Answers are grounded in this paper and cite their sources.
Underground coal mine operating scenes feature complex working conditions such as significant dynamic variations, frequent target occlusions, intense exposure interference, localized low illumination, and high concentrations of pervasive dust. Conventional object detection models tend to suffer from severe overfitting, leading to a substantial degradation in generalization performance for detecting slender drill pipes. To overcome the deficiencies of conventional models, this paper proposes a drill pipe detection algorithm based on the lightweight YOLO11n network, integrating large-kernel separation and coordinate locking mechanisms. At the end of the backbone network, Large Separable Kernel Attention (LSKA) and Coordinate Attention (CA) are integrated to construct the LSCA feature enhancement module; this module uses large-kernel convolutions to connect fractured features and alleviate frequent occlusions, and employs orthogonal pooling to precisely lock onto the physical edges of drill pipes against intense light, darkness, and dust interference. At the detection head, an interpolation-based flexible regression loss (InterpIoU) is introduced to reconstruct the bounding box localization mechanism; by establishing a transitional interpolation buffer, it suppresses gradient oscillations caused by extreme aspect ratios and effectively reduces the short-edge divergence sensitivity in slender drill pipe regression. At the validation stage, a validation set is constructed based on an Out-Of-Distribution (OOD) strategy with physical isolation in time and space. Experimental results demonstrate that: (1) The improved model exhibits superior generalization at unknown operating sites, where mAP@50 increases by 2.0% over the baseline (reaching 88.7%) and mAP@50-95 increases by 2.2% (reaching 43.7%); (2) The improved model demonstrates stronger anti-interference capability under complex conditions, with the miss rate reduced by 1.4% over the baseline (to 18.3%) and the recall increased by 1.4% (to 81.7%); (3) The improved model maintains lightweight characteristics, with parameters reduced by 0.029 M relative to the baseline (reaching 2.561 M) while maintaining high real-time inference speed (reaching 1303.8 fps); (4) When coupled with the ByteTrack tracker at the back end, the counting accuracy on five challenging video cases is improved by 9.1 percentage points over the baseline.
No takes yet. Share an insight, caveat, or question.
Xiao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: