Safety detection of power inspection personnel based on context decoupling and bounding box similarity measurements

In the power industry, wearing a safety helmet is an effective way of avoiding personal injury during the inspection process. This study proposes an improved YOLOv5 object detection method for the accurate detection of safety helmet wearing to address the numerous false positives and missed detections in complex environments. Firstly, a improved Task-specific context decoupling(TSCOD) is proposed in the detection head of YOLOv5. This module uses contextual information to perform feature fusion operations for the classification and localization tasks to enrich the abstract semantic expression of the classification results and supplement the position information lost in the convolutional neural network hierarchical transmission process as much as possible, improving the perception ability of the designed network to the object’s surrounding environment and detection accuracy. Secondly, a Normalized Gaussian–Wasserstein Distance(NWD) loss function is used to obtain the weighted sum of the distance between the prediction and target frames and the scalar value as the loss function of object detection. This function is based on the Wasserstein distance measurement modeling the bounding box as a two-dimensional Gaussian distribution to calculate the similarity between the corresponding Gaussian distributions. Finally, extensive experiments using a helmet dataset demonstrate its promising performance compared to state-of-the-art approaches.