基于深度学习的车辆轨迹重建与异常轨迹识别

车辆移动轨迹的不确定性及异常点段的存在使其在数字交通领域的应用面临挑战。本文构建基于数据增强的LSTM-AE-Attention深度学习模型，进行车辆轨迹重建和异常轨迹识别。首 先，使用对抗生成网络和贝塞尔样条曲线从样本量和种类两方面扩充数据集，实现数据增强；其 次，通过自编码网络与长短时记忆神经网络提取轨迹特征并完成轨迹重建；最后，结合自编码网络预训练和注意力机制构建异常识别模型。采用实际车辆轨迹数据测试，模型的评价指标明显优于支持向量机、随机森林和长短时记忆神经网络模型，重建实验中模型的决定系数为0.968，异常识别实验中模型的F1值较对比模型平均提升9.8%。结果表明，本文提出的模型可有效、可靠地运用于平滑车辆轨迹数据和纠正异常车辆轨迹。

Vehicle Trajectory Reconstruction and Anomaly Detection Using Deep Learning

The uncertainty and the existence of outliers of vehicle trajectory data makes it challenging to be used in the digital transportation systems. In this paper, we propose the LSTM-AE-Attention deep learning model for vehicle trajectory reconstruction and anomalous trajectory detection. To realize the data augmentation, the trajectory dataset is expanded in the sample size and sample type using the Generative Adversarial Networks and the Bezier Curve. The trajectory features are extracted by the Autoencoder and the Long Short-term Memory networks, and the trajectory reconstruction is completed. An anomaly detection model is then developed using the AE Pre-training and the Attention networks. The experiment has been performed based on a practical trajectory dataset. The results show that the proposed model produces several evaluation indicators that are better than the Support Vector Machines, Random Forest and Long Short-term Memory networks results. In the trajectory reconstruction section, the decision coefficient of the model is 0.968. In the anomaly detection part, the F1 score is 9.8% higher than the average score of the other models. The LSTM- AE- Attention model is an effective method for smoothing the vehicle trajectory and correcting abnormal trajectory.