高速公路车辆轨迹跟踪是指利用视频监控网络中的多个摄像头，对车辆在不同监控范围内的行驶轨迹进行跟踪。该技术可以为交通管理部门提供实时、准确的车辆位置和行驶轨迹信息，是车辆特征分析、驾驶行为提取的技术基础，也是提升道路交通安全、建设智慧城市的重要环节。论文聚焦于高速公路车辆轨迹跟踪这一应用需求，从实际场景出发，对车辆目标检测、目标跟踪、交通状态分析、车辆重识别等相关技术方法进行集成创新，提出了一种基于交通状态分析与深度学习计算机视觉的特定车辆轨迹跟踪方法，在对比车辆图像相似性的基础上充分利用摄像头时空信息，提升了车辆重识别的准确性，并在G6002贵阳绕城高速的实际应用场景中验证了该方法可行性。论文相关内容及创新性研究成果如下：（1）面向G6002贵阳高速公路的实际道路交通场景下的交通流表征参数及车辆行驶特征提取问题，基于高速公路视频监控系统所得的视频流数据，采用车辆目标检测技术，获得了贵阳高速公路交通流表征参数；采用AFCM聚类方法对以上数据进行聚类，生成了贵阳高速公路交通状态判别标准。基于该标准，结合实际场景与车辆所在路段交通状态，提出了一种目标车辆位置预测方法，从而获得了目标车辆位置置信度。（2）针对车辆重识别问题，基于“全局+局部”特征的联合学习思路，提出了一种多特征联合学习的车辆特征提取方法，从而获得了目标车辆图像置信度。该方法基于一个双分支深度学习模型，其中局部特征分支通过引入水平平均池化层对目标车辆图片前挡风玻璃处的粘贴标志、车辆内饰、驾驶员图像等个性化局部特征进行提取；全局特征分支则用于提取车辆的全局特征，包括车辆轮廓、车头、车灯等。我们对该模型在PKU-VD数据集的效果进行了检验，证明该模型相比传统人工特征模型及深度学习模型，均体现出了更好的性能。（3）针对高速公路目标车辆轨迹跟踪问题，基于前文提出的目标车辆位置预测方法及多特征联合的车辆重识别模型，提出了一种结合车辆位置置信度与车辆图像置信度的跨镜头车辆轨迹跟踪方法，通过车辆链式追踪与自动递进容错的错漏检纠正方法，实现了多摄像头内目标车辆的轨迹跟踪回放。最终，我们开发了高速公路交通状态判别及特定车辆轨迹跟踪系统，并在贵阳绕城高速的实际场景中验证了该系统的可用性。
Highway vehicle trajectory tracking refers to tracking the driving trajectories of vehicles in different monitoring ranges using multiple cameras in a video surveillance network. This technology can provide real-time and accurate vehicle location and driving trajectory information for traffic management departments. It is the technical basis for vehicle feature analysis and driving behavior extraction, and an important part of improving road traffic safety and building smart cities. This thesis focuses on the application requirements of highway vehicle trajectory tracking, and integrates and innovates related technologies such as vehicle target detection, target tracking, traffic state analysis, and vehicle re-identification from the perspective of practical scenarios. A specific vehicle trajectory tracking method based on traffic state analysis and deep learning computer vision is proposed, which fully utilizes the spatiotemporal information of cameras and improves the accuracy of vehicle re-identification based on comparing vehicle image similarity. The feasibility of the proposed method is verified in the actual application scenario of the G6002 Guiyang Ring Expressway. The relevant content and innovative research achievements of the thesis are as follows: (1) This thesis addresses the problem of extracting traffic flow characterization parameters and vehicle driving features in actual road traffic scenes on the G6002 Guiyang Ring Expressway. Based on the video stream data obtained from the highway video surveillance system, the thesis employs vehicle target detection technology to obtain traffic flow characterization parameters on the G6002 Guiyang Ring Expressway. By using the AFCM clustering method to cluster the data, a traffic state discrimination standard for the expressway is generated. Based on this standard, the thesis proposes a target vehicle position prediction method that combines the actual scenario with the traffic state of the road section where the vehicle is located, thus obtaining the confidence level of the target vehicle‘s position. (2) To address the vehicle re-identification problem, this thesis proposes a multi-feature joint learning vehicle feature extraction method based on the "global+local" feature joint learning approach. This method obtains the confidence level of the target vehicle image. The method is based on a dual-branch deep learning model. The local feature branch extracts personalized local features such as pasted signs on the front windshield, vehicle interiors, and driver images by introducing horizontal average pooling and dimension reduction convolution layers. The global feature branch is used to extract the vehicle‘s global features, including the vehicle contour, headlights, etc. We tested the model‘s performance on the PKU-VD dataset and found that the model exhibited better performance compared to traditional manual feature models and deep learning models. (3) To address the problem of target vehicle trajectory tracking on the expressway, this thesis proposes a cross-camera vehicle trajectory tracking method that combines the confidence level of the vehicle position with the confidence level of the vehicle image based on the target vehicle position prediction method and the vehicle re-identification model proposed earlier. By designing a fault-tolerant progressive tracking and forward-redundant matching, the method achieves trajectory tracking and playback of the target vehicle in multiple cameras. Finally, we developed a traffic state discrimination and specific vehicle trajectory tracking system for the expressway and validated its effectiveness in actual scenarios on the Guiyang Ring Expressway.
