부산대학교 도서관

Targeting the safety behavior detection issue of mine overhead occupant devices, this study introduces an end-to-end MSD-DETR model. Firstly, a lightweight LPMConv module is proposed, which not only reduces computational complexity and model parameters but also enhances performance. Secondly, a reconfigured channel spatial interaction attention mechanism is integrated into the LPMConv module to bolster information interaction among channels, thereby improving the model’s focus on the target. Lastly, the cascade grouping CGA structure is employed to redesign AIFI, further reducing parameters and computation to enhance model speed and accuracy. Experimental results demonstrate that the FPS of the MSD-DETR model increases by 20 frames, with detection accuracy and recall reaching 94.2 and 93.7%, respectively, and mAP achieving 97.3%. Hence, the end-to-end detection model proposed in this study is fully equipped for detecting the safety behavior of mine overhead occupant devices.