Helmet wear detection based on YOLOV5

Abstract

Safety helmet wearing detection is an important safety inspection task with widespread applications in industries, construction, and transportation. Traditional safety helmet wearing detection methods typically use feature-based classifiers such as SVM and decision trees, but these methods often have low accuracy and poor adaptability. In this paper, we propose an improved helmet detection method that uses a combination of SPD Conv, ASPP and BiFPN structures to increase the perceptual field to ensure maximum feature extraction from the helmet, and can ensure fusion between different feature layers to pass semantic information to deeper neural networks, effectively avoiding information loss and improving the performance of detecting helmets. Experimental results show that our method has a 1% improvement in the average accuracy of detection in the public dataset VCO2007 set compared to YOLOv5, which still allows for real-time detection and meets the needs of industry with some practicality.