Research On Algorithm Of Vehicle Re-identification System Based On Two-stage Strategy

With the increasing of Vehicle ownership and road cameras,Vehicle Re-ID based on computer vision is becoming more and more important in the road management mode of emerging cities.At present,most vehicle re-recognition algorithms identify vehicles based on the license plate and vehicle global features.However,due to the problems of road vehicles such as large inter-class similarity,small intra-class similarity and artificial occlusion of license plate,the recognition accuracy has gradually entered the bottleneck.In this paper,based on the public vehicle data set and self-made data set,the research on vehicle rerecognition in natural scenes is carried out to provide technical support for intelligent vehicle operation management.The vehicle color recognition algorithm based on machine vision is proposed based on RGB color normalization and Bo W(bag-of-words Model)due to the diversity of vehicle body color and the variation of light intensity or shooting Angle in natural scenes.First,the RGB color normalization algorithm was proposed to restore the vehicle color according to the atmospheric scattering theory.Then,based on Bo W idea,the color Coding rules of modified RBC(Radial Basis Coding)were used to encode the four parts of the vehicle,including the front,side,back and top.Finally,pixel spatial relation comparison support vector machine is used for color classification.In the vehicle color recognition experiment,its m AP reaches95.79%.Current vehicle re-recognition algorithms have low feature extraction ability for road vehicles,which are structured objects with large regional differences.This paper proposes a multi-branch vehicle re-recognition network algorithm which includes one global feature branch and two local feature branches.The global feature branch introduces the graph convolution module GMN into the Res Net50 backbone network to complete feature extraction of structured objects.The local feature branch is combined with the circular feature fusion segmentation branch and the decentralized spatial transformation branch to solve the influence of small local regional differences on vehicle re-recognition accuracy.Inspired by the process of human eye recognition,this paper proposes a two-step strategy re-recognition scheme based on color recognition and structure recognition.Firstly,the color recognition algorithm is used to eliminate the vehicles with different colors,and the actual color of the vehicles is restored.Then,the multi-branch vehicle re-recognition network algorithm is used to re-recognize and classify the processed data set.In the study,the proposed algorithm was compared with other algorithms using data sets such as Veri-776,Vehicle ID,Veri-Wild and VRe-Dataset,and its indicators were improved to some extent,indicating the effectiveness of the system.