Safety Belt Detection for Aerial Work Based on YOLOv5 and Few-Shot Learning
Aiming at the problem of safety belt wearing detection of high-altitude workers in power plant,most of the existing studies use deep detection model to detect direct,which needs a large number of samples to train the model.Moreover,due to the chaotic background of high-altitude operation and small personnel targets,it is difficult to detect.Therefore,a two-stage detection method based on target de-tection and few-shot fine-grained classification is proposed.Firstly,YOLOv5 is used to detect and extract the high-altitude workers in the video image,and then the few shot fine-grained classification method is used to realize the classification and recognition of whether safety belts are worn.In view of the characteristics of high similarity and the slight differences between people wearing and not wearing seat belts,the local descriptor is used to distinguish and express the subtle differences of images,a few-shot metric learning model is designed for seat belt recognition based on local descriptor,and the model is fine tuned with a small number of training samples on the basis of the pre-training model of public dataset.The fine tuned model is used for seat belt wearing recognition.The experimental results show that when the number of image samples in the support set is 60,the precision of the model after fine-tuning reaches 97.86%.The proposed method can realize the accurate detection of safety belt wearing of high-altitude workers in the case of few samples.
few shot learninglocal descriptorYOLOv5safety belt detectionaerial work
国家科技期刊平台
NSTL
万方数据
