Helmet Detection Algorithm Based on the Improved YOLOv5 and Dynamic Anchor Box Matching

To solve the problems of low recognition accuracy and undetectable helmet of small targets in helmet detection in complex scenes, a helmet detection algorithm based on improved YOLOv5 and dynamic anchor box matching is proposed to improve the detection efficiency of small helmets in complex scenes. Firstly, by adding a small target detection layer in the YOLOv5 network, the detection accuracy of small targets is preliminarily improved; Secondly, convolution block attention model (CBAM) is added to the feature extraction network to enhance the information transmission between feature layers and the recognition of foreground and background by the neural network; Finally, to further improve the detection rate of small target helmet, the accuracy of a priori frame matching is enhanced by dynamic topK anchor frame matching. The weight of pre-training on the COCO data set is fused for detection and recognition to improve the generalization and accuracy of detection. The experimental results show that in the helmet data set constructed in this paper, the detection accuracy of helmets is 98.2%, and the helmet detection of small targets is realized.