Safety Helmet Wearing Recognition Based on Improved YOLOv5

Abstract

In the industrial production of digital workshops, workers need to wear safety helmets at all times. However, the accuracy of target detection is not high enough due to the characteristics of different light, angle of view, and people easily obstructing each other. To solve this problem, the real-time detection of helmets is realized by improving the YOLOv5 algorithm. The Dahua spherical camera is used to collect the data set, and the network is trained on the self-made data set through manual annotation. Pooling is carried out through softpool, so that it can retain more information of the feature map; meanwhile, improve the network structure of YOLOv5, add a layer of 9*9 feature layer, improve the recognition rate of the detection target, and use the DIoU loss function. According to the experimental results, the following results can be obtained. The average accuracy of the improved YoloV5algorithm in self-made data sets has improved a lot, above 97.3%. which is 4 % higher than the original algorithm, and the target detection speed is also correspondingly improved. It can effectively and real-time detect the wearing of helmets Condition.