Adaptive activity recognition techniques with evolving data streams

2017-03-01T04:28:47Z (GMT) by Abdallah, Zahraa Said Emam Ammar
Activity recognition aims to provide accurate and opportune information on people’s activities by leveraging sensory data available in today’s sensory rich environments. Activity recognition has become an emerging field in the areas of pervasive and ubiquitous computing. The process of recognising activities flows through three key steps: sensing, modelling, and recognition. A typical activity recognition technique processes data streams that evolve from sensing platforms such as mobile sensors, on body sensors, and/or ambient sensors. Learning models in activity recognition are built from historical data and rely strongly on prior knowledge of activities. The learning model in this scenario is static and thus unable to cope with the evolving nature of activities in data streams. The evolving nature of activities arises for many reasons. Intuitively, people perform activities in different ways. “Walking” for one person could be “jogging” for another. Therefore, there is no model that fits all in activity recognition. To attain an accurate recognition, a learning model has to be tuned to suit a user’s personalised way of performing activities. Moreover, it is unrealistic to assume that the number of activities is static along the stream. While the learning model is built from historical data, novel activities may emerge and abandoned ones may disappear over time. This thesis develops adaptive techniques for activity recognition that dynamically change the learning model while activities evolve. These techniques apply an incremental and continuous learning approach for both personalisation and adaptation of the learning model. As a strategy to harness the potential of activity for pervasive environments, our techniques are capable of recognising activities that evolve from data streams. The first contribution of this thesis is to build a flexible, efficient, robust, and accurate learning model that enables personalisation and adaptation with evolving data streams. This learning model is the core for all our techniques developed in this thesis. Based on the developed learning model, we propose a technique for recognising activities efficiently. The recognition technique is an ensemble classifier that integrates with the learning model to recognise activities based on a hybrid similarity measure approach. The merit of this approach is to bring different perspectives together for more accurate recognition, especially across users. The ensemble classifier is evaluated on benchmarked datasets for activity recognition. The evaluation demonstrated the robustness, efficiency, and accurate recognition of activities. Our technique shows its best performance when applied across users and with noisy data. The accuracy is improved by more than 10% in these cases compared to other state-of-the-art techniques in activity recognition using benchmarked multidimensional datasets. The above activity recognition technique is extended to include incremental learning for personalisation with evolving data streams. This technique leverages the flexibility of the learning model for personalisation in real time to achieve an accurate recognition with the evolving activities. Furthermore, we deploy our technique on a mobile device to demonstrate its efficiency. Although the streaming environment imposes more constraints on the recognition process, the proposed recognition technique outperforms other benchmarked incremental techniques in activity recognition. Our technique shows its best performance when applied to data that contains noise with accuracy enhancement of about 15%. The last contribution is a technique that enables continuous learning to adapt the learning model. To fulfil this goal, our technique detects the arrival of new activities in data streams and/or the disappearance of abandoned ones. Moreover, it dynamically adapts the learning model with the detected changes for a future recognition. The developed technique is evaluated on benchmarked datasets to demonstrate its efficiency in recognising changes in activities and adaptation of the learning model accordingly. The recognition of novel activities varies depending on the characteristics of the datasets and the nature of the detected activity. This technique, as well as all techniques in this thesis, incorporates active learning to address the scarcity of labelled data especially in streaming environment by annotating only small amounts of the most informative data. Thus, this thesis takes a step forward in activity recognition dynamics in pervasive and ubiquitous computing by building efficient and adaptive techniques for recognising evolving activities. <div><br></div><div>Awards: Winner of the Mollie Holman Doctoral Medal for Excellence, Faculty of Information Technology, 2015.</div>