2I, and V2X and given as follows: 1) DSRC-V2V Communication Mode: V2V wireless communication enables vehicles exchange speed, location, and other data within 300 metres [24]. This information sharing alerts drivers to potential hazards such as road traffic congestion, accidents ahead, road repairs, and so on [25]. V2V is a wireless mesh network where each vehicle operates independently under three standards: 1) IEEE 1609 (WAVE) describes network operations and architecture. 2) IEEE 802.11p specifies DSRC PHY and MAC technologies. 3) SAE-J2635/J2945 define message packets [26]. WAVE uses the DSRC protocol suite for low-latency applications, analogous to TCP/IP over Wi-Fi.
2) DSRC-V2I Communication Mode: V2I mode connects vehicles to roadside infrastructure and uses the DSRC/WAVE protocol for bidirectional data transmission like V2V [19]. V2I lets vehicles interact with stationary RSUs for quick network acquisition, low-latency transmission, and frequent handovers [27]. The vehicle's OBU's communication, processing, and storage unit connects with onboard sensors and ports for V2I connectivity with RSUs and nearby vehicles. RSUs can sense and control traffic, cache data, facilitate vehicle handover, manage bandwidth, and prioritise communications [28].
3) DSRC-V2X Communication Mode: V2X includes V2V, V2I, vehicle-to-pedestrian (V2P), and billboards and traffic signals. V2X mode lets automobiles share information with surrounding vehicles, infrastructure nodes, and people [29]. 802.11p was developed nearly two decades ago, while 802.11n/ac/ax have enhanced MAC and physical layer technologies. A V2X study group formed IEEE Task Group 802.11bd (TGbd) in January 2019. Its criteria include a support vehicle communication range of 2000 m, a relative speed of 500 km/hr, and twice the throughput of 802.11p. 802.11p will be redesigned into 802.11bd, affecting system-level performance and launching a new era of V2X applications [30]. All the related acronyms are listed in Table II.

B. Cellular Network-Based Different VCN Modes
The 5G Automotive Association (5GAA) was founded by auto industry giants to support 5G cellular networks. 5G can enable large device connectivity, provide rapid connection, sustain connectivity at high mobility, ensure low latency for vital and safety applications, provide extreme dependability for autonomous vehicles (AV), and provide security [31]. As illustrated in Fig. 2, cellular technology provides three modes of communication, which are cellular-V2V, cellular-V2I, and cellular-V2X.
1) Cellular-V2V Communication Mode: V2V safety applications have strict latency, reliability, and coverage requirements. LTE-based V2V is cheaper and offers better coverage than DSRC-based. It can also be implemented on current infrastructure with little adjustment. Being centrally controlled, LTE can use traffic monitoring and control measures in addition to its wider coverage. LTE-supported device-to-device (D2D) communication as part of ProSe services in 3GPP Release 12 laid the groundwork for V2V communication via sidelink at the physical layer known as PC5 [32]. Release 14 supported two V2V connection modes as shown in Fig. 2.: PC5 and LTE-Uu (V2N). PC5 interface-based V2V communications enable two modes: cellular aided mode (Mode 3) and totally autonomous ad-hoc mode (Mode 4) [33]. PC5 Mode 3 uses a cellular network to schedule and manage V2V traffic via control signals over the Uu interface from a cellular BS (eNB or gNB) [34].
2) Cellular-V2I Communication Mode: Cellular-V2I connects vehicles to roadside infrastructure using cellular communications, as depicted in Fig. 2.3GPP Release 14 allows PC5 access to UE-type or eNB-collocated RSUs. LTE at sub-6 GHz spectrum can support V2I operational requirements [35]. LTE guarantees round-trip latency of less than 10 ms and 100 ms for user-plane and control-plane, respectively, for delay-sensitive vehicle applications [36].
LTE theoretically supports 350Kmph mobility with 50Mbps uplink and 100Mbps downlink data rates and 20 MHz bandwidth. LTE's data rate is 10 Mbps at 140 kmph, according to several testbeds. LTE can accommodate 1,200 vehicles per cell in a crowded environment by sending one cooperative awareness message (CAM) per second with an uplink delay of 55 ms [37]. Release 14 supported two V2V connection modes: PC5 and LTE-Uu (V2N). PC5 interface-based V2V communications enable two modes: cellular aided mode (Mode 3) and totally autonomous ad-hoc mode (Mode 4), as shown in Fig. 2. PC5 Mode 3 uses a cellular network to schedule and manage V2V traffic via control signals over the Uu interface from a cellular BS (eNB or gNB). 3GPP added and improved use cases for future vehicular communications in Release 15.
3) Cellular-V2X Communication Mode: V2X encompasses all entities that communicate data with vehicles in any form [38]. Some other specific communication types included in LTE-V2X are the V2P, vehicle to device (V2D), vehicle to grid (V2G), and vehicle to network (V2N) illustrated in Fig. 2. LTE V2X supports Uu and PC5 dual-mode connectivity. Cellular-V2X is a combination of a cellular network and network-less technology platforms that support long-and short-range direct communications, respectively [5]. The direct link also supports V2V, V2I, and V2P modes in the ITS 5.9 GHz band, while V2N uses the cellular licenced spectrum. PC5 also supports dynamic speed, direction, location, etc. sharing. Vehicle platooning, autonomous driving, sensor sharing, and remote driving are also advancing V2X.
In 3GPP Release 16, the new radio (NR) will allow LTE V2X essential safety applications and improved V2X (eV2X) applications like platooning, remote driving, and sensor sharing [39]. NR V2X would facilitate enhanced eV2X by providing more than 1 Gbps data rate for long-range communications and GHz bandwidth and throughput for shortrange communications using mmWave. NR V2X also supports 500 kmph communication. NR V2X offers two more unicast and group cast operations, broadcast modes, and the physical sidelink feedback channel (PSFCH). NR SL enhancement and UE relaying are Release 17 work items. Table III compares RAT technologies from different perspectives for better  understanding. III. VEHICULAR DATA/TRAFFIC OFFLOADING This section gives an overview of the vehicular data/traffic offloading domain and their types based on offloading direction, i.e., downlink/uplink. As mentioned above, vehicular communication uses V2V, V2I, and V2X modes; therefore, we categorise vehicular data offloading as V2V, V2I, and V2X, as shown in Fig. 1. Based on the categories indicated before, we further investigate downlink and uplink data offloading in the following sections and then sub-categorize each category according to their schemes' objectives. Each data/traffic offloading domain category is supplemented with a summary table to comprehend the basic ideas and gain a deeper understanding of the applicable scheme.
Overview of Vehicular Data/Traffic Offloading: The ubiquitous cellular connectivity and economical smart mobile devices dramatically transform our lives and enable us to use different Internet services while on the move, anywhere, and anytime [40]. However, there is undoubtedly a colossal burden on cellular networks due to vehicular traffic generated by innovative and immersive applications, which are data-hungry, computation-intensive, and delay-sensitive sensitive [14].
Data offloading in a vehicular environment is a technique used to alleviate network congestion and improve performance by transferring data traffic to different networks such as DSRC, C-V2X, and NR-V2X. These technologies help to optimize network performance, increase capacity, meet safety-critical applications, and reduce costs, thereby reducing the need for data offloading. Additionally, the industry is also exploring the use of edge computing to reduce latency and improve reliability for safety-critical applications.
Vehicular offloading through different VCN modes delivers a low-cost solution. For example, if a vehicle makes a data demand available with any proximity vehicle, we can opt for V2V offloading to forward the content to the requested vehicle in this situation. Hence, V2V-based offloading is to carry out data/traffic by determining a multi-hop inter-vehicle routing path without BSs and RSUs. While V2I-based offloading works on a mechanism of switching data/traffic between the vehicles and RSUs by avoiding BSs if it is under the coverage of RSUs. Otherwise, it is also forwarded to BS. Therefore, carrying out data/traffic through RSUs relaying mechanism is considered a low-cost offloading solution compared to directly accessing BS. Furthermore, unlike V2V and V2I-based traffic offloading, V2X-based offloading leverages V2V and V2I communication modes and other communicable objects to carry out data sharing from another vehicle or RSU/signage or signage to the other pedestrian/devices and even to the cellular network.
Data/traffic offloading involves transferring data from a primary network to a secondary network in order to reduce the load on the primary network. Applications for this include delay-tolerant services such as bulk data transfer, content caching, and vehicle data transfer, as well as delaysensitive services like safety-data delivery and vehicular crowdsensing. These offloading options are being researched for their potential to improve on-road safety, infotainment, and entertainment [14].
The use of RSUs and BSs that are managed and integrated together can facilitate trouble-free communication between CAVs and the underlying infrastructure. To guarantee that the vehicles are always linked to the best available network, a handover mechanism and multi-modal communication are used. Service providers can track data to make necessary adjustments, and they can coordinate to oversee the handoff between RSUs and cell sites. In addition, technologies and standards like 5G and C-V2X are being developed to better integrate and manage RSUs and eNBs/gNBs, allowing for more efficient and reliable communication between vehicles and infrastructure.
