面向群体行驶场景的时空信息融合车辆轨迹预测…

将车辆间时空交互信息融入卷积社会池化网络中，提出了一种面向群体行驶场景的有人驾驶车辆轨迹预测模型；使用长短时记忆(LSTM)网络预测群体车辆速度，基于此预测值计算群体车辆间的速度差；构造LSTM编码器捕捉群体车辆行驶轨迹的时间序列特征，设计卷积社会池化网络提取群体车辆间的空间依赖关系，使用LSTM解码器预测未来车辆各种动作的出现概率和相应轨迹，将具有最高出现概率的动作及其轨迹作为最终轨迹预测结果；使用真实轨迹数据集对所构建模型进行了参数标定和性能验证，测试了不同轨迹编解码与速度预测方法对模型性能的影响，确定了最优模型结构。计算结果表明:相较于历史速度，使用预测速度计算速度差作为模型输入可将均方根误差(RMSE)降低19.45%；相较于门控循环神经网络，使用LSTM进行速度预测可将RMSE降低4.91%；相较于原始卷积社会池化网络，所提出模型的轨迹预测误差在RMSE与负似然对数2个指标上分别降低了20.32%和21.04%，明显优于其他卷积社会池化网络变体；所提出模型与原始卷积社会池化网络计算耗时差距约3 ms，能够满足实时应用要求。 

Vehicle trajectory prediction based on spatio-temporal information fusion in crowded driving scenario

The spatio-temporal interaction information among vehicles was integrated into the convolutional social pooling network to formulate a human-driving vehicle trajectory prediction model in the crowded driving scenario. The long short term memory (LSTM) network was used to predict the speeds of the crowded vehicles. The prediction result was used to calculate the speed differences among the vehicles. The LSTM encoder was built to capture the time-series features of the crowded vehicle trajectories. The convolutional social pooling network was designed to captured the spatial dependence of the crowded vehicles. The emerging probabilities of all possible movements of the vehicles and corresponding trajectories were predicted by the LSTM decoder. The movement with the highest emerging probability and its trajectory were taken as final prediction result of trajectory. The real vehicle trajectory dataset was used in the parameter calibration and performance verification of the proposed model. Different methods of trajectory encoding/decoding and speed predicting were tested to figure out their influences on the model performance. The test results were used to identify the optimal model structure. Calculation results show that compared with historical speed, predicted speed used to calculate speed difference as model input can decrease by 19.45% in terms of root mean square error (RMSE). Compared with the gate recurrent unit, the LSTM network as speed predictor can decrease by 4.91% in terms of RMSE. Compared with the original convolutional social pooling network, the trajectory prediction errors of the proposed model respectively decrease by 20.32% and 21.04% in terms of RMSE and negative log-likelihood. The model performance is also significantly better than other variants of the original convolutional social pooling network. The computation time difference of the proposed model and original convolutional social pooling network is about 3 ms, which meets the request of real-time application. 8 tabs, 9 figs, 23 refs.