Object Detection Of Safety Helmet Wearing Based On Jetson Nano And Improved YOLO Algorithm

Accidents such as falling objects from high altitude often occur in construction sites,causing serious casualties.The supervision of wearing safety helmets plays a crucial role in promoting industrial safety.The use of artificial intelligence algorithm to detect the helmet-wearing provides a new solution.However,existing safety helmet-wearing detection systems typically require the transmission of surveillance videos to GPU servers for detection,which has poor portability,high hardware costs,and is difficult to apply in practice.This paper improved the YOLO series of algorithms for embedded devices,and deploys them on Jetson Nano to achieve low-cost real-time detection of helmet-wearing in construction site scenarios.The main work of this paper is as follows:(1)Aiming at the problems that the existing detection algorithms in the field of helmet-wearing target detection has slow inference speed,large model size,and high hardware requirements,YOLOv5,the existing object detection algorithm,is improved.Firstly,the structure of YOLOv5 network is modified.After comparing multiple lightweight network methods,the backbone of the YOLOv5 network was modified by using the group convolution and channel shuffle methods proposed by Shuffle Net V2.The Shuffle Net V2 module was added to the backbone part,which reduced the model complexity and model size and improved the inference speed of the model.Secondly,the model was optimized.The model quantification operation was used to reduce the weight,activation value and other parameters in the model from 32-bit float type to 8-bit int type.Use layer fusion operation to eliminate layers that are not output during forward propagation,and fuse some complex layers.After the model optimization operation,the computing power requirements for the model deployment are further reduced.After being deployed on the Jetson Nano,the average inference speed per image reached 32.2 ms,which was 84.7 ms faster than the YOLOv5 s model(116.9 ms)and achieved good lightweight performance.(2)Aiming at the problem that YOLOv7-tiny model has low detection accuracy in the field of helmet-wearing object detection,as well as the problem of false or missed detection when facing adverse conditions such as low lighting and background occlusion,we improve the YOLOv7-tiny algorithm in terms of accuracy.By adding CBAM attention mechanisms to the backbone connection part and feature extraction module of the network,the accuracy of detection is improved,and the probability of false detection and missed detection is reduced.The accuracy rate of the improved model reached 92%,a 1.5% improvement over YOLOv7-tiny(90.5%).After deployment on the Jetson Nano,we found that the improved method in this article can effectively suppress the issues of false detection and missed detection.This paper is based on the YOLO series of algorithms for improvement,resulting in a high-speed YOLOv5-SN model and a high-precision YOLOv7-CBAM model.Deploying these two models on Jetson Nano has achieved real-time detection of helmet wearing on construction sites,which helps to improve the intelligent level of site supervision.