Vehicular Downlink/Uplink Data: Vehicular communication involves the transmission of data between vehicles, road-side units (RSU), and passengers. This data can be related to vehicle safety, the environment, entertainment, infotainment, or smart roadside/mobility. It can be classified as either downlink (from the BS to vehicles) or uplink (from vehicles to the BS) data. Vehicles may fetch data for a better journey experience or generate data for infrastructure and other vehicles in proximity [5]. To reduce the load on the infrastructure and vehicles, different data offloading mechanisms can be used. The following section describes the available data offloading mechanisms to reduce the load from infrastructure and vehicles. These mechanisms also lead to load balancing, application efficiency, proper utilization of resources, and optimal data transfers in all three modes (i.e., V2V, V2I, and V2X) of vehicular communications under cellular and DSRC technologies. Vehicular communication systems have several data offloading use-cases that are designed to improve the efficiency of data transfer and reduce the load on the cellular network. These use-cases include bulk data offloading, cellular data offloading, maximization of content availability, and cost and energy optimization. These use-cases can be achieved through various scenarios, such as the use of content carrying vehicles, vehicular cloudlets/clusters, and vehicles relaying data.
A. V2V-Based Data/Traffic Offloading V2V data/traffic offloading is a mechanism in which only V2V communication mode is utilized to offload the data/traffic. Fig. 3 illustrates V2V data offloading scenarios in vehicular communication systems. It shows a conceptual realization of V2V data offloading, where the dotted purple, solid purple, and doted black link lines represent V2V, V2I, and infrastructure to cloud (I2C) communications, respectively. The Scenarios in Fig. 3-A 1 and 3-A 2 represent downlink and uplink data offloading. Where vehicles V a , V 4 , V 5 , and vehicles V b , V 1 , V 2 , and V 3 are making vehicular clusters in support of content downloading and uploading, respectively. V a in downloading cluster and V b in uploading cluster are the cluster heads, and are important vehicular nodes. These vehicles are the only nodes that are connected to the cellular network, while all the other proximity vehicular nodes take services from them. Whenever vehicles V 4, and V 5 need to access any content, it is being provided by the V a . Instead of requesting to the BS in far proximity, these vehicles request the content to the V a in near proximity. In this situation, V a may have the requested content, or it will grab that through the BS. On the other hand, V b in the uploading cluster provides similar services to vehicles V 1 , and V 2 , but for the content uploading. The downloading and uploading clusters in Fig. 3-A 1 and 3-A 2 are specific examples of how data offloading can be used to improve content availability, cellular network offloading, and cost, delay, and energy optimization. Additionally, both downloading and uploading clusters also help to reduce the number of nodes connected to the cellular network. This can help to increase the overall efficiency and reduce the load on the cellular network.
Another scenario of data offloading is depicted in Fig. 3-B, where a content relaying scenario is shown while addressing cellular network offloading. Vehicle V e downloads the content and forwards it to the vehicle V f , which then continues to vehicle V h through vehicle V g . Then vehicle V h uploads the content to the destination BS. Another content relaying path is also shown, from vehicle V e to the RSU connected to a data center through vehicle V i .
As discussed above, vehicular traffic comprises both downlink and uplink data, and therefore, the research contributions are segregated and categorized considering the downlink and uplink data offloading. Furthermore, to better understand data/traffic offloading, we sub-categorized the schemes according to their offloading objectives.
1) Downlink Data Offloading: In downlink data offloading, we want to reduce downloading nodes to relax the infrastructure and facilitate the vehicle itself. In this sub-section, we discuss the mechanisms of downlink data offloading. a) Bulk data transport: Modern vehicles are equipped with a handsome amount of storage resources, and these storage resources can be utilized to carry data from one location to another over road infrastructure. Focusing on bulk data transportation over road infrastructure, Gorcitz et al. [41] proposed a data offloading and capacity enhancement scheme for vehicular communication systems that utilizes V2V communications to opportunistically transfer bulk data. Vehicles are equipped with memory modules that can be wirelessly loaded, exchanged, or swapped with other vehicles at designated offloading spots. These spots are selected based on geographic location or the vehicle's destination, and the service provider tracks the vehicles, offloading spots, and data to be transferred. While the authors argue that the solution can achieve 200 times more data transfer. However, there are several non-realistic assumptions in their experimental analysis such as unhindered traffic scenarios and lack of consideration for data loss and routing issues.
Different from [41], Glacet et al. [42] studied the temporal vehicular connectivity in a store-carry-and-forward scenario using a graph-based solution under dynamic traffic conditions. They constructed an evolving graph using traffic traces and calculated the transitive closure using a temporal path. They found that the reachability ratio between store-carry-and-forward and connected forwarding was 91% and 60%, respectively. They also compared the technologypenetration ratio (vehicles with communication terminals vs. vehicles without) and found that the V2V store-carry-andforward solution provided better reachability and penetration even for sparse networks. However, the study used a strict-indirect model of store-carry-and-forward and could be evaluated more precisely with other models for a better comparison.
Many vehicles travel on the same path between two or more connected locations. Observing these vehicles, the private vehicle count is more than the total number of vehicles of other traveling agencies. Using this concept, Baron et al. [43] proposed a scheme for offloading large amounts of data through the use of electric vehicles (EVs) as data carriers, charging stations as offloading spots, and manufacturers as service providers. They created an overlay network using a mapping algorithm, and considered offloading stations as nodes with links characterized by capacity, delay, and data loss attributes. Data can be exchanged wirelessly between vehicles and offloading spots while charging. They found that this solution can transfer petabytes of data and is more efficient than a fixed dedicated data transfer vehicle solution. Moreover, they extended their work in [44] by incorporating an SDNbased logical central controller for more efficient route and vehicle selection. However, some assumptions such as constant throughput on all links and unconstrained offloading spots were made, which may not be realistic in practice.
On the other hand, Naseer et al. [45] investigated vehicular offloading in the environmental degradation aspect along with the network congestion and energy consumption perspectives. They proposed an energy-efficient framework for delivering delay-tolerant bulk data using public transport. The authors proposed Internet and vehicle-based energy consumption models for data transportation time, cost, and energy consumption calculations. The proposed vehicular data transportation system analysis showed a significant reduction in delay in contrast to internet-based data transportation on a similar amount of data transfer. They extended their work by considering the carbon emission factor in data transportation as the energy is consumed from resources like electricity or fuel.
In another work, Dias et al. [46] worked on the data offloading through public transport busses and taxis. A gridbased statistical information grid (STING) clustering algorithm is used for clustering analysis with low complexity. STING divides the city into GPS coordinate-based small square regions for the capacity analysis. First, clustering is performed considering the maximum vehicle percentage in each cluster. Next, vehicle activities are examined, the number of travels between clusters, stay time in each cluster, and travel time between clusters of vehicles are computed for probability density function (PDF) and cumulative distribution function (CDF). Finally, a data transfer comparison between 802.11n, 802.11p, and LTE has distinct size clusters. Analyzing the vehicle's activities and calculating the PDF and CDF for them with different cluster sizes shows that V2V communication can offload a considerable amount of data in a single day.
b) Optimization of cost and energy: Targeting the optimization of cost and latencies, Yao et al. [47] put forward another offloading and data dissemination scheme to offload extensive deadline-constrained bulk data from the cellular network to the V2V network. The scheme consists of a deterministic, contact-based, deadline-constrained deadline trimming contact graph (DTCG) for content delivery and linear programming (LP) based online offloading algorithm to construct the DTCG. The DTCG is constructed using predictions generated from offloading-coded (random linear network coding (RLNC) is used as coding scheme) contact events trace and meeting time. If the deadline of a particular event is over or entry of DTCG is invalid due to traffic dynamics, the LP formulation is used to rebuild a new predicted DTCG. A lazy-control mechanism is used to tackle the traffic-movement dynamicity, and a cumulative hurting indicator (CHI) is maintained for a defined threshold level. If CHI exceeds that, the DTCG is updated, and newly updated codes are assigned to the relevant vehicles. The CHI also has a constrained update policy to reduce computational load on the back-end servers. The performance of improved efficiency and accuracy is validated by comparing baseline schemes. However, considering the real-world mobility traces, this scheme's behavior is still a question to answer.
Jinglin et al. [48] explored data offloading for delayed contents and suggested a utility function-based greedy algorithm while focusing on vehicle movement estimate methods. A server-based system anticipates vehicle movement and contacts based on position, speed, direction, and road infrastructure. To maintain data distribution, secondary vehicle nodes are chosen. Any vehicle becomes a secondary node for requesting vehicles after receiving data. A node remains secondary until the content's maximum delay tolerance begins with request initiation. First, the request is sent to the central server, which reroutes it to the nearest one-hop secondary vehicle if found. Otherwise, the requester gets content directly from the cellular network and becomes a secondary node for vehicles without that resource. The proposed method has been evaluated against opportunistic traffic offloading using movement predictions and random selection algorithms. V2V communication offloads 70% of cellular traffic during performance validation. However, the approach assumed unrealistic data packet size, communication ranges, bandwidth, data interest, and synthetic traces.
Offloading data from the cellular network and storing videos in a vehicle's play-out buffer is a popular research topic. Vehicles can cache videos to save energy and service cost. In [49], the authors examined popular vehicle-stored video content. An analytical approach computes the appropriate number of content replicas, cache refreshing time, and temporal cache updating policy based on content popularity. The proposed protocol pushes content from cellular network infrastructure nodes to vehicular cloudlet, which store and cache content. User queries are sent to infrastructure nodes, then forwarded to vehicular cloudlet. The content is cached and sent to the requester if found. It waits for the next vehicle, which may have the content till TTL expiration. Cache misses occur when TTL expires, and infrastructure nodes fulfil user requests for content delivery.
