Research and application of intelligent detection and statistical method of bolt based on ROD-YOLOv8 algorithm

Addressing the challenges of high missed detection and false detection rates for multi-scale bolt targets, insufficient accuracy of intelligent detection in complex underground environments, as well as low efficiency and high safety risks of manual detection, an intelligent bolt detection method based on the ROD-YOLOv8 algorithm is proposed. Building upon the YOLOv8 algorithm, this method introduces the DynamicHead （a dynamic detection head）, which realizes adaptive receptive field adjustment through attention mechanisms across three dimensions: scale awareness, spatial awareness, and task awareness. This enables the model to dynamically focus on effective feature regions according to bolt scales. The RCS-OSA module is adopted to replace the original C2F module, integrating structural reparameterization and channel shuffling techniques to enhance feature reuse capability and cross-scale information interaction, thereby effectively improving the feature expression ability of small targets. Experimental results show that the mean average precision （mAP） of the proposed method on the self-built bolt dataset reaches 87.97%, which is 5.37% higher than that of YOLOv8. The precision and recall rate are 0.940 and 0.799 respectively, among which the detection accuracy for small-scale bolts is significantly improved to 78.34%. For the bolt counting problem in video sequences, a deduplication strategy based on ROI （Region of Interest） division and target tracking is designed, achieving accurate cross-frame target tracking and deduplication through Intersection over Union （IoU） matching. The field tests indicate that the maximum absolute error between the statistical results and manual counting does not exceed one bolt, with an average processing time of only 7.6 seconds per 10 meters of roadway. This represents an efficiency improvement of 263.8 times compared to manual detection （with an average time of 33.4 minutes）. The proposed method can effectively cope with complex conditions such as cluttered underground backgrounds and large target scale differences. While maintaining real-time processing capability, it significantly improves detection accuracy and statistical reliability, providing an effective technical means for the intelligent acceptance of mine roadway support construction quality and holding important application value for promoting mine safety production and intelligent development.