A Method for Detecting Dangerous Behaviors of Power Operation Personnel Based on Paddledetection

To address the demand for real-time safety monitoring of workers in power operation scenarios, a method for detecting dangerous behaviors of power operation personnel based on RT-DETR is proposed. By integrating object detection and behavior analysis technologies, it achieves a key point detection method for efficient and accurate identification and early warning of dangerous actions. Firstly, by introducing the hybrid encoder structure of RT-DETR, the perception ability for multi-scale targets in power operation scenarios is enhanced through the intra-scale interaction module (AIFI) and cross-scale fusion module (CCFM). Secondly, an IoU-aware query selection strategy is constructed, and combined with the power operation dangerous behavior detection dataset, the initial target query generation mechanism is improved. High-confidence candidate boxes are screened through IoU constraints to reduce false detections and omissions. Finally, a lightweight dynamic inference model is designed, and by integrating with IoT technologies such as BLE ranging, the collaborative analysis ability of personnel positions and actions is enhanced. Experimental results show that on the self-built power operation dangerous behavior dataset, the average precision (AP) of this method reaches 86.5 %, which is 1.5 % higher than that of PPYOLO, and the inference speed is stably maintained at 98 FPS on the T4 GPU, making it suitable for the actual needs of power operation scenarios.