A Probabilistic Model for Long-Term Trajectory Prediction in Multi-Agent Systems

The accurate prediction of vehicle trajectories in complex traffic environments is essential for ensuring the safety and effectiveness of autonomous driving systems. In this paper, we propose a novel probabilistic model for long-term trajectory prediction that consists of two components: a single-agent probabilistic model and a multi-agent risk-averse sampling algorithm. The single-agent probabilistic model is based on a dynamic Bayesian network, which considers the driver's maneuvering decisions and integrates surrounding lane information. In the multi-agent risk-averse sampling algorithm, feasible future positions are sampled simultaneously for all agents based on the probabilistic model, and a risk potential field model is then applied to reject the high-collision-risk samples. Eventually, a probability distribution of the combinations of long-term trajectories is predicted. After conducting experiments on the nuScenes dataset, our method achieved competitive performance in trajectory prediction compared with other state-of-the-art methods.

I. Introduction

Accurate prediction of a vehicle's future trajectory is crucial for autonomous driving systems, helping to make more accurate decisions while driving, avoid traffic accidents, create a smoother traffic environment, and make autonomous driving systems safe, stable, and efficient[1]. When predicting trajectories, the current physical state (such as speed, acceleration, heading angle, etc.) of the vehicle directly determines the trajectory in the near future[2]. As time goes by, the influence of the current physical state on trajectories will decrease, and the number of feasible trajectories will grow exponentially. The trajectory of the distant future is mainly determined by two factors: 1) Deterministic traffic environment information, such as static information like lanes and dynamic information like traffic lights; 2) Non-deterministic traffic participants, such as other vehicles, pedestrians, etc. The trajectory prediction problem falls into two categories: short-term prediction (a prediction time of less than 1 second) and long-term prediction (a prediction time of more than 2 seconds).