基于航拍视频构建风险指数的交织区拥堵识别方法

Congestion recognition method in weaving sections by constructing risk index based on aerial video

为了实时识别快速路交织区拥堵瓶颈的形成及其诱发因素,基于无人机航拍视频构建车辆轨迹数据,提出一种融合交通流不稳定性分析的交织区拥堵识别方法。识别方法由车辆轨迹提取、扰动感知模型和拥堵风险指数构建3个阶段构成。首先,通过YOLOv4(You Only Look Once,Version 4)网络训练航拍小目标权重检测俯拍车辆,关联外观与运动特征以跟踪车辆轨迹,从而提取无人机航拍视频中的精细车辆轨迹。然后,通过提取车辆微观速度、变道、冲突信息建立车速扰动和变道交织扰动感知模型。最后,采用熵值法结合扰动信息与平均车速构建归一化的拥堵风险指数,根据交织流的拥堵风险指数识别拥堵。本文采集广州大桥数据进行案例分析与测试验证。研究结果表明:学习了小目标特征的网络在航拍场景测试的误检率和少检率均低于5%,所提取的车辆轨迹连续稳定;在交织区拥堵识别评价中,本文方法的F1值达到97.85%,明显优于基本参数识别方法,在各路段中具有较高的识别准确度和算法鲁棒性;相比平均速度指标,所提出的拥堵风险指数能够更精细灵敏地反映短时和局部的拥堵,并能够从平均车速、个体车速差异和变道交织3个维度中识别多种因素引起的交织区交通瓶颈。研究结果可为城市重点路段交通诱导与优化提供技术基础。

In order to identify the formation of congestion bottlenecks and its inducing factors in expressway weaving sections in real time, based on vehicle trajectory data detected by UAV aerial video, a method of congestion recognition in weaving sections combined with the traffic flow instability analysis was proposed in this paper. The method consisted of three stages which are vehicle trajectory extraction, disturbance perception model and congestion risk index construction. First, the weight of aerial small object was trained by YOLOv4(You Only Look Once, Version 4) network for detecting vehicles from top view. The appearance and motion features were associated to track vehicle trajectories. So accurate vehicle trajectories in UAV(Unmanned Aerial Vehicle) video were extracted. Then, by extracting vehicle micro information including speeds, lane-changings and conflicts, the perception models for disturbance caused by vehicle speed and disturbance caused by lanechangings and weavings were established. Finally, the normalized congestion risk index was constructed by entropy method, combining disturbance information with average vehicle speed, to identify congestion according to risk index in weaving flows. In this paper, Guangzhou Bridge data were collected for case analysis and verification. The results show that: False rate and miss rate of the network with small objects learning features are both lower than 5% in the aerial scene test, and the extracted vehicle trajectories are continuous and stable. F1value of the method reaches 97.85% in congestion recognition for weaving sections, which is obviously higher than basic parameter methods and shows high recognition accuracy and algorithm robustness in each road section. Compared with average speed index, the proposed congestion risk index can more accurately and sensitively reflect short-term and local congestion. The index can also identify traffic bottlenecks in weaving sections caused by multiple factors from three parameter dimensions including average speeds, speed differences,lane-changings and weavings. The study results can provide a technical basis for traffic guidance and optimization in urban key road section.