Helmet Detection Based On Improved YOLO V3 Deep Model

Helmet wearing is very important to the safety of workers at construction sites and factories. How to warn/identify/certify workers “whether or not the helmet is worn” is often a difficult point for enterprises to monitor. Based on the YOLO V3 full-regression deep neural network architecture, this paper utilizes the advantage of Densenet in model parameters and technical cost to replace the backbone of the YOLO V3 network for feature extraction, thus forming the so-called YOLO-Densebackbone convolutional neural network. The test results show that the improved model can effectively deal with situations that the helmet is stained, partially occluded, or there are many targets with a low image resolution. In the test set, compared with the traditional YOLO V3, the improved algorithm detection accuracy increased by 2.44% with the same detection rate. The establishment of this model has important practical significance for improving helmet detection and ensuring safe construction.