A Cooperative Vehicle Localization and Trajectory Prediction Framework Based on Belief Propagation and Transformer Model
Feiyu Jin, Kai Liu, Chunhui Liu, Tongtong Cheng, Hao Zhang, Victor C. S. Lee
Abstract
The advancement of sensing, transmission, and computation technologies has transformed modern vehicles into ubiquitous consumer electronics. This paper presents a cooperative vehicle localization and trajectory prediction framework for enabling Intelligent Transportation Systems (ITS) applications. Specifically, the proposed framework consists of a Belief Propagation based Location Approximation (BPLA) algorithm for cooperative vehicle localization and a Transformer-based Vehicle trajectory prediction model called VFormer. The BPLA algorithm first establishes a factor graph based on the sensor measurements transmitted by vehicles, and then adopts a modified belief propagation procedure to approximate the posterior distribution of vehicles. On this basis, VFormer extracts the hidden features from historical positions estimated by BPLA and vehicle motion data to model long-term and short-term motion patterns of vehicles. Moreover, the multi-head attention layer in the VFormer is utilized to learn the spatial-temporal dependencies between the target vehicle and its surrounding vehicles at different timestamps to improve prediction accuracy. A comprehensive performance evaluation has been conducted based on the public vehicle trajectory dataset and the real-world system prototype. Experiment results demonstrate the superiority of the proposed framework on improving vehicle localization and trajectory prediction performance.