结合行人检测与单应性变换的安全社交距离估计

Safety Social Distance Estimation Based on Person Detection and Homography Transformation

为解决未标定摄像头监控视频中行人安全社交距离的估计问题,提出将行人检测、单应性与尺度估计相结合的方法,对单目相机中行人是否处于安全社交距离进行二分类。首先基于YOLOv5s框架,采用MSCO⁃CO数据集中只含有行人的数据训练得到鲁棒性较好的行人检测器;然后根据相机成像模型假设,推导出从场景地面到图像平面的单应性矩阵,再通过人类平均身高和图像行人检测框高度估计从场景地面到图像中行人局部区域的尺度信息,从而在图像上投影出行人椭圆形安全区域;最后通过计算图像中行人安全区域的重叠情况判断其是否违反安全社交距离。实验结果表明,当IoU=0.5时,该方法在MSCOCO只含行人数据的验证集上的行人检测准确率、召回率和AP分别达到81.39%、82.39%和76.52%;在OTC数据集上行人安全社交距离二分类的准确率、召回率和F1值分别达到98.99%、89.12%和93.79%。所提方法在一般监控场景下对行人安全社交距离违反情况的检测性能较佳。

In order to solve the problem of estimating the safe social distance of pedestrians in surveillance videos with uncalibrated cameras,a method combining pedestrian detection,homography and scale estimation was proposed to classify whether pedestrians were in the safe social distance in monocular cameras.Firstly,based on the YOLOv5s network,the pedestrian detector with better robustness was trained by using the data of MSCOCO containing only pedestrians.Secondly,according to the assumptions of the camera imaging model,the homography matrix from the scene ground to the image plane was deduced,and the scale information from the scene ground to the local area of the image was estimated by the average height of human and the height of the image pedestrian detection box,so as to project the elliptical pedestrian safety area on the image.Finally,the overlap of the pedestrian safety area in the image is calculated to determine whether the safe social distance is violated.Experimental results show that when IoU=0.5,pedestrian detection precision,recall rate and AP of MSCOCO dataset reach 81.39%,82.39%and 76.52%,respectively.The precision,recall rate and F1-score of pedestrian safety social distance classification of OTC dataset reached 98.99%,89.12%and 93.79%,respectively.The proposed method has high performance in pedestrian detection and safe social distance violation detection under normal security camera circumstances.