In another work, Pescosolido et al. [50] analytical vehicular offloading scheme considered vehicle speed, mobility, content popularity, and propagation model. Algorithms calculate energy usage and offloading efficiency on a central content delivery management system (CDMS). Content requesters-the vehicle or other devices-are always within an eNB's zone of interest. CDMS initially accepts requests and reroutes them to the nearest vehicle having the requested content. If the content request timeout expires, eNB routes it to the content server. The automobile that requests content becomes a content supplier for nearby vehicles. After determining the minimum energy and delivery cost, CDMS selects surrounding vehicles and eNB. The authors tested the scheme with different vehicle speeds and communication ranges and observed the optimal values for maximum transmission versus allowable energy consumption without affecting system efficiency.
One solution to solve cellular network overload is data offloading using VANETs (Vehicular Ad-hoc Networks). In this approach, vehicles act as mobile hotspots and can offload data from the cellular network to improve network performance. However, there are challenges in motivating vehicle owners to participate in this process, as they may be hesitant to use their bandwidth, storage, and energy resources to help other users. Thus, Yang et al. [51] formulated the interaction between a service provider and a requester as a Stackelberg game, considering the vehicle's mobility, BS bandwidths, V2V contact durations, the service provider and requester, and the content's popularity. Based on these conditions, the authors created a selection algorithm to let service requesters choose the lowest-priced provider. The technique reduces task downloading time and maximises service provider and requester utilities.
Guntuka et al. [52] devised a neural network-based offloading strategy to find efficient V2V offloading paths using queues. The technique determines V2V routes from the source vehicle to the next RSU. These paths and their costs are identified via a knowledge-defined network. One queue stores offloading paths and their prices, while the other stores RSUcovered vehicles and their stay times. These queues speed up re-computation after offloading interruptions. Offloading uses the cheapest V2V path. Average bandwidth, hop counts, latency, network utilisation, throughput, response time, and energy consumption determine the cost. The results indicate that the predicted V2V path made offloading easier.
The data in a communication network is delay-sensitive whenever real-time and or atomic data is delivered in any direction. The following subsection specifies the understanding of delay-sensitive data offloading schemes concerning the offloading objectives.
c) Maximization of content availability and throughput: Exploiting the vehicular storage resources, Malandrino et al. [53] explored the use of parked vehicles as relays to extend the capabilities of Road Side Units (RSUs) for content downloading and delivery. The authors proposed a time-expanded graph model to examine the dynamic capabilities and performance of parked vehicles as relays between RSUs and moving vehicles. They also proposed two optimization models for content freshness and radio resource utilization. The results of their study showed an increased ratio of content freshness and maximized channel idle time. However, the study assumed constant user-interest-time for all vehicles and perfect mobility prediction, which are both important factors for performance measurements.
In another work, Mezghani et al. [54] worked on a seed selection system that maximises commuters' interests in vehicular networks. SIEVE adds an interest-awareness aspect to seed selection to save resources. SIEVE lists time-constrained possible impact vehicles from future V2V contact forecasts, computes content utility based on interest, and seeds higher utility vehicles. These seed nodes then use the cellular interface to retrieve content from the content server and distribute it to other vehicles until the content's lifespan expires. This work assesses performance through the use of content utility. Seed selection is crucial to cover low-vehicular-density areas with various interests. In high-density settings, high content TTL values were seen. SIEVE outperformed randomness and future centrality at low TTLs. In another work, Zhu et al. [55] proposed a contact-aware resource allocation scheme for offloading data from the mobile network using V2V communication. The scheme uses a centralized controller to make decisions on data transfer and rerouting based on the predicted contact duration between vehicles. The vehicles act as relay nodes for other proximity vehicles and the data hosting and relaying process is reward-based. The scheme takes into consideration average contact rates and data interests, and showed improved performance in their experiments.
Different from the work in [55], Vigneri et al. [56] proposed a chunk-based video caching scheme in vehicular environment. First, the average per-content replication vector is calculated, then the chunk popularity factor and chunks inherent delays. Optimal per-content allocation is calculated based on fractional storage and the same popularity. Based on the previous calculations and assumptions, a queuing model for the play-out buffer is then proposed. Its assembly depends on available vehicles in the proximity of requesting users. A heuristic algorithm is also proposed for per-chunk allocation policy, taking per-content allocation as input. Chunks of these contents are reshuffled considering their inherent delays and their popularity afterward from the stored contents on vehicles. These chunks are kept for offloading purposes. They evaluated their scheme and analyzed a considerable gain of 10% to 25% against the content-based offloading scheme. However, their model has some weaknesses, e.g., the assumption of average content popularity, fixed and equal content size, assumption of identical vehicles, and fixed TTL for all contents.
Huang et al. [57] proposed a lifetime-based network state routing (LT-NSR) scheme that utilizes software-defined networking (SDN) to offload data traffic from the cellular network to the vehicular network through V2V communication. The LT-NSR, running on a central SDN controller, estimates a V2V path by collecting vehicles' GPS position, speed, and directional headway through the cellular network. The longest connection time is selected if more than one path is found. The LT-NSR offloaded 58% of cellular data compared to 45% of data offloaded by GD-NSR in moderate vehicle density scenarios. However, both schemes performed similarly in low and high vehicle density scenarios. Therefore, the authors extended their work in [58] and proposed a lifetime-based path recovery (LT-PR) algorithm for the recovery from the broken V2V paths and named it SDNi-MEC. However, the results for moderate vehicle density are pretty satisfying, but SDNi-MEC shows low performance in the case of high and low vehicle density.
Following the base architecture defined in the previous works as in [57] and [58], Huang et al. [59] proposed a k-hop based offloading scheme. According to the proposed scheme, the MEC server finds out the paths of offloading using the reported data. The paths having a higher lifetime while covering maximum road distance are selected to offload among all the other possible paths. An offloading agent is also selected based on maximum stay time under one RSU. The offloading agent vehicle fetches the data using the cellular network. Afterward, this data is transferred via V2V communication until the path or connection timeouts are ended. The proposed algorithm showed up to 4.07 times more data offloading while having a data loss of just 0.7%. Since this work considered only one MEC, this work can be extended further for multiple MECs. d) Offloading of cellular N/W: Motivated by the gap in building an offloading framework to offload the cellular network using V2V mode and provide a way of the samedata-content dissemination to multiple vehicles in a particular territory having different delivery deadlines. Yuan et al. [60] introduced a data offloading scheme considering Spatiotemporal constraints, named space and time-constrained data offloading scheme (STCDO) for location-dependent services on the IoV. First, a periodically updatable dynamic contact graph is constructed at the server to depict inter-vehicle transmission opportunities. Then, a greedy algorithm is proposed to select superior offloading vehicular nodes (i.e., seeds) for offloading and data dissemination. The algorithm keeps track of content dissemination evolution and limits the re-transmissions according to the reachability of the remaining subscribed vehicles. The proposed scheme can marginal data traffic offloading under Spatio-temporal constraints.
On the contrary, Bao et al. [61] proposed an incentive cache offloading scheme that utilizes the onboard storage of vehicles to offload data from the cellular network. The scheme uses an auction-based approach, where data requesting vehicles announce auctions for data and cache carrier vehicles submit bids based on the inter-requesting vehicle distance and the channel condition between the requester and the bidder vehicle. The requester chooses the bid with the highest reception probability and if the data can be received within the period, it will receive the data and issue the reward to the bid winner vehicle. The results of this scheme demonstrate the feasibility of offloading and the potential reward gains, but further verification is needed in different buffer spaces, request expiry time, and data selection for caching.
2) Uplink Data Offloading: V2V uplink data offloading refers to the process of transmitting data from vehicles to other vehicles or to the infrastructure, such as a RSU or a BS, using V2V communication. This process is used to alleviate the burden on the cellular network and improve the overall performance of the communication system. One approach to uplink data offloading is through the use of cooperative communication techniques. This involves vehicles forming a cooperative group and sharing their resources, such as data, to improve the overall communication performance. For example, the use of relaying and forwarding schemes can improve the coverage and capacity of the communication system. Another approach is the use of vehicular cloud computing, where vehicles act as mobile cloud nodes and provide computing and storage resources to other vehicles. This can be used to offload data and processing tasks to the vehicles, reducing the load on the cellular network and improving the overall communication performance. There are also various other solutions that have been proposed for uplink data offloading, such as the use of multi-hop communication, data compression, and coding techniques, and the use of MEC. The choice of solution for uplink data offloading will depend on the specific requirements of the communication system, such as the density of vehicles, the mobility patterns, and the available resources. A combination of these solutions may also be used to achieve the best performance. The following subsections provide a comprehensive study of V2V uplink data offloading.
a) Optimization of cost and resources: Targeting the service costs, Vare et al. [62] enlighten another aspect of vehicular data transfer, the closed-circuit television (CCTV) data offloading to reduce service costs. Vehicles produce a huge amount of CCTV data stored at their network video recording (NVR) unit. This proposed solution aims to reduce service costs by offloading CCTV data from vehicles to passenger's mobile devices. The data is split into small chunks and assigned to different mobile devices with incentives offered to passengers for participating. The passenger's device then offloads the data to a centralized ground system when an appropriate network is found. The proposed scheme is primarily a software-based solution and does not require additional equipment costs. The system's transfer capacity and offloading efficiency are also increased as it runs parallel to the previous system. Unlike [62], Moura et al. [63] introduced the smallest d-hop dominating set based greedy algorithm to offload the vehicular sensory data. The proposed algorithm works on the multi-hop closeness centrality to find out a subset of uploading vehicles. This algorithm runs on each BS to find out the subset of uploading vehicles from the set of its associated vehicles. Each vehicle from this subset collects vehicle sensory data from proximity vehicles, aggregates data, and uploads it to the core VANET servers through BS. The results show that if the network is a little fragmented, then the scheme works fine and helps to reduce a considerable cost. However, the scheme does not remain effective if the network is less fragmented or too much fragmented.
