A Novel Multimodal Behavior Prediction Method for Automated Vehicles

Mozhgan Nasr Azadani, Azzedine Boukerche

Abstract: Ensuring the safe and effective path planning of automated vehicles in uncertain conditions hinges on the precise and dependable anticipation of future movements of nearby vehicles, coupled with a comprehensive grasp of the surrounding environment. This difficulty escalates significantly in dynamic and complex scenarios, such as unsignalized intersections, where there are no traffic lights to regulate vehicle behavior, and where multiple lanes are not available to infer drivers intentions based on their chosen lane. In this work, we design a novel method based on deep learning to anticipate vehicle behaviors at unsignalized intersections. Our approach accounts for the inherent uncertainty and multimodality of vehicle behavior by generating multiple potential outcomes. Our introduced model combines temporal convolutions with a mixture density layer to achieve this. We further cluster the obtained potential modes into feasible maneuvers, ranking them according to their probabilities. To evaluate the performance of our multimodal behavior prediction model, we conducted comprehensive evaluations using a substantial real-world dataset comprising over 23000 trajectories. The evaluation results underscore the better performance of our behavior prediction approach when compared to various baseline and state-of-the-art models.