An efficient attention-based detection network for underground miner pose recognition

Abstract Visual pose recognition of underground miners is essential for safety monitoring, yet it faces a critical conflict between high-precision measurement under severe visual degradation and the stringent computational and thermal constraints of intrinsic-safe edge devices. To reconcile these conflicting demands, this study proposes MinerPoseYOLO, a lightweight detection framework that prioritizes architectural-level optimization over post-hoc model compression. The proposed architecture integrates an HGNetV2 backbone for dense multi-scale feature representation, an ADown module to attenuate spatial information loss during downsampling, and a parameter-efficient PAttBottle attention mechanism to adaptively recalibrate pose-related features with negligible computational overhead. Furthermore, the MPDIoU loss is leveraged to enforce geometric consistency and improve bounding box localization performance. Experimental results on the DsLMF dataset demonstrate that MinerPoseYOLO achieves a mean average precision @0.5 of 92.6% while reducing the model parameter count by approximately 40% compared to the baseline, verifying its effectiveness under strict resource constraints. To further assess its applicability in heterogeneous underground scenarios, the model is retrained on the CMU-Drill dataset, which represents a distinct underground visual detection task with similar environmental challenges, where stable performance is maintained. Crucially, embedded deployment experiments on an NVIDIA Jetson Xavier NX platform substantiate the feasibility of low-power, real-time inference in practical industrial environments. These results indicate that MinerPoseYOLO offers a reliable solution for underground visual measurement tasks, effectively balancing detection performance, efficiency, and deployability under harsh environmental and computational constraints.