In another work, He et al. [64] exploited a software-defined vehicular network (SDVN) that aims to improve the costeffectiveness of data routing in vehicular networks. The SDVN uses a scheduling scheme that takes into account the different interfaces available on the on-board units (OBUs) such as DSRC, cellular, and Wi-Fi, to select the most cost-effective and high-bandwidth option. The SDVN also uses a predictable vehicular trajectory and a greedy algorithm to estimate the cost and bandwidth of different interfaces and find the most efficient routing solution. The proposed scheme was shown to be more cost-effective than existing baseline algorithms to improve the overall performance of the system.
In contrast, to [64], Manzo et al. [65] worked on floating content (FC) in the vehicular environment and introduced a machine learning (ML) technique for setting up the anchor zones (AZ), which is a geographically constrained area concerning FC. An AZ configuration algorithm is designed to assume the central resource management in a heterogeneous smart city environment by employing a convolutional neural network (CNN) approach. The algorithm sets AZ configuration by considering the communication network status, time partitioning, time-slot-based road subsets, and road correlations. A link feature vector is extracted first, and considering the required performance and resource usage, the machine learning algorithm learns from this vector and provides a global AZ configuration as a final step. The proposed scheme is validated and provides a more accurate and optimal global AZ configuration.
b) Maximization of content availability and connectivity: Maximizing the content availability, Manzo et al. [66] put forward an analytical approach to evaluate the performance of FC in a vehicular context. The authors considered vehicles moving in the district mobility model and mapped on a modified random waypoint (RWP) model. Then, the model is set considering a vehicle having FC (called seed) is within an AZ. Whenever a seed generates FC data within an AZ, other vehicles in AZ copy that data when they come into contact with the seed. Every vehicle entering the AZ gets a copy of FC, and on leaving, it is discarded. In this way, FC data persists in AZ probabilistically for a longer time. Results showed that taking a small AZ radius is more suitable for real-time and short-lifetime applications/services. On the other hand, the scope of AZ can be extended to the requirements of delay or relatively long-term services, and it will be more beneficial in low-traffic areas.
On the contrary, Huang et al. [67] proposed an offloading time-based k-hop limited V2V offloading scheme, considering a vehicle has a V2V path in the rear or ahead direction linking it to one of the system's RSU/AP. While designating vehicles as offloading or relaying, the V2V multi-hop link is established with the help of MEC. The V2V k-hop path is shrinking, expanding, and recovery algorithms are also proposed to keep an offloading session as elongated as possible. The shrinking and expansion of an offloading session depend on the data lifetime threshold. The results with low, medium, and high vehicle densities show that 1.42 to 4.07 times more data can be transferred than the traditional schemes, i.e., under-only-RSU.
Focusing on content availability, D'orey et al. [68] put forward an optimal way of floating car data (FCD) information flow where vehicles are divided into long (cellular) and short (DSRC) range communicating vehicular nodes. A neighboring vehicular table is constructed using single-hop CAMs and shared with the geo-server for virtual infrastructure and cluster heads (CH) selection processes. When the CHs are selected, they can operate as the data collector, relay, and dissemination bodies operating on the instructions from the geo-server. The results showed a marginal message penetration with reduced overhead and improved network utilization by offloading network traffic to short-ranged V2V networks.
Kolios et al. [69] investigated the interplay of vehicle speeddensity and density congestion concerning data communication and its applications to offloading. They formulated a scheme to identify on-road regions of low-speed highcontention (LSHC) and high-speed low-contention (HSLC). In the HSLC region, inter-vehicle space increases and contention decreases on RSU coverage, while in the LSHC region, vehicle density increases and inter-vehicle space decreases, leading to increased contention under one RSU coverage. The study found that V2V relaying was more feasible in LSHC regions and that the relaying scheme performed better in LSHC regions as compared to HSLC regions, where V2V relaying is difficult to achieve due to larger inter-vehicle space.
Sensor data influences ITS decisions. In this article [70], vehicle cluster relays (VCR) are used to increase the number of associated devices without the high capital and operating costs of specialised relays. Due to regularly varying channel interference and restricted communication coverage, a movement-and fairness-aware heuristic (MFAH) strategy is provided to ensure sensor device transmission rates and fairness. MFAH sequentially conducts two unique channel allocation algorithms, exclusive and compatible, to swiftly allocate channels and optimise channel utilisation, enhancing device associations while ensuring transmission rates. A device association technique that considers the number of existing affiliations between devices and their proximity to the target VCR makes uploading sensor data equitable for all connected devices.
c) Offloading of cellular N/W: Kolios et al. [71] proposed a graph-based RAN-load-balancing offloading mechanism to address the challenge of transferring large amounts of beacon messaging (BM) data produced by vehicles via cellular RAN to vehicular back-end servers. The mechanism involves every vehicular node finding a vehicular node α, which is always connected to any infrastructural nodes in range, to transfer its data. The α node collects data from neighboring vehicular nodes and decides the amount of data to be transferred to the connected infrastructural node at an instantaneous time and how much data should be transferred in the next possible exchange. The proposed mechanism aims to introduce α nodes to have a load-aware estimation to transmit the collected data, avoid congestion and provide load balancing in a large network scenario.
In contrast to the RAN-load-balancing offloading mechanism proposed by Kolios et al. [71], Stanica et al. [72] presented a distributed offloading scheme using V2V communication to reduce the load of vehicle-data-uploading from cellular network infrastructure. The scheme involves a small group of vehicles, formulated as a minimum dominating set (MDS) problem, collecting and distributing the entire network's FCD data. The authors applied greedy MDS, lexicographically first maximum independent set (LFMIS), and polynomial-time approximation scheme (PTAS) algorithms to calculate the CDF of non-offloading cars for different MDS heuristics. The analysis was done using the degree-based (DB) mechanism, then a degree-based with confirmation (DB-C) method was introduced based on cellular but with increased gain and very close coverage compared to greedy MDS. Then, a reservation-based (RB) scheme was developed to overcome the deficiencies of DB-C. This solution results in a constant MDS approximation in an ideal scenario where each node reserves a different slot, but in reality, reservation collisions occur. The RB mechanism confirms its performance near to the LFMIS scheme under the impact of collision scenarios. It provides coverage of all nodes utilizing 20% fewer nodes than the DB-C solution and approaching greedy MDS in a dense vehicular environment.
LTE-based vehicular crowdsensing (VCS) puts an extra burden on the LTE resources and degrades the QoS. Considering this viewpoint, Nunes et al. proposed a EUCLIDEAN scheme, which is a vehicular geo-clustering based solution that aims to reduce the burden on LTE resources and improve QoS in VCS. It uses two algorithms, one for formulating attraction areas under the ROI of crowdsensing and the other for formulating vehicular clusters from these attraction areas. Each vehicle records neighboring vehicle locations, radius, and centroids of attraction areas mapped to ROIs. These vehicles then use this information to decide on a CH through which the VCS data is sent. The CH then forwards the data to the VCS server. The performance of EUCLIDEAN has been shown to result in a 92.98 percent reduction in LTE resource demands and fulfill VCS resource requirements through V2V communication [73] Summary: The data/traffic offloading using V2V communication mode is either done through DSRC or PC5 links under IEEE 802.11p and LTE technologies. The motivation to introduce V2V-based offloading is to complement the cellular network load and utilize the resources optimally. Therefore, we have sub-categorized the V2V data/traffic offloading into downlink, and uplink data offloading and summarised it in Table IV.
The state-of-the-art research provides several offloading solutions. Apart from the different uplink and downlink data offloading schemes, another difference lies in considering individual vehicles or clusters of vehicles as data carriers. Moreover, some schemes use vehicle mobility-based contact graphs data/traffic routing or SDN-based data/traffic routing solutions. Recently, some schemes employed machine learning to predict future contacts of vehicles and data-deliverysuccess-probabilities.

B. V2I Based Data/Traffic Offloading
Many studies have focused on data offloading techniques leveraging V2I communication mode. Furthermore, most studies are based on V2I data offloading architecture design, RSUassisted data offloading optimization problem formulation, and data transmission algorithms design. Fig. 4 demonstrates the concept of V2I data offloading scenarios, where solid blue link lines represent V2I, and doted black link lines represent infrastructure to infrastructure (I2I), and I2C communications. The scenario in Fig. 4-A represents an example of store-carry-andforward mechanism while Fig. 4-B represents cellular network offloading while using RSUs and WiFi-APs opportunistically. Vehicle V a in Fig. 4-A follows as a store-carry-and-forward mechanism for bulk data offloading, carrying data from a filling station to its destination point, where it offloads the data to a destination RSU linked data center. This data source and destination point could be any location, such as a filling station, garage, parking lot, RSU or BSs, linked to data centers. Vehicles Fig. 4-B scenario tries to improving the efficiency of data transfer and reducing the load on the cellular network while opportunistically switching among RSUs and WiFi-APs. Both of Fig. 4-A and Fig. 4-B scenarios are adopted to improve content availability, cellular network offloading, and energy, cost, and delay optimization.
