改进YOLOv5的油田作业现场安全着装小目标检测

Small Target Detection in Oilfield Operation Field Based on Improved YOLOv5

针对油田作业现场监控视频中的工人安全着装小目标检测效果较差的问题,提出了改进YOLOv5的油田场景规范化着装检测方法Cascade-YOLOv5(C-YOLOv5).首先搭建YOLO-people与YOLO-dress级联的小目标检测网络,定位行人目标,然后裁剪出行人区域并进行尺度变换,最后对行人进行安全着装检测;为了充分融合浅层与深层特征信息,在各级网络中使用4个不同尺度的卷积特征层来预测待检测目标.最后在原始图像中用不同颜色的框标出行人以及行人的着装部件类别,从而判定行人是否着装规范.实验证明,相比原始YOLOv5算法,C-YOLOv5方法不仅满足实时性的要求,而且检测的mAP提升了2.3%.同时,融合了深浅层信息的改进方法有效地增强了特征的表征能力,提高了小目标的检测精度.

Given the poor performance on the small target detection of clothing safety in video surveillance for oilfield operation,this study proposes a standardized clothing detection method based on Cascade-YOLOv5(C-YOLOv5),an improvement from YOLOv5.Firstly,a small target detection network cascading with YOLO-people and YOLO-dress is built to locate the pedestrian target.Then the pedestrian area is cut out and transformed in scale to detect the clothing safety of pedestrians.To fully integrate the shallow and deep feature information,this study adopts four convolutional feature layers with different scales to predict the undetected targets.Finally,in the original image,different color frames are used to mark the types of pedestrians and their clothing parts,determining whether the pedestrians are dressed properly.Experimental results show that compared with the original YOLOv5 algorithm,the C-YOLOv5 method not only meets the real-time requirement but also improves the detection mAP by 2.3 percentage points.At the same time,the improved method of fusing deep and shallow information effectively enhances the representation ability of features and promotes the detection accuracy of small targets.