Real-time monitoring of railway construction personnel at the edge based on YOLOv12-MobileViTv1

In railway maintenance work, the safety of construction personnel is of utmost importance. Among the series of algorithms in the YOLO model, the YOLOv12 algorithm represents an innovative model architecture. It not only employs the traditional CNN-based methods used in previous YOLO models but also incorporates a novel self-attention mechanism. This study aims to achieve real-time edge-based monitoring of construction workers, with construction workers, safety helmets, and reflective vests as the objects of detection. A detection model based on YOLOv12- MobileViTv1 is constructed by modifying the backbone network of the YOLOv12 model and replacing it with the MobileViTv1 module. Combined with the X-AnyLabeling annotation software, this enhances the accuracy of image recognition and generates a weighting file. In the laboratory, 3,222 images with a resolution of 1920x1080 were used to train the parameters of the YOLOv12-MobileViTv1 module. The accuracy, recall rate, mAP50, and mAP50-90 metrics were estimated. The results indicate that the improved YOLOv12 model, when combined with the MobileViTv1 module, achieves higher recognition accuracy for target images and more precise identification of construction workers working in hidden corners, thereby providing reliable safety assurance for railway construction personnel.