Font Size: a A A

Design And Implementation Of Multi-scenario Vehicle Identification System Based On Deep Learning

Posted on:2020-11-08Degree:MasterType:Thesis
Country:ChinaCandidate:L H WanFull Text:PDF
GTID:2428330590450610Subject:Software engineering
Abstract/Summary:PDF Full Text Request
In recent years,with the rapid development of image processing technology,more and more video surveillance systems have added intelligent detection and recognition systems based on image processing technology,which greatly reduces the consumption of human resources and improves the reliability of the monitoring system.Among them,the detection and classification of vehicles has always been an important branch.Through the automatic detection and identification of vehicles,it can play an important role in intelligent traffic and safety monitoring.This topic comes from a smart optical cable monitoring project of a technology company.According to the surveillance video installed above the underground optical cable,it detects whether there is heavy vehicle entering,and instantly recognizes the alarm to prevent the underground fiber from being damaged by the heavy vehicle.Firstly,the research status of vehicle detection and classification at home and abroad is introduced and summarized,and the overall system workflow is designed for the actual scene of cable protection.The advantages and disadvantages of various motion detection algorithms are analyzed and compared.The hybrid Gaussian back-static modeling method is used to extract the moving target of the input surveillance video.The pre-processing module analyzes and compares various de-fogging algorithms and dark-light enhancement algorithms,and uses a dark channel prior theory to defogg the target area,and optimizes the time efficiency of the method.In terms of dark light enhancement,the logarithmic transformation method is used to improve the image quality in the weak illumination environment.Finally,based on the deep learning method,the characteristics between Alex,VGG and GoogLeNet networks are compared.GoogLeNet,which has better features and computational efficiency,is used as a training network,and then collects 10,000 pieces of image data covering seven models.Through the Caffe platform,a model for classification identification of vehicles is obtained.After testing the vehicle detection system,the results show that the accuracy rate is about 90% in the normal scene and about 88% in the complex scene.The processing speed can basically meet the real-time requirements,which basically meets the actual use.
Keywords/Search Tags:Vehicle identification, Deep learning, Caffe
PDF Full Text Request
Related items