Improved Yolov5s-Based Helmet Recognition in Complex Scenes

Abstract

Enforcing helmet use can be important in improving traffic safety and reducing head injuries, particularly in the difficult context of the real world. The traditional detection systems are often not able to find tiny objects, occlusion, and dynamic lighting, thus, lowering its reliability and efficiency. A helmet detection dataset is provided with annotated pictures of both the helmet and no-helmet categories, and was borrowed off the Helmet Detection YOLOv8 Dataset v3. When maintaining parallel inputs in the classification and detection tasks, preprocessing involves the organization of the dataset, annotations of bounding boxes in the YOLO format, normalization, and the creation of a data configuration file. Some of the deep learning models have been applied, including YOLOv5s6u, YOLOv5x6u, YOLOv8, YOLOv9, YOLOv11, and YOLOv26 to detect effectively, and Faster R-CNN as a reference. The performance is measured with precision, recall, and mAP. By having the largest mAP of 0.930 and equal precision (0.868) and recall (0.874), YOLOv5s6u is the best-performing model, showing great detection ability in difficult conditions. In the case of helmet compliance monitoring systems, the proposed approach significantly enhances detection and real time applicability.