In this subsection, we review the offloading techniques considering V2I based communications and further sub-categorize V2I based data offloading schemes into downlink and uplink data offloading domains, respectively, which are discussed as follows: 1) Downlink Data Offloading: Mobile users put a massive load on the cellular infrastructure while downloading content. In particular, vehicular users participate in information retrieval on the move. ITS can play an essential role in offloading the cellular networks in this context. The V2I-based  We categorize the state-of-the-art techniques according to their objectives, such as maximization of content availability, optimization of cost and delay, and offloading of the cellular network. The following is a detailed discussion of these objectives: a) Maximization of content availability: In vehicular networks, the duration of communications is short and rarely happens. Thus, maximizing the amount of data transfer for each contact opportunity is critical. Silva et al. [74] proposed scheme to utilize the users' trajectory information and predict future requests and cache the content accordingly. This improves the content availability and reduces the delay in data delivery. The two forwarding strategies and the neighborhood discovery protocol work together to efficiently deliver the cached content to the users. However, the constantly changing nature of vehicular networks makes it challenging to deliver the content efficiently. Similarly, Silva et al. [75] examined the development of content delivery networks (CDN) in the view of vehicular networks. RSUs help communication by delivering and replicating content to vehicles under their coverage range. The authors devised a policy for estimating content delivery performance. The proposed metric is used to design a deployment strategy that identifies the best locations to deploy RSUs to adequately support the dissemination of content, each of which needs specific performance levels.
b) Optimization of cost and delay: Access to content from the Internet on cellular infrastructure involves cost and delay. Optimizing these, Cheng et al. [76] proposed a heterogeneous offloading framework, which explores vehicle-assisted D2D networks to present delay-tolerant data in a store-andforward manner. In addition, a dynamic mode selection and resource allocation algorithm were developed to maximize the overall average delivery ratio while ensuring fairness for smart grid users.
Cheng et al. [77] presented an approach to predict WiFi offloading potential and access cost in a vehicular environment. Two offloading mechanisms are presented, i.e., auction & congestion game-based offloading, for vehicular users to offload cellular traffic over the WiFi network efficiently. With the implementation of the proposed mechanism, cellular traffic can automatically and transparently be offloaded for vehicular users. Thus, the cellular network congestion could be alleviated. However, it is challenging to utilize the Internet by APs in moving vehicles as WLAN APs have a short-range and are generally not deployed to cover roads.
In another work, Chen et al. [78] analyzed the utilization of extra resources in intermittently connected vehicular networks (ICVNs) to support target vehicles to download and process files from a remote server. A fog-enabled scheme is introduced for ICVNs to distribute the pooled resources among vehicles. Specifically, each target vehicle is served by a virtual mobile fog node whose resources are extracted from a set of supporting vehicles. The corresponding supporting vehicles download parts of data from the server and then perform preprocessing. The authors formulated the download problem as a mixed-integer scheduling problem, which reduces the target vehicles' time to receive the entire processed data while considering each support vehicle's V2V communication interference, computing, and storage capacity limitations. Additionally, a heuristic algorithm is devised to schedule the resource pool to help target vehicles.
Evaluating the tradeoff between the downloading delay and cost, Wang and Wu [79] provided a decision mechanism to determine whether it should download the data from the cellular network or RSUs. The authors unified the downloading cost and downloading delay as the users' satisfaction. The authors presented an adaptive algorithm to examine the problem of interface scheduling in cellular networks and VANETs and designed downloading approaches based on historical contacts among vehicles and RSUs. Similarly, Zhang et al. [80] proposed 5G enabled on-road BS as self-sustaining caching stations (SCSs). This approach can enhance the network's capacity in a cost-effectively way. The authors recommended the deployment of these SCSs, while the traditional macro BSs operate to ensure coverage and provide control signaling. Moreover, a hierarchical management scheme is devised to address the challenges of renewable energy supply and dynamic traffic allocation management. To understand the vehicular network in 5G, the authors conducted a case study on size optimization of cache and sustainable energy traffic management.
c) Offloading of cellular N/W: The alternative road network is envisioned to play a vital role in reducing the burden on cellular infrastructure. RSUs can take advantage of mobility prediction to determine what data to be obtained from the Internet and to schedule transmissions to vehicles. Malandrino et al. [81] analyzed the data transmission scheduling and content prefetching from RSUs of finite-horizon and inaccurate mobility prediction in a realistic scenario. They found that if the prediction error is not overwhelming, ITS can effectively serve vehicles, either by direct download from RSUs or relays, thereby alleviating the cellular network from the download traffic. Furthermore, they presented a probabilistic graph-based representation of the system to explain the uncertainty of the prediction, which results in near-ideal offloading efficiency. Lee and Lee [82] proposed an offloading algorithm that examines historical patterns of the user's mobility to determine whether to transmit data to the WLAN rather than switching to the cellular network. They considered user application usage patterns to predict available WLANs. However, the algorithm's performance needs to be measured for high-speed vehicles. Similarly, Malandrino et al. [83] presented a model to calculate different degrees of prediction. The proposed model can reproduce the prediction accuracy of Markovian techniques. A graph-based representation is proposed that incorporates the prediction information and improves content prefetching and transmission scheduling.
Zhioua et al. [84] studied to assess the potential of VANET to offload the cellular traffic analytically. The offloading decision analyzes the constraints associated with the availability of the V2I link, capacity and quality of the V2V link, channel and medium contention, data flow volume, and the link connectivity span of the vehicle and the RSU. The data offloading is affected by the data volume of flows and the vehicular links' quality. In another work, Zhioua et al. [85] also utilized VANETs to offload cellular data traffic and modeled an optimization problem to choose a maximum target set of flows to offload data to the destination via VANETs. Nevertheless, different from the work in [84], the authors studied the link availability constraints and channel load by analyzing medium contention. The results revealed that the traffic data volume, channel load, and link quality substantially impact the performance of data offloading.
Like [81], [82], [83], [84], and [85], Zhioua et al. [86] investigated the use of ITSs for offloading cellular data traffic. This model intends to choose a full target set of flows to offload a part of cellular data over VANETs. In this paper, the link quality of V2V, V2I, and VANET capacity on data offloading decisions is examined. Moreover, a cooperative traffic transmission under the 4G's collaboration is presented, i.e., LTE-A, cellular network, and VANETs. The results showed that the performance of data offloading is highly related to the data volume, path quality of the traffic, and channel load. Emerging 5G communication models have ensured that real-time infotainment systems can run efficiently by offloading and opportunistic spectrum sharing for vehicle nodes. Harris et al. [87] pointed out the advantages of offloading to WiFi via Drive-thru internet in automotive applications. The authors presented an approach of opportunistic access relying on beacon interval white time to macro-cellular WiFi APs to maintain an active connection during moving in an urban area and overcome the inefficiencies of successive handovers and increase data transfer to the vehicle.
In order to promote data offloading, the authors of [88] adopted the Manhattan mobility model and focused on the network architecture for data transmission to reduce the overall transmission time and the effect of retransmission as anticipated by the temporal convolutional network (TCN) model. They created a new data transmission delay model leveraging fixed-route automobiles as data carriers between data centers utilizing a genetic algorithm based on a RL mechanism (RLGA) to pre-allocate resources for offloading requests. Consequently, fewer transmission hops and retransmissions, and a shorter overall data transmission time when applied to the Seattle traffic count dataset were achieved. The development of novel network topologies, in particular space-air-ground integrated networks may help enormous amounts of data transmission might then be the subject of research.
2) Uplink Data Offloading: This sub-section investigates the different schemes proposed for uplink data offloading utilizing the V2I mode of communication. These offloading schemes can help when network connectivity is scant or with no end-to-end connectivity while targeting both delay-tolerant and delay-sensitive data. Therefore, these schemes can be further categorized into objectives like optimization of delay and cost and offloading of cellular networks. a) Optimization of delay and cost: The work by Cheng et al. [89] proposed an analytical relationship between offloading and the average service delay in a vehicular environment. They designed queuing and mobility models to avoid delays greater than the expected service delay and proposed an analytical model to offer offloading instructions to both vehicular users and network operators. However, the work only considers WiFi access network infrastructure and does not take into account continuity issues. Additionally, the planned deployment of WiFi access points could improve offloading performance.
Different from [89], Bazzi and Zanella [90] proposed two routing protocols named greedy forwarding with available relays (GFAVR) and greedy forwarding with virtual RSUs (GFVIR). To deal with the VANET's local minima problem, which occurs when the GF approach is utilized for stationary RSUs. The GFAVR algorithm is distributed, and no prior knowledge is required except a single overhead bit. In contrast, the GFVIR demands a preliminary design phase to individualize the local minima and alternate routes. Besides, GFVIR requires an increase in the RSU database, not providing an additional signaling overhead. Although GFAVR is easy to implement and independent of the specific scenario, GFVIR delivers better performance in most cases. However, with frequent network partitions, limited link lifetimes, and high vehicle velocity, providing accurate traffic information remains a challenge [91].
Data offloading complements the limited cellular spectrum and bandwidth against users' data-hungry application demands. To make offloading more robust and attractive, Raja et al. [92] proposed a reward-based offloading scheme. The scheme comprises RSU selection, reward-based offloading, traffic splitter, and RAT selection algorithms. The RSU selection algorithm runs at a centralized SDN controller and utilizes the network and vehicle information. The ML algorithm is designed for reward-based offloading, where non-safety applications are first split among RSU, WiFi, or cellular nodes. Next, the RAT selection algorithm identifies the traffic to be offloaded. Q-learning is used in the RAT selection algorithm that calculates the rewards for RSU and the cellular selections in user/vehicle proximity. User/vehicle selects the possibility of having maximum reward value to perform offloading. The experiments validated the scheme performance, reduced the RAT selection process time to 17%, and improved the overall throughput by around 15%.
