基于改进DETR的电力作业现场安全行为识别方法OACSTPCD

Aiming at the problem that the image samples collected from the electric power operation site are few and the distribution of positive and negative samples is not balanced.This paper proposes an improved detection transformer(DETR)method for recognizing safety behaviors at power operation sites.On the one hand,the generalization performance of the DETR model is improved by means of multiple pre-training.On the other hand,the Adapter module is incorporated into the DETR model and a small number of samples are utilized for fine-tuning.The experimental results show that the proposed method achieves high detection accuracy in helmet wearing recognition,short-sleeve shorts recognition,and seat belt wearing recognition scenarios.Under the condition of using 100 samples for training,the average accuracy is 0.81,which is 0.02 and 0.03 higher than that of YOLOv5 and Faster R-CNN,which are the state-of-the-art target detection models at the present stage,respectively;under the condition of using 300 training samples,the average accuracy is 0.84,which is 0.01 and 0.03 higher than that of YOLOv5 and Faster R-CNN,respectively.In addition,end-to-end training is realized based on the improved DETR method,and fast migration to other task scenarios can be achieved through Adapter fine-tuning.