Advancing Vehicle Trajectory Prediction: A Probabilistic Approach Using Combined Sequential Models

Abstract In the domain of autonomous vehicle control, the foundation often lies in vehicle dynamics models, which simplify the complex behavior of vehicles to design effective controllers. However, these simplified models can introduce biases, particularly in extreme driving scenarios. Additionally, environmental factors introduce uncertainty into vehicle trajectories. Even with identical control actions, vehicles can display varying behaviors in different scenarios. This paper proposes methods for predicting the probabilistic distribution of a vehicle’s trajectory based on initial conditions and control action sequences. We leverage sequential learning and bias learning techniques to enhance learning efficiency. Our systematic study compares the Seq2Seq model, the particle-based model, and the hybrid model regarding prediction accuracy. The performance of these methods is evaluated and compared using both point error metrics and probabilistic prediction metrics. This research introduces novel approaches to probabilistic modeling of vehicle trajectories by combining physics-based and data-driven models, which can further contribute to the development of more robust and adaptable autonomous vehicle control systems that can navigate complex, uncertain, and extreme driving scenarios.