Real-time predictive modeling and robust avoidance of pedestrians with uncertain, changing intentions

Abstract

To plan safe trajectories in urban environments, autonomous vehicles must be able to interact safely and intelligently with other dynamic agents. Due to the inherent structure of these environments, drivers and pedestrians tend to exhibit a common set of motion patterns. The challenges are therefore to learn these motion patterns such that they can be used to predict future trajectories, and to plan safe paths that incorporate these predictions. This thesis considers the modeling and robust avoidance of pedestrians in real time. Pedestrians are particularly difficult to model, as their motion patterns are often uncertain and/or unknown a priori. The modeling approach incorporates uncertainty in both intent (i.e., where is the pedestrian going?) and trajectory associated with each intent (i.e., how will he/she get to this location?), both of which are necessary for robust collision avoidance. A novel changepoint detection and clustering algorithm (Changepoint-DPGP) is presented to enable quick detection of changes in pedestrian behavior and online learning of new behaviors not previously observed in prior training data. The resulting long-term movement predictions demonstrate improved accuracy in terms of both intent and trajectory prediction, relative to existing methods which consider only intent or trajectory. An additional contribution of this thesis is the integration of these predictions with a chance-constrained motion planner, such that trajectories which are probabilistically safe to pedestrian motions can be identified in real-time. Hardware components and relevant control and data acquisition algorithms for an autonomous test vehicle are implemented and developed. Experiments demonstrate that an autonomous mobile robot utilizing this framework can accurately predict pedestrian motion patterns from onboard sensor/perception data and safely navigate within a dynamic environment