Abstract: To further improve the detection accuracy of safety harness wearing status during high-altitude operations and ensure the safety of underground coal mine personnel, a RT-DETR-based model for detecting safety gear compliance during elevated operations in coal mines. First, the original backbone network ResNet in RT-DETR is replaced with the lightweight network ShuffleNetv2, significantly reducing the model parameters while largely maintaining detection accuracy and improving inference speed. Second, a Focaler-MPDIoU loss function is constructed by integrating Focaler-IoU and MPDIoU, which addresses the issue of the GIoU loss function ignoring geometric deviations in two perpendicular directions and reduces the missed detection rate. Furthermore, quantization-aware training is applied for fine-tuning to further optimize model performance. The results show that the proposed method achieves a detection accuracy of 97.2%, with a weight file size of 17.9 MB. Compared with the original model, the weight file size is reduced by 88.9%. While ensuring detection accuracy, the inference speed is effectively improved. In comparison with current mainstream target detection models, the proposed model demonstrates superior performance and practical applicability.