Stochastic pedestrian models for autonomous vehicles

<p>This dissertation addresses the modeling of pedestrians in dynamic and</p> <p>urban environments interacting with autonomous vehicles. The collision</p> <p>avoidance system of an autonomous vehicle has contrary safety and effi-</p> <p>ciency requirements. On the one hand, there might be collisions in following</p> <p>a risky driving policy. On the other hand, a safe driving policy might bring</p> <p>the passenger very slow to the target to prevent all kinds of risks. The au-</p> <p>tonomous vehicle does not know or perceive all relevant information, such</p> <p>as the unknown intention and the environmental and situational factors</p> <p>influencing the pedestrian’s behavior. There is a resulting decision dilemma</p> <p>for autonomous vehicles between road safety for all road users and efficient</p> <p>motion planning in environments with vulnerable road users. There also</p> <p>exists a lack of knowledge by predicting the future movements of pedestri-</p> <p>ans, where one could compute worst-case reachable state-sets. The areas</p> <p>of possible reach sets could get very large. An autonomous vehicle is not</p> <p>allowed to drive into these areas, making motion planning inefficient. The</p> <p>adaption to real-world scenarios is not trivial. The decision-making process</p> <p>in motion planning is challenging due to the enormous variety of situations</p> <p>and the uncertainty of predicting future human movements with absolute</p> <p>certainty. There is a potential risk of accidents in adapting and predicting</p> <p>human locomotion. These problems influence the trust and acceptance of</p> <p>autonomous vehicles with additional technological and legal challenges.</p> <p>This dissertation aims not to ensure total safety because of pedestrians’</p> <p>technical and diverse physical, cognitive, situational, and environmental</p> <p>complexity. This work uses a new method that combines machine learning</p> <p>with reachability analysis (resulting in an adaptive funnel, hull, or belief set</p> <p>computation). Machine learning adapts the reachability analysis to current</p> <p>situations. Therefore adaptive reachability analysis and corresponding mo-</p> <p>tion planning are presented and evaluated in vehicle simulations. Adaptive</p> <p>hull computational methods for adaptive implementation of reachabilianalysis lead to risky pedestrian bypassing. These computational methods</p> <p>provide a trade-off between safety and efficiency in motion planning. How-</p> <p>ever, the proposed approach cannot guarantee an exact threshold for overall</p> <p>safety due to the complexity of the problem (unknown intent and environ-</p> <p>mental factors). A very intuitive approach predicts the future situation’s</p> <p>maximal velocity, acceleration, and jerk of a pedestrian. Afterward, it com-</p> <p>putes the adaptive reach sets with conventional methods. Adapting classical</p> <p>worst-case reachability analysis could drastically reduce the cumulative</p> <p>volume of adaptive reachable sets compared to classical reachable sets (less</p> <p>than 70 percent of the cumulative area compared to classical computation</p> <p>in a presented use case). The results show a massive potential to reduce</p> <p>the areas from reach sets to belief sets. The causal inference could model</p> <p>the intention change. Intent changes are modeled with causal inference</p> <p>without considering structural learning and tested with model predictive</p> <p>control and adaptive hull computation in a simulation environment. This</p> <p>thesis also presents a similarity of human locomotion with (Partially Observ-</p> <p>able) Markov Decision Processes (MDPs and POMDPs). Nevertheless, many</p> <p>causal relationships describing pedestrian behavior are still unexplored.</p> <p>Concepts for building novel test environments provide an outlook on fun-</p> <p>damental research. New test environments might bring new insights into</p> <p>the causal information structure between cognition and human locomotion</p> <p>in different settings to improve accident statistics further. These new test</p> <p>environments could help make the adaptive reachable sets more robust and</p> <p>provide a quantitive guarantee for trusting novel hull computations.</p> <p><br></p>