A Hybrid Trajectory Prediction Framework for Automated Vehicles With Attention Mechanisms

The driving safety of automated vehicles is largely dependent on accurately predicting the motions of surrounding vehicles. However, the existing approaches invariably neglect the impact of the ego vehicle's future behaviors on the surrounding vehicles and lack model explainability for the prediction results. To tackle these issues, a hybrid trajectory prediction framework based on long short-term memory (LSTM) encoding is proposed. It introduces a reactive social convolution structure to model the planned trajectory of the ego vehicle with the historical trajectories of the surrounding vehicles to reduce uncertainty in potential trajectories. Furthermore, a spatio-temporal attention mechanism is presented to quantitatively describe the contributions of historical trajectories and interactions among the surrounding vehicles to the prediction results by appropriate weights setting. Finally, the proposed scheme is comprehensively evaluated based on the NGSIM and HighD datasets. The results demonstrate that the proposed approach can elucidate the prediction process from a spatio-temporal perspective and outperform other state-of-the-art methods under different traffic scenarios. The root-mean-square errors on the NGSIM and HighD datasets are reduced to less than 3.65 m and 2.36 m over a time horizon of 5 s, respectively. The qualitative analysis on the reliability and reactivity is also presented.