基于YOLO_V3的侧视视频交通流量统计方法与验证

A YOLO_V3-based Road-side Video Traffic Volume Counting Method and Verification

为了研究因受限于观测点位的临时性和不确定性而导致自动化技术手段无法在临时交通观测中实用的难题,提出了一种适用于侧视角度拍摄视频、可快速识别车辆并实现交通流量统计的方法,该方法克服了传统视频识别技术无法满足侧视视角交通视频识别的困难。采用基于深度学习的YOLO_V3方法,以临时观测的路侧采集视频为对象进行车辆检测,提出基于车辆检测区域和流量计数区域的二级目标物检测框架,建立卡尔曼滤波+匈牙利分配+透视投影变换的交通流量计数模式,实现车辆的快速和高精度追踪。采集多组实际视频数据,从拍摄相机与道路相交角度、相机架设高度、道路车流密度3个指标,分析了不同条件下方法的有效性,结果表明:在相机高度为3 m,与路侧夹角为30°的环境中,车流计数精度在95%左右,但当公交、货车等大型车辆占比较高时,精度降为90%左右。在windows10 x64操作系统,2080Ti显卡,64 G内存,i7-7820XCPU的环境下,利用1080P视频流进行执行效率测试,显示相机架设角度和高度均对程序运行效率无显著影响,而车流密度则影响较大,在低密度流量下,FPS值约为44,而高密度流量下,FPS值降为33左右,表明该方法仍然具有较高的执行效率,可用于实时视频流量计数。

In order to study the problem that the automated technical means cannot be used in temporary traffic observation due to the temporary and uncertainty of the observation points,a method suitable for shooting videos from side-view angles,quickly recognizing vehicles and realizing traffic flow statistics is proposed.This method overcomes the difficulty that the traditional video recognition technology cannot meet the traffic video recognition from the side view angle.Taking the temporary observed roadside video as the object,the vehicle detection is carried out by the deep-learning-based YOLO_V3 method,a secondary target detection framework based on vehicle detection area and traffic counting area is proposed,and a traffic flow counting mode with Kalman filtering,Hungarian allocation,and perspective projection transformation is established to realize fast and high-precision tracking of vehicles.Multiple sets of actual video data are collected,and the effectiveness of the method under different conditions is analyzed in terms of intersection angle between camera and road,camera erection height,and road traffic density.The result shows that in an environment where the camera height is 3 m and the angle between camera and roadside is 30°,the traffic flow counting accuracy is about 95%,but when large vehicles such as buses and trucks occupy a relatively high proportion,the accuracy drops to about 90%.In the environment of windows10 x64 operating system,2080Ti graphics card,64G RAM and i7-7820XCPU,the execution efficiency test is performed using 1080P video stream.It shows that the installation angle and height of camera have no significant influence on the efficiency of the program,while the traffic density has a greater influence on the efficiency of the program.The FPS value is about 44 under low-density traffic,while the FPS value drops to about 33 under high-density traffic,indicating that this method still has a high execution efficiency and can be used for real-time video traffic counting.Under low-density traffic,the FPS value is about 44,while under high-density traffic,the FPS value drops to about 33,indicating that the method still has high execution efficiency and can be used for real-time video traffic counting.