Mine Personnel Counting Algorithm Based on Monocular 3D Detection

WANG Zhe; FU Zhe; CAO Pengjun; LI Qing; WEI Jing

doi:10.16545/j.cnki.cmet.2025.03.004

Abstract：Using deep learning to conduct real-time statistics on coal mine tunnel personnel can grasp the distribution of production personnel and ensure coal mine production safety.However,popular 2D object detectors rely on the spatial surface features of the target,which limits detection accuracy when the target is similar to the background or obstructed.To address this issue,a mine tunnel personnel counting algorithm based on monocular 3D object detection was proposed.Firstly,to address the issue of confusion between the target and background,additional geometric information was introduced.Based on a sparse query 3D object detector,a 3D object detector(M3D-HL)with a hybrid branch matching mechanism(HBMM)and a position aware classification module(LASM)was proposed.HBMM assigns multiple queries to each real sample through an additional one to many matching branch,which can obtain more predicted box information.LASM is used to select the top K predicted boxes with high classification prediction scores during updates,resulting in more accurate prediction results.Secondly,by utilizing the 3D information detected by M3D-HL and combining it with DeepSORT,robust 3D information inference can be achieved,thereby alleviating the problem of local occlusion and missed counting.Finally,the direction of personnel entering and exiting was determined by the order in which the center point of the side of the 3D detection box crosses the dual observation lines,and counting was achieved.The experiment showed that the algorithm could achieve efficient counting of personnel in coal mine tunnels and could be used in practical scenarios.