Transformer-Based Multi-Modal Probabilistic Pedestrian Prediction for Risk-Aware Autonomous Vehicle Navigation

Over the past decade, the field of assisted and autonomous driving has experienced significant advancements. However, autonomous driving systems are still challenged by the complexities of dynamic urban environments, especially when it comes to predicting and responding to the often stochastic behavior of pedestrians. Current approaches largely concentrate on most likely predictions but tend to ignore their inherent probabilistic nature. Our research introduces a novel Transformer-based multimodal probabilistic prediction model that utilizes a Gaussian Mixture Model (GMM). This approach is simpler than its predecessors, yet it maintains competitive performance, capable of inferring prediction uncertainties using GMM parameters. Additionally, we demonstrate how our prediction model can be incorporated into a riskaware behavior planner, based on the Chance-Constrained Stochastic Shortest Path (CC-SSP) framework. This planner uses probabilistic trajectory predictions as a Markov transition function to modulate the speed of the autonomous vehicle, effectively keeping the probability of collision below a defined threshold. Our implementation is available at https://github.com/Murdism/Probabilistic Pedestrian Trajectory Prediction-PPTP.git.