Abstract: Accurate prediction of road vehicle behavior is of great significance for improving road safety, optimizing traffic flow and realizing autonomous driving. The vehicle trajectory prediction model in this paper uses long-short term memory (LSTM) encoder and decoder. In order to further improve the prediction accuracy, 3D convolution is introduced into the model to better capture the interaction effects of dynamic changes between different vehicles, a 3D-LSTM model is established. Compared with the CS-LSTM model, the input speed and acceleration are added. The model can comprehensively consider the historical state of the target vehicle and the dynamic information of the surrounding vehicles, and realize the accurate prediction of the trajectory. In this paper, the NGSIM data set is used to evaluate the model. Compared with the CS-LSTM model, the root mean square error (RMSE) of the proposed model decreased by 26%, 15.5%, 12.6%, 11.4% and 10.2%, respectively, in 1 to 5 seconds. Simulation results show that the proposed model can significantly improve trajectory prediction accuracy.

Key words: vehicle trajectory prediction; driving intention recognition; deep learning; LSTM

摘要： 道路车辆行为的准确预测对于提升道路安全、优化交通流量以及实现自动驾驶具有重 要意义。文章提出的车辆轨迹预测模型使用长短时记忆（LSTM）编码器和解码器，为了进一 步提高预测精度，在模型中引入 3D 卷积，以更好地捕捉不同车辆间动态变化的交互影响，建 立了 3D-LSTM 模型。与 CS-LSTM 模型相比，在模型输入上增加了速度和加速度。该模型能 够综合考虑目标车辆历史状态以及周围车辆动态信息，实现对轨迹的准确预测。文章使用 NGSIM 数据集来对模型进行评估，模型的均方根误差（RMSE）与 CS-LSTM 模型相比，1 到 5 s 分别降低了 26%、15.5%、12.6%、11.4%和 10.2%，仿真结果表明，模型可以显著提高轨 迹预测精度。

关键词: 车辆轨迹预测；驾驶意图识别；深度学习；LSTM