Research On Vehicle Re-identification Based On Deep Learning

Nowadays,with the rapid development of computer vision,vehicle re-identification(re-id)technology has gradually become an important part of intelligent transportation system.Vehicle re-id task aims to the retrieval technology of identifying the same vehicle image across cameras.Vehicle re-id is facing great challenges due to vehicle attitude,high similarity of vehicle types and changes in lighting.Therefore,this paper mainly studies how to use deep learning method to solve the difficult problem in vehicle re-id task.To this end,we propose Pose-Aware based Multi-components Fusion(PAMF)method.The main work and innovations are as follows:(1)Research on vehicle re-id methods can be divided into two main directions: feature-based representation and distance metric-based learning.This paper introduces the common algorithms in each kind of methods.At the same time,the data set used in the experiment and the evaluation criteria are introduced in detail,and the performance of various methods is analyzed by comparing the experimental results.(2)The Pose-aware model(PAM)is proposed to classify the vehicle's attitude information within the group,measure the intra-class difference of the same ID vehicle image with the pose information i.e.viewpoint information,and design a multi-loss function based on attitude perception,which is used to optimize the model algorithm by combining weighted summation.This method can effectively solve the problem of intra-class differences in vehicle re-id tasks,that is,the visual appearance of the same ID vehicle image is quite different due to factors such as photographic angle or illumination conditions.(3)A feature extraction method based on multi-component fusion is proposed.The upper and lower regions of the vehicle are extracted separately,and the similarity calculation is carried out based on the complete image features of the vehicle.This method makes full use of the salient areas of vehicles,such as annual inspection labels and hangings of windows,and effectively solves the problem of similarity between classes in vehicle re-id tasks,that is,the appearance of different ID vehicle images is very similar due to vehicle type similarity and other factors.In this paper,experiments are conducted on vehicleID and VeRi-776,including vehicle re-id and vehicle retrieval task.It achieves better experimental results than existing methods.