Lightweight Network for Anomalous Human Behavior Detection in Railway Track Areas

Manual patrols for detecting anomalous behavior in heavy-haul railway track areas have limited coverage, while deep learning methods require significant computational resources. To address these limitations, this study builds upon the YOLOv9-s algorithm. We introduce an optimized backbone network by integrating the Efficient Multiscale Attention (EMA) mechanism with Faster Blocks, proposing the novel EMA-Faster-RepNCSPELAN4 module. Additionally, the lightweight dynamic upsampling operator DySample is employed to replace the original upsampling method. This integration constructs a lightweight network for detecting anomalous human behavior in railway track areas. Experimental results demonstrate that the improved model achieves a 21.4% reduction in parameters, a 22.2% decrease in computational complexity (FLOPs), and a significant 24.5% reduction in model size compared to the baseline model. The proposed network exhibits strong performance on the railway human behavior recognition dataset, effectively balancing efficiency and accuracy. This advancement facilitates the practical deployment of the model for detecting anomalous human behavior in real-world heavy-haul railway scenarios.