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