Data-Driven Reachability Analysis of Pedestrians Using Behavior Modes : Reducing the Conservativeness in Data-Driven Pedestrian Predictions by Incorporating Their Behavior

Abstract: Predicting the future state occupancies of pedestrians in urban scenarios is a challenging task, especially considering that conventional methods need an explicit model of the system, hence introducing data-driven reachability analysis. Data-driven reachability analysis uses data, inherently produced by an unknown system, to perform future state predictions using sets, generally represented by zonotopes. These predicted sets are generally more conservative than model-based reachable sets. Therefore, is it possible to cluster previously recorded trajectory data based on the expressed behavior and perform the predictions on each cluster to still be able to provide safety guarantees? The theory behind data-driven reachability analysis, which can handle input noise and model uncertainties and still provide safety guarantees, is quite recent. This means that previous implementations for predicting pedestrians are theoretically probabilistic and would not be appropriate to implement in actual systems. Thus, this thesis is not the first of its kind in predicting the future reachable sets for pedestrians using clustered behavioral data, but it is the first work that provides safety guarantees in the process. The method proposed in this thesis first labels the historically recorded trajectories into the behavior also referred to as mode, the pedestrian expressed, which is done by simple conditional statements. This is done offline. However, this implementation is designed to be modular enabling easier improvements to the labelling system. Then, the reachable sets are computed for each behavior separately, which enables a potential motion planner to decide on which modal sets are relevant for specific scenarios. Theoretically, this method provides safety guarantees. The outcomes of this method were more descriptive reachable sets, meaning that the predicted areas intersected areas that it reasonably should, and did not intersect areas it reasonably should not. Also, the volume of the zonotopes for the modal sets was observed to be smaller than the volume of the implemented baseline, indicating fewer over-approximations and less conservative predictions. These results enable more efficient path planning for Connected and Autonomous Vehicles (CAVs), thus reducing fuel consumption and brake wear.