Probabilistic Pedestrian Trajectory Prediction for Autonomous Vehicles in Crowded Environments

Abstract

In the recent decade, there has been a rapid growth in the field of assisted and autonomous driving, with notable advancements in the commercialization of assisted driving. Although commendable, automated driving systems face a huge challenge in highly dynamic urban areas where there are many interacting agents such as other vehicles and pedestrians. Among those, pedestrian behavior is more dynamic and therefore harder to predict. Furthermore, incorporating the predicted behavior into decision-making is another challenge. Current methods mainly focus on the first aspect of the prediction while ignoring its probabilistic aspect. In this research, probabilistic multimodal trajectory prediction model is proposed to help solve those orthogonal problems. To this end, Gaussian Mixture Model (GMM) with a transformer model is proposed. Our results show that training transformer to learn GMM parameters could produce plausible multi-modal trajectories. Moreover, the model could infer uncertainty related to the predictions based on GMM parameters. The addition of probability to the multimodal trajectory predictions is crucial in making informed and risk-bound decisions by an autonomous vehicle.

Date of Award	Apr 2023
Original language	American English
Supervisor	Majid Khonji (Supervisor)