基于YOLOv5 DeepSORT和虚拟检测区的车轴时空定位方法

Spatio-temporal axle localization method based on YOLOv5 DeepSORT and virtual detection area

为获得道路桥梁上汽车车轴的分布状况,基于YOLOv5 DeepSORT机器视觉技术对监控视频中车轴时空定位的方法进行研究。首先,根据监控视频中车轴多尺度、小目标的特点,提出基于Faster R-CNN算法的图像半自动标注方法,快速构建车轴目标检测数据集;利用YOLOv5算法检测视频中的车轴目标,并对YOLOv5系列算法性能进行评估;然后,提出在视频监测区域中设置虚拟检测区,先利用卡尔曼滤波算法对车轴目标的位置和状态进行预测,再分别利用重识别算法、匈牙利算法和级联匹配方法实现前后2帧车轴目标的匹配,完成基于DeepSORT算法的车轴多目标跟踪,生成车轴轨迹;最后,利用多目标跟踪结果,结合直接线性转换和基于匀速假定的位置推定,实现了对桥上所有车轴的时空定位。结果表明:在目标检测方面,YOLOv5s6模型表现最优,准确率达到96.42%,检测时间19.2ms/帧,对车轴具有高准确率和更快的检测速度;在多目标跟踪方面,基于虚拟检测区和YOLOv5 DeepSORT的多目标跟踪方法具有更好的检测和跟踪效果,与不设置虚拟检测区对比,多目标跟踪精度(MOTA)和识别精确率与识别召回率的调和平均数(IDF1)分别提升了14.7%和10.1%,被跟踪目标身份发生改变的次数(IDS)减少了108次。基于YOLOv5 DeepSORT和虚拟检测区的车轴定位方法能在车轴发生遮挡、目标较小和驶出检测区域等情况下准确检测和定位车轴,可为移动荷载识别提供准确的位置信息,同时也为桥梁动态称重方法的车轴信息检测提供了一种新思路。

In order to obtain the distribution of car axles on road and bridge,the method of spatiotemporal localization of car axles in surveillance video was studied based on YOLOv5 DeepSORT machine vision technology.First,based on the characteristics of multi-scale and small targets of axles in surveillance video,a semi-automatic image annotation method based on Faster R-CNN algorithm was proposed to quickly construct the axle target detection dataset.The YOLOv5 algorithm was used to detect the axle targets in the video,and the performance of YOLOv5series algorithms was evaluated.Then,a virtual detection area was proposed to be set up in the video monitoring area,first the position and state of the axle target using Kalman filtering algorithm,and then matching the axle targets in the front and rear frames were predicted,using the reidentification algorithm,Hungarian algorithm and cascade matching method,respectively,to complete the axle multi-target tracking based on DeepSORT algorithm and generate the axle trajectory.The multi-target tracking results,combined with direct linear transformation 1 Tf9 and position presumption based on uniform speed assumption were used,and the spatio-temporal localization of all axles on the bridge was achieved.Finally,the spatio-temporal localization of all axles on the bridge was achieved by using the multi-objective tracking results combined with direct linear transformation and position presumption based on uniform speed.The results show that the YOLOv5s6model has the best performance in target detection,with 96.42%accuracy and 19.2ms per frame detection time,which has high accuracy and faster detection speed for axles.In multi-objective tracking,the multi-objective tracking method based on virtual detection area and YOLOv5 DeepSORT has better detection and tracking effect,compared with no virtual,the MOTA and IDF1are improved by 14.7%and 10.1%,respectively,and the IDS is reduced by 108times.The axle localization method based on YOLOv5 DeepSORT and virtual detection zone can accurately detect and locate the axle in the cases of axle occlusion,small target and driving out of the detection area,which can provide accurate position information for moving load identification and also provide a new idea for axle information detection in dynamic weighing method of bridges.4tabs,11figs,29refs.