Vehicle Motion Prediction Using Locally Conditioned Trajectory Sets

Abstract

Vehicle motion prediction aims to regress a continuous future trajectory from past motion data for all agents in the scene. However, not all future trajectories are viable for each agent. Vehicle trajectories are often limited by factors such as current vehicle dynamics, an agent’s location in the scene, and legal traffic maneuvers. We look to use this idea and frame the vehicle motion prediction problem as a classification problem across a set of trajectories that are ensured to be viable. While other methods typically only consider vehicle dynamics, we look to also explicitly factor in lane geometries and traffic rules as constraints in our trajectory set generation. Although we are unable to achieve similar performance to SOTA, we show that even with minimal training data, this approach allows for generalization to completely new scenes with no retraining and minimal performance loss.