Non-Autoregressive Vehicle Trajectory Prediction Based on Spatial-Temporal Trend-Aware Attention

Accurate vehicle trajectory prediction is of paramount importance for enhancing the stability and safety of intelligent transportation systems. Despite extensive research efforts, several challenges persist, including extracting spatial-temporal features from trajectory data, identifying driving intentions and making longer-term predictions. In response to the problem of only utilizing the last-moment features during spatial information interaction, less integrating driving intention and mostly using autoregressive methods for short-term trajectory prediction, this paper introduces a novel non-autoregressive spatial-temporal trend-aware attention model (NASTTA) organized in a hierarchical manner to address the issues. Experimental results on the NGSIM data set demonstrate that compared to the current state-of-the-art algorithms, the performance of the proposed model is comparable in short-term prediction but exhibits more noticeable improvements in long-term prediction, up to 50%. Furthermore, ablation experiments are conducted to evaluate the performance of each module, revealing that the absence of any module results in a decrease in prediction performance.