A Lightweight YOLOv5s-Based Target Detection Method for Railway Workers

The precise detection of workers through on-site surveillance videos is crucial for enhancing the intelligent management of construction safety. This paper introduces an improved lightweight network, Yolov5s-NSFnet, which builds upon the original Yolov5s model to effectively address the challenges associated with small object detection and real-time processing on embedded devices in surveillance settings. To mitigate the difficulties posed by long monitoring distances and the detection of workers as small targets, we enhance the Yolov5s architecture by replacing its backbone network with the more efficient FasterNet. Moreover, we have incorporated SE channel attention mechanism into the FasterNet block to enhance the feature extraction capability of critical channels. Finally, we use the Normalized Wasserstein Distance (NWD) loss function instead of the traditional loss function to improve the accuracy of small object detection in the dataset. The experimental results show that our method not only effectively reduces the number of network parameters and computational load, but also improves the detection speed of the model on embedded devices while maintaining good detection accuracy.