Enabling big data analytics in ITS involves using important traffic patterns and trends to make more accurate and time-bound smart decisions, especially in CAV environments. Hu et al. [93] presented an integrity-oriented data offloading policy with variable modulation and coding schemes for the relay vehicular network. This analytical model on link duration considers the effect of the mobility characteristics and fading channel, which is used to describe the general, intermittent Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. vehicular network mathematically. The relay transmission can help decrease offloading delays, and an efficient scheduling method is necessary to maintain transmission integrity and timeliness when the RSUs are sparsely deployed.
Tassi et al. [94] developed an infrastructure based on fog computing, which can achieve an agile offload of CAV data. Since CAVs are expected to generate large quantities of data, it is impractical to consider the completion of the data offloading process when CAV is within the communication range of RSU as it is expected that CAV is only within the RSU range for a limited time. The authors proposed an agile fog computing infrastructure to solve this problem, which interconnects all the RSUs and collects sensor data transmitted by CAVs. Both RSUs and CAVs communicate via ETSI's ITS-G5 communication. Therefore, the data reconciliation is solved efficiently by deploying the random linear network coding technique. The obtained results illustrated the feasibility of the solution and showed its efficiency when running on a large city test platform. Different from [94], Darwish et al. [95] introduced a traffic-aware data offloading (TRADING) method for big data-centric ITS applications in CAV. A gateway discovery algorithm is designed to select an offloading gateway depending on the load at gateways, network, and traffic status. The gateways advertise their traffic and network status in the vicinity. The scheme assigns a value to the gateways based on their load, traffic, and network status. A gateway having a higher score value is selected for offload. The TRADING balances offloading data traffic between gateways by focusing on vehicular traffic and the state of the network in the proximity of gateways. Also, TRAD-ING alleviates the impact of the advertisement burden by releasing the transmission channels to transmit big data traffic. The proposed scheme showed an advantage in terms of 2% reduced offloading delays, 98.39% overall end-to-end delays, and reduced 81.6%, 84.95%, and 79.94% gateway overhead under low, average, and high vehicle densities, in comparison to the existing schemes.
A QoS provisioning data offloading scheme is proposed by Saleem et al. [96]. Three functions, traffic classification, overload control, and admission control, are doing QoS provisioning. The first function is for data prioritization, then the RSU load balancing is done through the overload control function, and the offloading vehicles and RSU service administration is done using the admission control function.
Urban tracking is vital to ITS safety, but existing monitoring systems consume a lot of bandwidth and computing resources to store, process, and analyse data in the cloud. For minimising sensor data offloading while ensuring tracking success, in this paper [97], an ITS crowd-AI hybrid urban tracking system uses a graph-optimized data offloading scheme with two elements: trajectory prediction and task allocation. The task allocation approach selects individuals and cameras to cover projected regions of interest based on object trajectories from state graph analysis. Using real-world datasets, the proposed scheme outperformed the compared schemes and also revealed shortcomings in the suggested strategy that could be solved with more investigation. b) Offloading of cellular N/W: Obtaining environmental information from vehicles is likely to overburden cellular networks. The V2V and V2I wireless communication technologies are being utilized for offloading cellular networks. Visible light communication (VLC) is also a breakthrough in this direction. Bazzi et al. [98] studied the feasibility and the efficiency of VLCs for cellular network offloading in vehicular networks. They found that using traffic lights as RSUs for VLC communication can offload more than 90% of cellular traffic, but VLC alone is not enough to replace RF systems such as DSRC and cellular, which allow for long-range non-line-ofsight links [99].
Liu et al. [100] analyzed a network recommendation system for offloading in VANETs that uses a big data analysis framework to study traffic status, network conditions, user preferences, and service applications. They developed an android application that allows vehicles to access the network based on the recommendations provided by the system. This scheme can effectively choose the best vehicle network and fully utilize network resources. However, this study only focuses on the performance of recommender systems in specific fields, and may not be suitable for other advertising systems in VANET.
Summary: The cellular users lay a massive load on cellular infrastructure while content downloading and uploading, especially when considering it for ITS. The ITS-V2I infrastructure includes RSUs to provide on-road data exchange services and applications to complement the cellular infrastructure. Therefore, accompanying offloading utilizing V2I mode with ITS also counterparts the cellular infrastructure. Several offloading schemes are discussed in the literature falling into the V2I offloading category, in which [74], [75], [76], [77], [78], [81], [82], [83], [84], [85], [86], [87] focused on downlink offloading for delay-tolerant data, while [79], [80] considered delay-sensitive data offloading. Similarly, for uplink, delaytolerant data offloading schemes are investigated in [89], [90], [91], [92], [98], [99], and [100], while [93], [94], [95] focused on the delay-sensitive offloading schemes. While investigating the opportunistic data offloading solutions, some schemes have considered mobility prediction for content caching, fetching, and prefetching. Similarly, some researchers consider link availability, resource pooling, and content delivery estimation for partial data offloading. Also, in the literature, cellular interface selection/scheduling, queuing, and RSU-V2V-RSU relaying schemes are also discussed. For ready reference, we summarize the state-of-the-art research contributions in this category in Table V. C. V2X Based Data/Traffic Offloading V2X based communication is a versatile mode, and several research efforts focused on data offloading techniques using V2X communications. Here, we group together all the methods that utilise V2V, V2I, V2N, or V2P modes or that frame problems in such a way as to demonstrate the simultaneous usage of various modes, such as V2I for communication with the RSU and V2V for communication with another vehicle or V2N for data offloading. The illustration in Fig. 5 shows    Fig. B is an example of store-carry-and-forward mechanism, mostly adopted for bulk data transfers. The scenario in Fig. C follows a relaying mechanism. The scenarios in Fig. A and Fig. C are commonly adopted to improve content availability, cellular network offloading, and energy, cost, and delay optimization. several scenarios of data offloading using V2X technology, where doted and solid purple lines represent V2V and V2I communications, doted and solid red and green lines represent V2N, V2P, and person to infrastructure (P2I) links, respectively, and blue lines represent I2I links. Each scenario in the illustration can be applied to multiple use cases and falls into various application categories. The scenario in Fig. 5-A shows an example of uplink and downlink data offloading, where black and green vehicles act as cluster heads for blue and red vehicles, respectively, and the green vehicle also serves nearby pedestrians. Both black and green cluster heads can form clusters for uplink or downlink offloading on demand. The scenarios in Fig. 5-B is used in V2I cases, while the scenarios in Fig. 5-A, and Fig. 5-C falls in the category of V2X cases. The scenario in Fig. 5-B is commonly used for bulk data transfers, and the scenario in Fig. 5-C is typically applied to offload cellular networks, increase content availability, and optimize costs, energy, and delays.
In addition, some researchers investigated V2X-based data offloading architecture design, formulated optimization problems, and designed data transmission algorithms. In this subsection, we review the research efforts contributed towards V2X based data offloading techniques by sub-categorizing this category into downlink and uplink data offloading schemes, respectively, which are discussed as follows: 1) Downlink Data Offloading: V2X communication is a technology that allows vehicles to communicate with other vehicles, infrastructure, and devices in their surroundings. In the context of IoT and smart cities, V2X can be used as a solution for downlink data offloading, where data traffic is transferred from cellular networks to V2X networks in order to alleviate congestion and improve performance. This can include bulk data transfer, content availability and delivery, network offloading, and optimization of energy, cost, and delays.
a) Bulk data transport: Exploiting the benefits of cognitive radio (CR) merger in vehicles' OBUs, Ding et al. [101] proposed a vehicular CR capability harvesting network (V-CCHN) framework, which works on the principle of storecarry-and-forward for a short period. The architecture of V-CCHN includes cognitive radio-capable vehicles, cognitive RSUs, and a virtual service provider (VSP). In the proposed framework, the delay-tolerant data can be loaded on available OBU's free space and transferred between data networks and end devices using CR technology. Hence, the delaytolerant data can be offloaded opportunistically whenever an idle spectrum is available. An extension of this work is provided in [102] by focusing on the data routing problem at intersections. A spectrum-aware (SA) data routing scheme is proposed. The data delivery is formulated as a Markov decision process (MDP) at a centralized secondary service provider (SSP) unit. The SSP solved this MDP process using dynamic programming and optimal routing decisions. The proposed scheme considered Shanghai's traces and simulation results outperformed the baseline schemes against the minimum delay routing algorithm (MDRA) and GPSR. This work can be enhanced by assuming different users' activities on licensed and unlicensed bands.
