Vision Uav Security Inspection System Design Based On Neural Networks

Safety production is the most important subject in engineering operations.Although there are strict regulations on the wearing of helmets in the construction environment,there are still many engineering personnel who are negligent in their work,lack safety awareness,and do not wear safety helmets when entering dangerous areas.With the development of computer vision technology in recent years,automated detection has attracted people's attention.Computers can use target detection technology to determine whether workers are wearing hard hats through the images captured in the camera,thereby forming a set of efficient and intelligent hard hat detection systems.For civil engineering building construction,fixed cameras cannot monitor the real-time situation of the engineering personnel on the varying situations,which inevitably increases the difficulty and cost of supervision.Taking the advantages of drone technology is apparently an efficient safety check measure.The main work of the paper is as follows:1.Based on the research background of the problem of hard hat detection and the research achievements made at the current stage of computer vision,a visual UAV security inspection system based on neural network detection algorithm was designed.The overall design of the security inspection system is given,including the composition and control of the quadcopter,the reception and detection of video images,and the system workflow of the helmet detection.2.The experiment designed a high-speed target detection framework for helmet detection.The overall framework of the detection model design,the backbone network design of the model and the design of the prediction module are given,and a network model with balanced performance and fast detection speed is continuously optimized through sufficient experiments.3.Based on the advanced lightweight network structure,an efficient convolution feature extraction module TSU(Transform Shuffle Unit)is proposed.This is a convolution module that combines the advantages of Shuffle Net and Mobile Net.It can effectively extract feature information with fewer parameters.While enhancing the characterization capability of the model,the network can have a faster forward running speed.4.On the basis of the classic neural network detection model structure,a target detection module suitable for hard hat detection is proposed.The lightweight detection structure adopts an anchor frame distribution scheme optimized for hard hat detection and a detection structurewith dilated convolution.On the one hand,it is used to improve the speed of model detection,on the other hand,it balances the accuracy of model detection.