Application Of Object Detection Based On YOLO V4 In River Sand Mining Supervision

In recent years,the problem of illegal sand mining is becoming more and more prominent.It is imperative to carry out effective supervision and control.However,river sand mining and other illegal behaviors have strong dynamic change,and relying on traditional manual inspection or video surveillance can not find problems in time,and low efficiency,high cost.Video monitoring of water conservancy is widely used in river,so it is possible to build an application platform of object detection technology based on deep learning.In this paper,a new algorithm improvement model is proposed based on YOLOv4 algorithm based on the collected sand mining supervision objects(sand mining ship,sand mining vehicle,sand mining machine,etc.).After experimental verification,the optimized and improved YOLOv4 algorithm model achieves the best balance between speed and accuracy at the present stage,the computational complexity is reduced by 3.6G,and the reasoning speed is increased by8 FPS.Finally,based on the improved YOLOv4 algorithm model,the platform of river sand mining supervision system is researched and developed,which greatly improves the efficiency of river sand mining supervision.This paper mainly studies the optimization of YOLOv4 algorithm and the development of river sand mining supervision platform.1.Optimization and improvement of YOLOv4 algorithm: Firstly,the channel shuffling idea of light-weight convolutional neural network model Shuffle Net V2 was used for reference,the number of parameters was reduced by half,and the information interaction of channel feature diagrams of different groups was strengthened.Secondly,the idea of Rep VGG structure reparameterization is used to realize multi-branch training and single-branch reasoning.Training multi-branch structure can learn richer semantic information and spatial location information of image,reasoning single branch structure can realize fast reasoning,so there is no complicated tensor calculation,and can reduce the computing memory of hardware,so that the network has high performance and efficient reasoning speed.Finally,Shuffle Rep VGG module is constructed by Shufflenet V2 and Rep VGG,and a new network structure module Csp SRB is proposed by combining the idea of csp.2.Development of river sand mining supervision system platform:Based on the improved YOLOv4 algorithm,real-time detection and alarm are carried out for the targets of sand mining vessels,sand trucks and sand mining machines and tools in the supervised river channel,so as to realize the 7×24 hours non-stop automatic inspection of the river channel,and timely notify relevant responsible persons to timely deal with problems through the system page,mobile phone messages and APP.