程凯

由于近些年来机动化进程不断加快，道路交通安全问题层出不穷，给国家带来了巨大的生命、财产损失，因此对交通风险的分析和交通事故的预防十分必要。考虑到车辆轨迹信息涵义的丰富性，本研究基于包含车辆轨迹信息的无人机航拍视频对交通风险展开研究。传统的事故率方法被广泛地用于交通风险研究中，但其存在数据样本难以获取等不足，而交通冲突技术的理论和仿真的手段可以弥补事故率方法的缺陷。因此对道路交通风险进行研究，可以通过车辆轨迹提取及其特征分析、基于轨迹的交通冲突识别、基于仿真的交通冲突分析来实现。

首先，本研究采用无人机及其配套设备在研究路段上进行了4k级高分辨率视频的自主拍摄，针对视频图像中的车辆进行了目标检测与追踪方法的应用。利用MATLAB编写改进相关图像处理算法步骤，实现了车辆的边缘检测识别、追踪器训练、坐标转换、数据质量控制、阈值参数校正及结果输出功能。通过对输出的轨迹结果进行小波变换处理得到最终去除白噪声的轨迹文件及更新后的车辆检测文件。基于George 2.1提取出的人工轨迹数据，进行了轨迹精度验证，最终该轨迹提取及小波去噪的方法，能够应用于无人机航拍视频的高精度车辆轨迹提取，为交通风险研究提供有效的支撑。

然后，本研究利用提取到的车辆轨迹进行交通流特征分析和交通冲突分析。车流组成规律、车辆行驶特性以及换道行为特征为后续交通冲突分析及其验证提供理论背景，同时为仿真模型建立提供数据。为了更好地描述冲突的风险水平，本研究基于交通冲突技术构建了一种衡量冲突风险的指标PCC，使其在适用于二维空间冲突识别时，充分考虑到边界到达风险、持续暴露风险和两车驾驶状态风险，赋予交通冲突技术更丰富的内涵。经过与传统TTC在德国inD数据集中的对比分析，验证了本研究中指标的有效性，以及在追尾和交叉场景中的优越性。另外，利用该指标对不同主体和不同类型下的交通冲突进行了特性分析，得到了快速路中有入口匝道汇入主线且有货车时的交通冲突分布规律。

最后，本研究利用VISSIM仿真平台搭建研究路段的微观仿真模型，并将前面章节中统计得到的交通流信息作为基本参数输入到模型中。另外，针对驾驶行为中的跟车模型参数和换道模型参数进行了研究，通过正交实验设计出驾驶行为参数的校正方案，选取了与行程时间和平均车速相关的评价指标后，得到最优调参方案。经过仿真可以得到车辆运行的轨迹文件，将轨迹文件作为替代安全评价模型SSAM的输入，最终输出基于仿真的交通冲突识别结果。根据冲突识别结果，得到了不同主体、不同类型下冲突的特征规律，验证了前面章节中的分析结论。同时，通过仿真得到的结果可以应用于冲突数据样本的扩展，而仿真的手段也可以为交通安全的改善和治理提供理论基础。

关键词：无人机视频，轨迹提取，交通冲突分析，冲突风险，仿真

As the process of motorization has been accelerating in recent years, road traffic safety problems have emerged one after another, which has brought huge losses of life and property to the country. Therefore, the analysis of traffic risks and the prevention of traffic accidents are very necessary. Taking into account the richness of the meaning of vehicle trajectory information, this research is based on the UAV aerial video containing vehicle trajectory information to study the traffic risk. The traditional accident rate method is widely used in traffic risk research, but it has disadvantages such as the difficulty of obtaining data samples. The theory and simulation methods of traffic conflict technology can overcome the shortcomings of the accident rate method. Therefore, the research on road traffic risk can be realized by vehicle trajectory extraction and feature analysis, trajectory-based traffic conflict identification, and simulation-based traffic conflict analysis.

First of all, this research uses unmanned aerial vehicle and its ancillary equipment to complete 4k-level high-resolution video’s autonomous shooting on the research road, and applies the target detection and tracking method to the vehicle in the video image. Using MATLAB to compile and improve related image processing algorithm steps, realize the functions of vehicle edge detection and recognition, tracker training, coordinate conversion, data quality control, threshold parameter correction and result output. By performing wavelet transform processing on the output trajectory results, the trajectory file’s noise is removed and the updated vehicle detection file are obtained. Based on the artificial trajectory data extracted by George 2.1, the accuracy of the trajectory was verified. Therefore, the method of trajectory extraction and wavelet denoising can be applied to high-precision vehicle trajectory extraction from drone aerial video, so as to provide effective support for traffic risk research.

Then, this research uses the extracted vehicle trajectories to analyze traffic flow characteristics and traffic conflict analysis. The composition of traffic flow, vehicle driving characteristics and lane-changing behavior characteristics provide a theoretical background for subsequent traffic conflict analysis and verification, as well as data for the establishment of simulation models. In order to better describe the risk level of conflicts, this study builds an indicator PCC to measure conflict risk based on traditional traffic conflict technology, so that when it is applied to two-dimensional space conflict identification, it can also fully consider the risk of boundary arrival and the risk of continuous exposure, and the two-vehicle driving state risk, giving the traffic conflict technology richer connotations. After a comparative analysis with the traditional TTC in the German inD data set, the validity of the indicators in this study and their superiority in rear-end and cross scenarios are verified. In addition, this indicator is used to analyze the characteristics of traffic conflicts under different subjects and types, and obtains the distribution characteristics of traffic conflicts when there is an on-ramp into the main line on an expressway and there are trucks.

Finally, this study uses the VISSIM simulation platform to build a microscopic simulation model of the research section, and inputs the traffic flow information obtained by statistics in the previous chapters as basic parameters into the model. In addition, the car-following model parameters and lane-changing model parameters in driving behavior are studied, and a correction scheme for driving behavior parameters is designed through orthogonal experiments. After selecting evaluation indicators related to travel time and average vehicle speed, the optimal scheme is obtained. Through simulation, the trajectory file of the vehicle operation can be obtained, and the trajectory file is used as the input of the surrogate safety assessment model SSAM, finally the simulation-based traffic conflict recognition result is output. According to the results of conflict recognition, the characteristics of conflicts under different subjects and different types are obtained, which verifies the analysis conclusions in the previous chapters. At the same time, the results obtained through simulation can be applied to the expansion of conflicting data samples, and the means of simulation can also provide a theoretical basis for the improvement and governance of traffic safety.

Key Words: UAV video, trajectory extraction, traffic conflict analysis, conflict risk, simulation