SW-SAN:基于Seq2Seq结合注意力机制与滑动窗口的车辆轨迹预测模型

SW-SAN: Vehicle trajectory prediction model based on Seq2Seq combined with attention mechanism and sliding window

针对长时间内4~5 s车辆轨迹预测精度较差的问题,提出基于Seq2Seq结合注意力机制与滑动窗口的车辆轨迹预测模型(SW-SAN)。首先,使用滑动窗口的方法更新历史轨迹状态集合,利用编码器对目标车辆的历史轨迹数据编码,得到历史轨迹特征向量;其次,经过注意力机制计算历史时间内各时刻的关联性得分、时间注意力权重因子和历史时间相关性特征向量;最后,解码器将历史时间相关性特征向量作为输入,多次循环解码层,输出目标车辆的未来预测轨迹。实验结果表明,SW-SAN模型在4 s和5 s时预测轨迹的RMSE误差为1.99 m和1.94 m,SW-SAN模型在较长时间4~5 s的预测误差更低,在车辆轨迹预测问题上性能更强。

A vehicle trajectory prediction model(SW-SAN)based on Seq2Seq combined with attention mechanism and sliding window is proposed to address the issue of poor accuracy in predicting vehicle trajectories over a long period of time(4~5s).The sliding window method is used to update the historical trajectory state set,and an encoder is used to encode the historical trajectory data of the vehicle(the object)to obtain the historical trajectory feature vector.The attention mechanism is used to calculate the relevant scores,time attention weighting factors,and historical time related feature vectors for each moment in the historical time.The decoder is used to take the historical time correlation feature vector as the input,and the decoding layer is cycled for multiple times to output the future predicted trajectory of the vehicle(the object).The experimental results show that the root mean square error(RMSE)of the SW-SAN model in predicting trajectories at 4s and 5s are 1.99 m and 1.94 m,respectively,so the prediction error of SW-SAN model is relatively lower over a longer period of time(4~5s),and its performance in vehicle trajectory prediction is better.