基于监控视频的交通事故自动识别算法

An Algorithm for Automatic Traffic Incident Detection Based on Surveillance Video

为了充分发挥现有交通监控网络提供的有效信息,提高交通事故发生后的救援、责任认定和交通疏导效率,以交通监控视频网络为数据源,采用了精简的YOLOv3目标检测网络实现监控视频中车辆与行人的实时多目标检测。在多目标跟踪方面,为了提高deepSORT算法实时性和稳定性,提出了基于无向图的特征描述矩阵,通过对车辆类目标的二次特征提取进行deepSORT级联匹配。基于交通时空冲突判定设计了事故自动识别算法,对各交通参与者运动特征进行了甄别处理,实现了交通事故的自动识别和事故特征的自动化提取。通过对不同角度的监控视频进行了数据提取,人工标注了13000多张训练数据,完成了对精简YOLOv3网络的训练。采用50个典型交通事故案例对事故自动识别算法进行了测试。结果表明:精简YOLOv3网络能够实现对包括轿车、客车、行人等8种常见交通元素的识别平均精度超过80%;事故自动识别算法达到了80%的识别成功率和8%的虚警率。提取的事故信息不仅包含事故的时间和地点等简单信息,而且包含事故各方的速度、轨迹等高维运动特征和视频片段,对实现交通事故的远程快速处理和救援、缓解由于交通事故引起的拥堵压力等方面具有重要的意义。

In order to fully utilize the effective information provided by the existing traffic monitoring network and improve the efficiency of rescue,responsibility identification and traffic diversion after traffic accidents,a simplified YOLOv3 object detection network is used as the data source to achieve real-time multi-objective detection of vehicles and pedestrians in monitoring videos.In terms of multi-target tracking,in order to improve the real-time performance and stability of the deepSORT algorithm,a feature description matrix based on undirected graphs is proposed,which performs deepSORT cascade matching by extracting secondary features of vehicle class targets.A traffic automatic incident detection algorithm is designed based on the judgment of traffic spatio-temporal conflicts,which screen and process the motion characteristics of each traffic participant,achieving automatic recognition of traffic accidents and automatic extraction of accident features.The result shows that(1)by extracting data from over 50 different angles of surveillance videos and manually annotating over 13000 training data,the simplified YOLOv3 network can achieve an average recognition accuracy of over 80%for 8 common traffic elements,including cars,buses,pedestrians,etc;(2)automatic accident extraction testing is conducted on 50 typical traffic accident cases,achieving an 80%recognition success rate and an 8%false alarm rate.The extracted accident information includes not only simple information such as the time and location of the accident,but also high-dimensional motion characteristics such as the speed and trajectory of all parties involved in the accident,as well as video clips.It is of great significance for achieving remote and rapid processing and rescue of traffic accidents,alleviating congestion pressure caused by traffic accidents and other aspects.