基于时-空注意力机制的车辆轨迹预测

城市交通环境中车辆的驾驶行为随机性较高，且驾驶人驾驶风格迥异。为了解决复杂交通环境下车辆行驶轨迹难以精确预测的问题，在社会生成对抗网络(Social GAN)的基础上，考虑车辆的行驶速度、加速度、航向角等行驶状态参数和形状尺寸，建立车辆间交互影响力场模型，提出一种基于时-空注意力机制的车辆轨迹预测算法(SIA-GAN)。根据受到场景中其他车辆交互影响力的大小赋予其他车辆不同的空间注意力权重因子，重点关注对自车行驶影响较大的车辆信息，并结合时间注意力机制挖掘自身车辆对观测时段内历史轨迹特征向量的时间依赖性，得到车辆预测轨迹。为验证所提算法的有效性，在开源数据集上对算法进行迭代训练，并与LSTM、Social LSTM、Social GAN三种轨迹预测算法进行对比分析。研究结果表明：SIA-GAN不仅在训练时的收敛速度上有较大提升，且与现有其他轨迹预测算法相比在平均位移误差、最终位移误差、平均速度误差、平均航向角误差等评价指标均有大幅下降，预测3.2 s时各项指标平均降低了51.25%、60.1%、37.84%、13.75%;预测4.8 s时各项指标平均降低了52.78%、61.47%、3...

Vehicle Trajectory Prediction Based on Spatial-temporal Attention Mechanism

Vehicle movements can be highly random, and the driving style used becomes complex in urban traffic environments. To overcome the difficulties in accurately predicting vehicle trajectory in complex traffic environments, the Social Generation Adversarial Network (Social GAN) machine-learning model was used to develop a vehicle trajectory prediction algorithm named SIA-GAN. This developed algorithm was based on a spatial-temporal attention mechanism by considering a vehicle's speed, acceleration, course angle driving state, and shape size, and an interaction influence force field between the different vehicles was derived. Based on the magnitude of the interaction influence force that characterized each vehicle at the scene, different spatial attention weighing factors were assigned to the vehicles, along with a component of stressed "attention" that incorporated the information of vehicles having a greater impact on each other's driving pattern. The time attention mechanism was then combined to mine the time dependence of the vehicle under consideration in terms of the trajectory's feature vector during the observation period. To verify its effectiveness, the proposed algorithm was iteratively trained on an open-source dataset and compared with three trajectory prediction algorithms (long short-term memory (LSTM), Social LSTM, and Social GAN). The results show that SIA-GAN not only improves the convergence speed during training but also significantly reduces the average displacement error (ADE), final displacement error (FDE), average velocity error (AVE), and average course angle error (ACAE) when compared with other existing algorithms for trajectory prediction. At predicts 3.2 s, each of the aforementioned indexes decreases by 51.25%, 60.1%, 37.84%, and 13.75%, respectively, on average. The average reduction at predicts 4.8 s is 52.78%, 61.47%, 35.92%, and 9.57%, respectively. Thus, the proposed SIA-GAN trajectory prediction algorithm can accurately and effectively reflect complex spatial interaction characteristics between vehicles, enhancing the accuracy, rationality, and interpretability of trajectory predictions.