Probabilistic Trajectory Prediction in Intelligent Driving

Abstract

For intelligent vehicles, trajectory prediction for their surrounding traffic participants is the basis of their ego trajectory planning. Since there exist various uncertain factors that influence the driving processes, the participants' future trajectories are uncertain either. Unique deterministic prediction results provided by traditional methods do not involve the probability information of different trajectories, and thus are insufficient. To solve this problem, a probabilistic trajectory prediction model is proposed here, and uncertain future trajectories are designed to be described with position probability distributions at discrete moments. In the proposed model, the predicted participant's motion is decomposed into lateral and longitudinal motions which are independent with each other. Probabilistic trajectories are calculated in these two directions respectively and then are combined into complete ones. Furthermore, the model considers two driving modes(Free driving and Vehicle-following modes) in the longitudinal motion, and thus computes trajectories in different ways. Besides, we also modeled the features of different types of participants and the interaction between their trajectories. Model parameters are identified off-line based on historical trajectory data, therefore only iterative computations are needed when the model is applied on-line. The experimental results show that the proposed method has higher prediction accuracy than the traditional ones, and can meet the demands of real-time applications.