Abstract: In coal mine safety production, real-time monitoring of the number and spatial distribution of personnel is a key basis for ensuring miners’ life safety and preventing accidents. However, current computer vision-based personnel counting technology still faces numerous challenges in complex coal mine environments. Given the problems of false detection, missed detection, and inaccurate counting, which are caused by insufficient multi-scale feature fusion and focusing capabilities of existing methods that result from multi-scale variation phenomena and non-personnel target interference in complex coal mine environments, a coal mine personnel tracking and counting method based on CDM-RTDETR is proposed. The RTDETR detection model is optimized through three aspects: within the backbone network, the C2f module is optimized using PSConv, and the backbone network CSP-AP is constructed to enhance feature extraction capability; in the encoder, the deformable attention mechanism is combined with the intra-scale feature interaction module to design the Deformable Attention-based Intra-scale Feature Interaction(DAIFI) to improve the focusing capability on the key features of personnel; additionally, the Multi-scale Feature Aggregation and Diffusion Pyramid Network(MADPN) is proposed to optimize the fusion and transfer of multi-scale features. On the basis of the optimized detection model, the ByteTrack target tracking algorithm is incorporated to achieve stable tracking of personnel trajectories and accurate counting. Experimental verification is conducted on the constructed coal mine personnel detection dataset and coal mine personnel tracking dataset. Results show that in terms of detection performance: compared with the original RTDETR detection model, the precision and recall are improved by 2.3% and 3.8% respectively, the number of parameters is reduced by 12.4%, and the mean average precision (mAP@0.5) reaches 97.8%. In terms of tracking performance, the Multiple Object Tracking Accuracy(MOTA) reaches 90.8%, and the Multiple Object Tracking Precision(MOTP) reaches 84.1%. In terms of counting accuracy, the comprehensive accuracy of personnel counting reaches 95.3%.