The freeway accident detection and classification have attracted much attention of researchers in the past decades. With the popularity of Global Navigation Satellite System (GNSS) on mobile phones and onboard equipment, increasing amounts of real-time vehicle trajectory data can be obtained more and more easily, which provides a potential way to use the multi-vehicle trajectory data to detect and classify an accident on freeways. The data has the advantages of low cost, high penetration, high real-time performance, and being robust to the outdoor environment. Therefore, this paper proposes a new method for accident detection and classification based on the multi-vehicle trajectory data. Different from the existing methods using GNSS positioning data, the proposed method not only uses the position information of the related vehicles but also tries to capture the development tendencies of the trajectories of accident vehicles over a period of time. A Deep Convolutional Neural Network model is developed to recognize an accident from the normal driving of vehicles and also identify the type of the accident, and the six types of traffic accidents are considered in this study. To train and test the proposed model, the simulated trajectory data is generated based on PC-Crash, including the normal driving trajectories and the trajectories before, in, and after an accident. The results indicate that the detection accuracy of the proposed method can reach up to 100%, and the classification accuracy can reach up to 95%, which both outperform the existing methods using other data. In addition, to ensure the robustness of the detection accuracy, at least 1 s of duration and 5 Hz of frequency for the trajectory data should be adopted in practice. The study will help to accurately and fastly detect an accident, recognize the accident type, and future judge who is liable for the accident.

高速公路事故检测与分类在过去几十年中引起了研究人员的广泛关注。随着全球导航卫星系统（GNSS）在手机和车载设备上的普及，越来越多的实时车辆轨迹数据可以越来越容易地获得，这为利用多车辆轨迹数据进行检测提供了一种潜在的途径。并对高速公路上的事故进行分类。数据具有成本低、渗透率高、实时性高、对室外环境鲁棒性强等优点。因此，本文提出了一种基于多车辆轨迹数据的事故检测与分类新方法。与现有使用 GNSS 定位数据的方法不同，该方法不仅利用了相关车辆的位置信息，还试图捕捉事故车辆在一段时间内的轨迹发展趋势。开发了深度卷积神经网络模型以从车辆的正常驾驶中识别事故并识别事故类型，本研究考虑了六种类型的交通事故。为了训练和测试所提出的模型，基于 PC-Crash 生成模拟轨迹数据，包括正常驾驶轨迹和事故前、中、后的轨迹。结果表明，所提方法的检测准确率可达100%，分类准确率可达95%，均优于使用其他数据的现有方法。此外，为保证检测精度的鲁棒性，实际应用中轨迹数据的时长至少为1 s，频率为5 Hz。该研究将有助于准确快速地发现事故，识别事故类型，以及未来判断事故的责任人。