Research And System Implementation Of Mobile Terminal Safety Helmet Detection Algorithm

With the rapid development of China’s economy and heavy industry,China has made remarkable achievements in infrastructure.In order to ensure the personal safety of workers,it has become a mandatory requirement for personnel entering the construction site to wear safety helmets.The traditional construction site safety helmet supervision is a manual monitoring method.However,the long-term manual monitoring is prone to missed detection and low efficiency.In recent years,with the development of artificial intelligence technologies such as deep learning,the hard hat detection method based on deep learning has been widely used in industry.The helmet detection method based on deep learning is obtained through deep neural network training,requires strong GPU computing power support,and is expensive to deploy.In response to this problem,this paper implements a lightweight helmet detection model Helmet,which can be installed on mobile devices as an application,and designs and implements a web system for viewing alarm pictures and managing detection equipment.The details are as follows:1.Based on the Nano Det model,this paper adopts the improved Lc Net network as the backbone network.First,in order to reduce the amount of computation,only two depthwise separable convolutions are reserved for each layer in the middle layer,and a Shuffle module is added to the last layer of the network.The detection accuracy of the improved backbone network is increased by 1.2%,but the amount of computation is lower.Then,in the Neck part,a feature collection and redistribution module is added before the PAN,and the multi-layer feature maps from the backbone network are fused and then distributed into different sizes.After merging with the original feature maps,they are input into the PAN.The optimized Neck improves the detection accuracy by 1.8 %.Finally,using Mosaic data augmentation to enhance the diversity of the rich dataset,the model detection accuracy after data augmentation is improved by 1%.The final optimized helmet detection model achieves a detection accuracy of 92.8% in the validation set and 91.4% in the test set,which is 4% higher than the 87.4% of the original Nano Det model.2.In order to facilitate the deployment of the detection model,this paper builds the trained model into an installation package for installation on mobile devices.According to the test,the average detection speed of the detection application in Huawei Mate30 Pro 5G can reach 38FPS(Frames Per Second).The detection application will issue a voice reminder when it detects a target without a helmet,and can send an alarm picture to the outside.3.Use Go+Vue to implement a web system to receive and store the alarm pictures sent by the detection application.Managers can view the alarm pictures through the Web system,learn about the latest site conditions,and manage the testing equipment through the Web system.