Abstract: Aiming at the problem of complex scenes, diverse targets and difficulties in intelligent safety monitoring due to partial occlusion of safety control under power site operations, we propose an improved algorithm based on YOLOv7-Tiny. Firstly, the YOLOv7-Tiny detection network is built, and the attention mechanism of channel reorganization is fused in this algorithm framework to effectively improve the interaction ability between channels and to enhance the saliency of target regions in complex scenes. Secondly, in the feature fusion stage, Res-PANet, a multi-scale feature fusion structure based on residual hopping, is constructed to effectively fuse multi-scale targets and to improve the multi-target detection capability in the scene. At the same time, the Swin-Transformer module is combined in the output detection head of the model to enhance the perceptual field of the model, to achieve enhanced global perception of the feature map by the model, and to improve the detection ability of the model in the case of partial occlusion. Then, an improved Mosaic data enhancement is adopted during training to enhance the number of small target distributions, to achieve the purpose of enriching the target scenes, and to improve the generalization ability of the model. Finally, the wearing of safety helmets and safety clothing of electric personnel, electric fences and electric warning signs are taken as the monitoring objects of safety operations for the verification of the improvement algorithm, and the heat map analysis based on Score-CAM is also adopted to further verify the effectiveness of the model improvement. The experimental results show that the average detection accuracy of the fusion-improved model can reach 90.1%, the image detection speed is 46 frame/s, and the test inference delay is 75 ms on the embedded hardware Jetson NX, which can effectively meet the requirements of power safety field detection accuracy and detection speed.