Abstract:

Most of the existing vehicle trajectory predictions are single-target trajectory predictions, without two-way interaction and relational reasoning, and interactive modeling of mixed entities cannot be achieved. To solve the above problems, a fully scalable trajectory prediction model called Q-LSTM is designed by combining the Q-learning algorithm of reinforcement learning and the LSTM network of deep learning. In Q-LSTM model, the LSTM network captures the time characteristics of the vehicle trajectory, and the Q-learning algorithm represents the interaction process of multiple vehicles. Therefore, the Q-LSTM model can realize multi-interaction modeling of a random number of vehicles, and the accuracy of predict the long-term interactive vehicle trajectory is guaranteed. In addition, the relationship between vehicle length and width and coordinates is considered in the model to avoid abnormal collision phenomena, and it is suitable for scenarios of multi-type vehicle trajectory prediction. The performance analysis experiment of the model is carried out on the public data set HighD. The experimenal result proves that the Q-LSTM model has certain advantages in terms of long-term interactive vehicle trajectory prediction accuracy and reduction of collision phenomena.

Key words: intelligent transportation, trajectory prediction, long short term memory, reinforcement learning

摘要：

现有的车辆轨迹预测大多是单目标轨迹预测，无双向交互和关系推理，不能实现混合实体的交互建模。针对上述问题，结合强化学习的Q-learning算法和深度学习的LSTM网络，设计一个完全可扩展的轨迹预测模型Q-LSTM。该模型中，LSTM网络捕获了车辆轨迹的时间特性，而Q-learning算法则表示了多车辆的交互过程，因此Q-LSTM模型可以实现随机数量车辆多交互建模，并且在长期交互车辆轨迹预测中保证精确度。另外模型中考虑了车辆长宽与坐标之间的关系，避免出现异常的碰撞现象，适合用于多类型车辆轨迹预测的场景。在公开数据集HighD上进行了模型的性能分析实验，实验结果证明Q-LSTM模型在较长期交互车辆轨迹预测精度和减少碰撞现象等方面具有一定优势。

关键词: 智慧交通, 轨迹预测, 长短记忆模型, 强化学习