Safety-related data/information in VANET is always tiny and needs less bandwidth; however, such traffic has a high transmission frequency. Focusing on the poorly used bandwidth and VANETs-Internet collaborations, Qin et al. [103] proposed a delay-tolerant data offloading scheme by intelligently utilizing the underutilized network resources. First, delay-bound data categorization is performed. Then, the possible V2I and V2V routes between source and destination are determined. Finally, routes satisfying the delay requirements are filtered, and the data to be transferred is scheduled accordingly using a defined scheduling algorithm for both V2V and V2I links. Routes filtration is based on congestion avoidance, data delay-bound, and traffic dynamics. Vehicles play the role of data loading, carrying, and forwarding in this scheme. While incorporating handling-of-expired data, their experimental setup under diverse VANET dynamics demonstrated satisfactory performance.
b) Maximization of content availability and delivery: A single network can be inadequate while facing the proliferation of mobile devices and data heterogeneity services. Hence, increasing content availability and access time offloading to other networks is an effective solution. Chen et al. [104] proposed a model for data offloading in mobile networks that addresses the issue of fleeting connection time among vehicles and RSUs. The authors proposed a two-phase resource allocation scheme that employs wireless and backhaul link patterns for instant bandwidth allocation and link scheduling, with the goal of optimizing system throughput. Additionally, the authors highlighted the challenge of geographical unevenness of traffic density in 5G networks, which results in hot spot overloads and poor resource utilization in under-loaded areas. Edge caching is like offloading the network backbone; Ndikumana et al. [105] exploited self-driving AI-supported OBUs and proposed a deep learning-based edge caching scheme for infotainment services in the vehicular edge. A multi-layer perceptron (MLP) CNN model is developed to predict the probability of infotainment data. Moreover, a communication model is introduced to access infotainment data in a cache, and a management model is also developed. Besides, a block successive majorization-minimization (BS-MM) based optimization problem is proposed to link proposed models aiming to reduce the downloading delay. The results demonstrated that their caching solution boosted the system significantly and reduced about 61% traffic load from the system backhaul. In another work, Huang et al. [106] also proposed an edge caching scheme in a vehicular network. The authors proposed a delay-aware content fetching algorithm to optimize the pre-fetching and caching decisions at RSUs, and the vehicle's V2V and V2I associations. The scheme works for both non-handover and handover scenarios. The proposed scheme provided a better cache hit ratio under diverse traffic scenarios.
Bhover et al. [107] studied a non-DSRC or WAVE method to V2X communication to increase content availability. Offloading cellular data with ZigBee, WiFi, and cellular networks reduced cost. Cellular network connectivity and vehicle monitoring are proposed. WiFi enhances content availability, throughput, and power usage. Unlike [107], Ko et al. [108] presented a service architecture using cooperative V2I/V2V communications to coordinate centralised scheduling within RSU coverage and ad hoc data scheduling. This architecture allows RSUs to cooperate and load balance to promote ad hoc data exchange across vehicles travelling in different directions. To exploit V2I and V2V communication, the authors used centralised data broadcast and ad hoc data sharing to create a hybrid data scheduling problem. RSU cooperation-based adaptive scheduling considers centralised RSU scheduling, vehicle ad hoc scheduling, and cluster management.
Data delivery is one of the prime concerns in communication networks. Jiang et al. [109] made a relation of content delivery with deep CNN in the vehicular communication network. A two-dimensional coverage grid is generated from GPS trajectory data. The grid value determines the possibility of V2V and V2I communication, leading to the delivery ratio calculation. In the content delivery scheme based on a maximum flow-directed network, the deep CNN utilizes this delivery ratio and grid data to generate training images and training targets. The obtained results depicted that the deep CNN models accelerated the delivery-ratio-calculation process. ResNet 50 was the best model, which provided less average calculation time than the other models. Hence, the delay in content offloading decisions can be significantly reduced using the stated model.
The edge caching can accelerate the content access response. However, content popularity varies over time. Therefore, it is a challenge for traditional content-popularity methods to perform cache decisions. To accommodate on-road content popularity dynamics, Wang et al. [110] introduced a request predicting cooperative caching scheme in VANETs. The proposed approach performed prediction based on historical cache hit counts and time series analysis. Vehicles are clustered using k-means according to their positions and speeds, ensuring the content transmission stability through vehicle dynamics. Content popularity is obtained using Zipf's law, content hit history, and time series. Following that, the content hit ratio maximization function is solved using Q-learning. Contents are cached at the edge (i.e., RSUs), then cluster heads (CH) collect copies of the cached contents. Later, both the RSUs and the CHs respond to the requests of proximity vehicles only when the requested content is already cached. The results demonstrated that the proposed scheme achieved 8-10% and 5-7% better performance in terms of delay minimization and cache hit ratio maximization, respectively. However, the proposed scheme has limitations in inter-RSUs and inter-CHs autonomous cache decisions.
In [111] the authors proposed a V2I and V2V scheduling technique to minimize the total number of content distribution time slots. First, the RSU serially transmits integrity content to vehicles during the V2I phase based on the vehicular network topology and transmission scheduling mechanism. Then, fullduplex communications and concurrent transmissions are used in the V2V phase to achieve content sharing between vehicles while improving transmission efficiency.
c) Offloading of cellular N/W: Feng et al. [112] introduced a vehicle assisted offloading (VAO) scheme that utilizes a vehicle queue standing at the red signals to offload cellular data from a busy road cell to its nearby idle cells. Besides, free roadside cell resources can be used to reduce cross-cell congestion, which dramatically enhances wireless resources. The authors conducted a theoretical study on the efficacy of the suggested scheme by analyzing road traffic flow and vehicle and pedestrian communications. The authors adopted fluid theory to simulate road traffic and analyzed the VAO scheme's theoretical performance. However, this is still challenging to persuade the vehicles to participate in the offloading.
Vehicle applications set distinct QoS constraints on information exchange. As a result, the performance required for the services differs significantly concerning latency, bandwidth, and communication reliability. To tackle this issue, Brahmin et al. [113] presented a vertical handover (VHO) management algorithm, where two network interface types are defined, i.e., the IEEE 802.11p and LTE as a primary and secondary interface, respectively. The authors proposed an algorithm that enables the vehicle to be transferred to the secondary interface if the packet loss of the services surpasses the permissible threshold. Then timer initiates the time for the user for as long as it stays connected to the interface. When the timer goes off, and the approximate primary interface packet loss ratio is not more than the maximum acceptable threshold, the vehicle is handover to the primary interface. Similarly, Huang et al. [114] conducted offloading from the cellular network to the DSRC network under SDN architecture. The authors suggested offloading with handover decisions based on the SDN model. The controller collects all vehicles' and RSUs' information, such as speed, location, direction, and senses nearby RSUs' IDs to determine whether it is helpful to offload or not. In addition, the proposed scheme examined the quality of the network and the expected vehicle residence time within the RSU for making the offloading decision.
The authors in this paper [115] proposed edge caching and immune cloning methods to provide IoV information faster and cheaper. The RSU predicts content popularity using a forward neural network fed by vehicle node queries' histories. After that, it actively caches popular content from the BS to speed up fetching and enhance the hit rate. The requesting node chooses content sources during content distribution. Optimizing with immune clones is suggested. The proposed content distribution improves network efficiency, hit rate, and vehicle user node needs.
2) Uplink Data Offloading: With increasing vehicles and their services, cellular networks are experiencing a tremendous traffic burden. Offloading traffic on VANET is an encouraging way to address this overloading problem. Different V2X-based uplink data offloading schemes are discussed in the following subsection.
a) Minimization of cost, energy, and delay: In vehicular communications, roadside infrastructure such as WiFi APs often requires a significant investment. Yang et al. [116] proposed a solution called "Mobile vehicular offloading (MoVe-Off)," which does not require additional investment but allows data to be transferred from on-board devices to mobile devices of drivers and passengers for uploading to the internet. The data is offloaded in a delay-tolerant manner when the users arrive at places where they can access WiFi. The authors built a realistic system to study users' travel routines and WiFi usage, and used deep learning to predict individual mobility. They also developed a long short-term memory (LSTM) model to encourage collaborative offloading, and a routing scheme for inter-vehicle communication. In this scheme, every vehicle transmits its expected offloading probability and delay, and the information is dynamically delivered to the nodes providing an additional and stable offloading service for delay-tolerant data. The proposed MoVeOff approach can provide a costeffective solution for vehicular data offloading by utilizing the mobile devices of drivers and passengers, instead of investing in additional roadside infrastructure.
In ITS, vehicle data collection within a target area is a concern for many applications. Mir et al. [117] put forward a broadcasting scheme applying transmission power adjustment. A multi-tiered architecture is defined for hybrid vehicular networking. Besides, a centralized transmission power adaptation technique is introduced. Given the current topology state, the contextual information on vehicle mobility is obtained in estimated link expiration time. Then transmission power is allocated to each vehicle that guarantees connectivity with long-term links. The proposed technique results in a much lower packet collision rate and delays while keeping a shorter transmission range. However, it should be evaluated through real big traffic data sets.
Wang et al. [97] presented a graph optimized data offloading algorithm that uses a crowd-AI hybrid method to reduce the data offloading cost while ensuring the reliable urban tracking result. The authors solved the challenge by dividing it into trajectory prediction and task allocation. First, the trajectory prediction algorithm computes possible tracking areas of the target object by taking advantage of the state graph. Then, the task allocation algorithm using the dependency graph minimizes the data offloading cost separately.
b) Offloading of cellular N/W: Offloading the overburdened cellular network, Wang et al. [118] examined offloading in VANETs with fixed and mobile offloading nodes. An offloading model is proposed to estimate the offloading capacities of APs and vehicles by a connectivity graph. The graph determines which node, i.e., AP or vehicular, will be chosen as the offloading node. The authors proposed combined optimization to facilitate offloading and mitigate mobile traffic. The authors handled mixed-integer programming to achieve optimal solutions while satisfying global QoS. To this end, the authors modeled data offloading as a multi-objective optimization problem to minimize mobile data traffic and provision QoS-aware service.
In order to support the communication transport of FCD, LTE-VANET hybrid architecture is envisaged. Salvo et al. [119] proposed FCD message flows in LTE-VANET hybrid architecture. This framework reduces the LTE channel capacity needed by previous approaches substantially. FCD collection is achieved by connecting each vehicle with its established channel across the LTE-RAN. The authors adopted a distributed procedure that utilizes the capability of vehicles to communicate with each other through a V2V channel based on VANET to choose represented nodes. There is an evident performance gain, as the results confirm the number of LTE resource blocks saved.
