A Convolutional Transformer Network for Vehicle Trajectory Prediction in Urban Traffic Scenarios

Motion optimization for intelligent connected vehicles (ICVs) requires predicting the trajectories of surrounding vehicles to achieve obstacle avoidance, which can be facilitated by the Intelligent Vehicle Cyber-Physical System (IVCPS). IVCPS builds digital twins of real-time traffic scenarios in the dynamic occupancy grid map (DOGM) format to support ICV trajectory prediction. To this end, this study integrates convolutional neural networks and the transformer to develop the convolutional multitarget trajectory transformer (CMTT). The CMTT utilizes parallel convolution blocks with prime-size kernels to extract spatial features from input DOGMs and generates independent feature map sequences through progressive receptive field expansion. It then employs an encoder–decoder structure to fuse and translate them into trajectory predictions with the guidance of preaccessible information. To comply with IVCPS specifications, this study combines SUMO and CARLA to construct a joint simulation platform that utilizes OpenStreetMap and real-world trajectory data from the pNEUMA and CitySim datasets to generate DOGM-format digital twins for CMTT training and validation. Experimental results indicate that CMTT surpasses all advanced baseline models in various urban traffic scenarios, showcasing its exceptional predictive performance and adaptability. Subsequently, ablation studies demonstrate the contribution of key components of CMTT to enhance model performance. Furthermore, this study determines the most suitable parameter configuration and pruning strategy for CMTT through parameter analysis and computational complexity analysis, achieving the optimal balance between prediction accuracy and computational efficiency. This study proposes CMTT, which demonstrates high prediction accuracy and computational efficiency, indicating potential for optimizing motion planning in real-world ICV applications.