Vehicle trajectory prediction based on spatio-temporal Transformer feature fusion
In complex traffic environments,autonomous vehicles must thoroughly analyze the motion direction,speed,and other information of surrounding traffic objects to accurately predict future trajectories.A network model based on spatio-temporal Transformer was proposed to address this issue.The framework initially employs a spatial self-attention mechanism to capture the spatial interactions between vehicles at the same moment,achieving precise modeling of the spatial relationship interactivity among multiple vehicles.Subsequently,a temporal self-attention mechanism was uti-lized to extract the temporal dependencies between consecutive frames,thereby generating a set of spatiotemporal fea-tures that reflect the dynamic behavior of vehicles.These features were then fed into a decoder to predict the motion tra-jectories of vehicles over the next 5 s.The proposed model was trained and validated on the publicly available NGSIM dataset.Compared to other state-of-the-art schemes,our scheme demonstrates greater accuracy and precision in trajec-tory prediction over the subsequent 5 s.The long-term forecasting accuracy is increased by 14.6%compared to the ad-vanced schemes.
万方数据
维普