To make intelligent decisions for data offloading, Turcanu et al. [120] proposed a distributed clustering algorithm for intelligent data offloading in connected and autonomous vehicles. This algorithm selects forwarder vehicles to send their data and the data of their neighboring vehicles through LTE, and uses DSRC to decrease the LTE channel utilization significantly. The authors also analyzed the channel quality indicator to evaluate the performance of the LTE uplink in this approach. The algorithm aims to optimize the use of resources in the network and makes the system more efficient by reducing the congestion on the LTE channel and offloading the data traffic to the vehicles in the network. Designing an adaptive approach that can handle the challenges of unreliable connectivity and rapid topological changes due to vehicles' high speed is a difficult task. Aujla et al. [121] proposed a 5G-enabled SDN-based data offloading approach that uses VANETs to offload cellular traffic. The SDN-based controller in this approach manages data offloading through the use of a priority manager and a load balancer. The proposed approach is able to perform efficiently even in the case of an overloaded network by utilizing these two SDN-based controller managers. The approach uses the flexibility of SDN to manage the traffic and reduce collisions and delay. This approach can effectively solve the problem of unreliable connectivity and rapid topological changes in vehicular networks. Similarly, Bentley et al. [122] studied the feasibility and profitability of transferring data from one point to another in a resource-constrained and RF-challenged environment, instead of relying on a slow direct link. They found that transferring files of even relatively small sizes to a proximal vehicle in harsh communication environments is appropriate. The study provides insight into situations where existing network infrastructures are limited and/or have no access to services from most cloud vendors. The study suggests that data transfer via proximal vehicles can be a viable solution in such scenarios, allowing for more efficient use of resources and improving overall performance.
Lin et al. [123] proposed a k-hop vehicle to roadside unit (VVR) data offloading path to offload the LTE cellular network's data traffic to the vehicular network. The SDNbased proposed approach employs the idea of time-extended prediction inside the MEC server controller to determine the best VVR data offloading path, which exists during the timeextended prediction period.
Summary: V2X communication mode is considered the most robust type for data offloading as it facilitates both V2V and V2I communications. With V2X mode, a vehicle can communicate with nearby vehicles, local RSU/BS, or both, relying on functional channel conditions. The literature contains several V2X-based data offloading schemes. For the downlink data offloading category, some schemes are designed to handle delay-tolerant data in [101], [102], [103], [104], [105], [106], [112], [113], and [114], while the others [107], [108], [109], [110] focused on delay-sensitive data. In the same manner, the following schemes [116], [118], [119], [121], [122] focused on uplink delay-tolerant data offloading, while delay-sensitive data offloading is handled in [117] and [120]. Some of the reviewed offloading schemes considered spectrum-aware cognitive radio networks and SDN-based approaches to offload cellular traffic over VANET's momentary connections. In contrast, some schemes contemplated vehicle-assisted offloading and handover schedules to complement the cellular traffic over V2X. Moreover, deep learning-based CNN models, delayaware edge caching, store-carry-and-forwarding, VANET clustering, and FCD-based schemes are also presented in some papers. Finally, for the reader's facilitation, we summarize the state-of-the-art research contributions in this category as shown in Table VI.

IV. INSIGHTS GAINED AND OPEN ISSUES
The purpose of this article is to gain an understanding of how VCN modes can be utilised for data offloading in vehiclar environment in order to improve the overall performance of vehicular applications and services. To be more specific, and in contrast to other surveys that have been conducted in the underlying domain, our goal is to examine the VCN-based vehicular data offloading and provide insights into critical issues and challenges in the design of the system.
• Cellular BS or RSU deployment needs to address issues, such as higher installation costs, limited coverage, and network maintenance. Vehicles are usually large in number, so network expansion is required to support high traffic loads while maintaining stable and acceptable network performance, such as vehicle users' QoE. On the other hand, when the RSUs are sparsely deployed and an efficient scheduling method is necessary to maintain the transmission's integrity and timeliness, the relay transmissions can help increase the offloading capacity. In terms  of network scalabilit y, efficient techniques like clustering, load balancing, and synchronized multi-processing need to be effectively exploited to meet the demands of vehicle customer services and applications.
• VANET offloading is an encouraging solution to the conflict between the inadequate capacity of the cellular network and the collection of big data. The practical choice of network is essential to guarantee the QoS of vehicles. 5G communication models have ensured that real-time infotainment systems run efficiently using offloading and opportunistic spectrum sharing of vehicle nodes. However, in the offloading scenario, data size optimization and sustainable energy traffic management are also important points that need to be addressed. Furthermore, providing accurate traffic information remains a challenge with frequent network partitions, short lifetime communication links, and high vehicular speeds.
• Many of the previous research works have made unrealistic assumptions, including perfect mobility prediction, fixed and equal size of contents, identical vehicles, average or fixed road speed for all vehicles, fixed vehicle density, constant throughput, constraint-less environment, and fixed TTL. Similarly, user activities are assumed on licensed and unlicensed bands, fixed communication ranges, and average contact rates for all vehicles. All the assumptions are critical factors in performance measurements. However, with synthetic mobility traces, one does not know the behavior of these schemes in real-world mobility traces? Also, neglecting the power consumption and ignoring the probability of data loss may affect the overall system performance.
• Furthermore, as already highlighted, accessing accurate traffic information is a challenge with limited link lifetime and fast vehicle speeds. Therefore, some practical models need to be developed to fill this gap.
• To meet such real-time data applications' requirements, exploiting V2V data offloading with mmWave is still an open research direction. The probability of change in vehicular topology is very high; furthermore, devices' re-configurations, re-calculations of routing paths, and supervision of in-vehicle data are still challenging, time and resource-consuming overburdening tasks that need to be addressed.
• Managing the limited spectrum resources available for data offloading in the vehicular environment is a major challenge, as the high mobility of vehicles can lead to rapid changes in network conditions.
• Deploying and managing data offloading infrastructure in the vehicular environment is a another significant challenge, as it requires coordination between multiple stakeholders such as vehicle manufacturers, service providers, and government regulators.
• In automotive networks, communication duration is concise; providing content in such dynamic networks is not a trivial task. Therefore, it is crucial to maximizing data transfer for each communication opportunity. Internet access through AP for moving vehicles is challenging because of WLANs' low coverage, which is not usually deployed to cover roads. However, VLC is another technology apart from mmWave to be used. Though VLC brings high bandwidth and secure communication, it can still not replace high-speed RF communication systems (i.e., DSRC, LTE, and NR), which allow long-range nonline-of-sight links.
• Vehicular data offloading through V2X is a hot research and application area. The practical network information, including vehicle mobility, locations, congestion situations, social relationships, traffic rates, and link connections, must be obtained and predicted in time. Which is essential for the associated network components to design data offloading schemes.
• Providing consistent and reliable QoS for data offloading in the vehicular environment is a major challenge, as the dynamic nature of the vehicular environment can lead to fluctuating network conditions.
• Managing network congestion is a major challenge, especially in urban areas where there are high densities of vehicles. This includes developing mechanisms to prioritize traffic, as well as identifying and utilizing underutilized resources to offload traffic.
• Developing strategies for energy-efficient data offloading is a significant challenge, particularly for electric and hybrid vehicles. This includes identifying methods for reducing the energy usage of the vehicle's communication subsystems and offloading data to conserve energy.
• Using AI algorithms, the process of data offloading in a vehicle context can be further optimised. Specifically, Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.
AI can be used to monitor network conditions, device capabilities, and data trends, and make dynamic adjustments to increase the data offloading process's efficiency, security, and scalability. This may involve choosing the optimal offloading approach, controlling network load, anticipating traffic circumstances, detecting and preventing security threats, allocating network resources, assuring interoperability, and maximising scalability.
• The use of aerial vehicles, such as drones or balloons, as relay stations or mobile network extenders is considered a promising solution for improving network performance and reducing congestion in vehicular environments. The use of drones or balloons can help to provide additional coverage in areas where ground-based networks may be limited, such as in remote or rural areas or in areas with challenging terrain [124], [125]. They can also be used to support data offloading, by providing a secondary network for transferring data traffic and reducing the load on the primary cellular network. This can help to improve network capacity and reduce congestion, which can be especially important for safetycritical applications in vehicular environments. Additionally, the use of aerial vehicles can also help to improve the reliability and availability of communications in these environments [126]. However, there are challenges that need to be overcome such as regulatory, safety, and security concerns.
• For retrieving network information, security & privacy are also notable concerns to consider because some network components are reluctant to share their information with other devices [127]. Therefore, in-depth research is required for secure communication schemes concerning data caching and transmission protocols to enable the practicality of the schemes.
• Motivating vehicles to participate in the data offloading process remains an open problem. In the vehicular environment, if two vehicular users lack in social-trust or have limited resources, these vehicles may show reluctance to offload data. Therefore, there is a need to develop trust and effective incentive mechanisms in this situation. For designing an effective incentive mechanism, the following incentives can be leveraged that include social incentives, pragmatic incentives, and indirect incentives [128], [129].

V. CONCLUSION
This paper presented a comprehensive literature review on CAV data offloading considering VCN modes. We endeavored and focused on a systematic overview and fundamentals of vehicular communication technologies, including DSRC and cellular technologies based on V2V, V2I, and V2X communication modes. We have categorized the vehicular data/traffic offloading schemes considering above mentioned VCN modes. Each category of CAV data offloading was supported with a summary to motivate more in-depth studies in the area. Finally, we highlighted the open research challenges in this field and predicted future research trends.