Improved Yolov5S normative identification algorithm of power place operators

In the process of electric power construction, there are a variety of non-standard or abnormal behaviors, such as not wearing a safety helmet, smoking, talking on the phone, falling, etc., which should be discovered in time to avoid safety accidents. Aiming at the existence of false detection and missing detection in the recognition of small targets such as smoking and telephone, an improved yolov5 method is proposed. Add Coordinate Attention mechanism in Neck of yolov5 to enhance the ability of the model to accurately locate the target. Accordingly, a small object detection layer is added to prevent the loss of shallow semantic information to some extent. Aiming at the problem of insufficient feature fusion, the original FPN+PANet was replaced by weighted bidirectional feature pyramid BiFPN to enhance the expression ability of multi-scale features. SIoU border loss function is used to improve the model recognition accuracy. The experimental results show that, compared with the yolov5s model, the improved yolov5s model has 3.7%, 2.0% and 1.9% improvement in recall rate, mean mean of average accuracy (mAP) and accuracy, respectively, among which the improvement of small targets such as smoking and phone call is the most significant.