부산대학교 도서관

Vehicle trajectory prediction plays a vital role in intelligent driving modules and helps intelligent vehicles travel safely and efficiently in complex traffic environments. Several learning-based prediction methods have been developed that accurately identify vehicle behaviour patterns in actual driving data. However, these methods rely on manually curated structured data and are difficult to deploy in intelligent vehicles. In addition, modular information channels that perform vehicle detection, tracking, and prediction tasks encounter error propagation issues and insufficient computing resources. Therefore, this paper proposes a new multitask parallel joint framework in which vehicle detection, state assessment, tracking, and trajectory prediction are performed simultaneously according to raw LIDAR data. Specifically, a multiscale bird's eye view (BEV) backbone feature extraction model is proposed and combined with the designed vehicle state identification branch to distinguish dynamic and static vehicles, which is used as a strong prior for trajectory prediction. In addition, a spatiotemporal pyramid model with convolutions and a backbone residual network is used to generate high definition (HD) maps with strong constraints and guidance capabilities, thereby improving the trajectory prediction accuracy. The experimental results on the real-world dataset nuScenes show that the proposed multitask joint framework outperforms state-of-the-art vehicle detection and prediction schemes, including ES3D and PnPNet.