Pedestrian detection in coal mines is crucial for video surveillance systems. Limited computational resources pose challenges to deploying large models, affecting detection efficiency. To address this, we propose a lightweight pedestrian in coal mine detector with multi-level feature fusion. Our approach integrates the backbone network with coordinate attention, introducing a bidirectional feature pyramid network and a thin neck technique to enhance multi-scale detection capability while reducing computational load. We also employ regression loss with a dynamic focus mechanism for bounding box regression to minimize model errors. The Linkage Channel Pruning method enforces channel-level sparsity on the designed detector to achieve network slimming and secondary lightweight development. Results on a proprietary dataset demonstrate our method's parameters (0.61 M), computational load (2.0 GFLOPs), model size (1.48 MB), detection accuracy (0.966), and inference time (2.1 ms). Compared to the baseline, our method achieves a 4.96 × reduction in parameters, a 4.05 × reduction in computational load, a 4.02 × reduction in model size, a 59.62% reduction in inference time, and a 1.2% accuracy improvement. Experimental validation on proprietary and public datasets confirms that our method exhibits state-of-the-art lightweight performance, accuracy, and real-time capability, demonstrating significant potential in practical engineering applications. The insights gained provide technical references and real-time accident prevention for coal mine video surveillance systems.