Cooperative Probabilistic Trajectory Forecasting Under Occlusion

Anshul Nayak, Azim Eskandarian

Keywords: Cooperative Perception, Trajectory Prediction, Probabilistic Inference

TL;DR: A cooperative and reliable way of trajectory forecasting with safety guarantees under occlusion

Abstract: Perception and planning under occlusion is necessary for safety-critical tasks. Cooperative planning requires communicating the information of occluded object to the ego agent. However, communicating rich information across multiple agents under adverse conditions and limited bandwidth may not be always feasible. Relative pose estimation between interacting agents sharing a common field of view can be a computationally effective way of communicating location of occluded objects. In this study, we use cooperative perception to reliably estimate the current states in the reference frame of ego agent and then predict the trajectory of the occluded pedestrian. Experimentally, we show that the uncertainty-inclusive predicted trajectory by ego agent using vision-based relative pose estimation is almost similar to the ground truth trajectory predicted by the ego agent assuming no occlusion. The current research holds promise for uncertainty-aware navigation among multiple interacting agents under occlusion.

Submission Number: 10