Kernelized Convolutional and Transformer Based Hierarchical Spatio-temporal Attention Network for Autonomous Vehicle Trajectory Prediction

In autonomous driving, predicting the future trajectories of adjacent traffic participants is critical for safe motion planning, navigating through challenging traffic environments, and improving safety. This work introduces a novel spatio-temporal attention-based model to predict the long-term trajectory and behaviour of a target vehicle over a long time horizon. The model employs a multi-head attention mechanism inspired by human reasoning, a Temporal Convolution Network (TCN) to establish temporal correlation in terms of motion state, and a kernel technique to construct a highly non-linear correlation for each vehicle in the driving scenario. Finally, a Transformer-based decoder is utilized to predict the target vehicle's long-term trajectory using an encoded spatio-temporal environment. The experimental findings using the NGSIM dataset demonstrate that the suggested model performs better than the leading spatio-temporal networks, achieving a 12% reduction in Root Mean Square Error (RMSE) when predicting a trajectory lasting 5 s. Additionally, the study analyzes the effectiveness of the proposed model by evaluating the effect of driving behaviour.