Research On Vehicle Recognition Optimization Based On Deep Learning

Due to the rapid development of China's economy and the people's living standards continue to increase,the number of motor vehicles has also increased.With the continuous increase of motor vehicles,it has also caused various social problems.In China,the Intelligent Transport System(ITS)has also become a focus of attention.This system can increase the control of vehicle safety in the traffic police department team,and restore time to places where traffic accidents occur.Greatly improved the effective protection of people's safety and property.Object detection is one of the core functions of the ITS system.The traditional methods have achieved high development for video cameras and surveillance images.However,under various complicated backgrounds such as light intensity,unclear snow,rain and haze,etc.,they have a huge impact on vehicle detection and recognition.Traditional detection often because of these The reason is the problem of incorrect recognition and low recognition accuracy.This paper summarizes the traditional vehicle detection methods.In view of the shortcomings of traditional algorithms,a deep learning detection algorithm is used to detect vehicle targets.The focus is to solve the problem of low vehicle classification and recognition efficiency of traditional methods.The main tasks to be completed in this article are as follows:1)The structure and principle of the traditional vehicle detection algorithm are studied,and the performance analysis is carried out through experiments.The advantages and disadvantages of the current traditional vehicle detection algorithm are learned,which provides a basis for vehicle detection using deep learning algorithms.2)The working principle of convolutional neural network in target detection is studied,and the representative R-CNN,Fast R-CNN,Faster R-CNN and other algorithms in the current stage are studied,and the four algorithms are compared.The experimental verification on the multi-vehicle data set found the advantages of Mask R-CNN in multi-vehicle detection,and proposed using Mask R-CNN as the main algorithm in this paper,The first vehicle detection research using Mask R-CNN.3)The structure analysis of Mask R-CNN module is carried out.The defects of Mask R-CNN in vehicle detection are verified through experiments,and the optimization design of Mask R-CNN is proposed for the defects,and it is applied to the multi-vehicle detection data set.4)The experimental design is combined with the research background.In this paper,deconvolume operation(Deconv)is used to upsample conv3.The pool and deconvolution operations can leave more object details in the image.Then apply the conv4 and conv5 layers to take these three feature maps as input to combine these three feature maps(conv1,conv2,and conv3).These two levels produce more abstract semantic features and compress the three feature maps into a unified space.Local response normalization(LRN)is used to normalize the multi-layer feature maps.Then,in order to speed up the calculation,we add an average pooling layer to the candidate window,and its features are set to 8 * 6.The algorithm in this chapter verifies the feasibility of the algorithm in accuracy and speed by comparing the unoptimized Mask R-CNN.