A Self-Supervised Approach for Cooperative Neighboring Vehicle Positioning System based on Spatial-Temporal Learning Techniques

Precise vehicle positioning is the key foundation for advancing vehicle automation technology beyond level three. However, the conventional global positioning system (GPS) is susceptible to inaccuracies caused by environmental interference. Existing works for improving positioning accuracy either require fundamental infrastructure modification to replace GPS or utilize prior knowledge of environmental information to reduce interfer-ence, where both are impractical in the real world. To improve the GPS-based vehicle positioning system to provide more accurate coordinate estimates without prior knowledge of environmental information, we propose a self-supervised learning architecture composed of four learning methods: hierarchical density-based spatial clustering (HDBSCAN), graph convolution network (GCN), domain-adversarial neural network (DANN), and long short-term memory (LSTM). The proposed framework utilizes both spatial and temporal information in vehicle positioning. The simulation results within our proposed comprehensive framework demonstrate a significant improvement in the accuracy of vehicle coordinate estimates, with the estimation error mean decreasing by 46 % and the error standard deviation decreasing by 34 % compared to the baseline.