Research On Vehicle Re-identification Based On Joint Deep Learning And Re-ranking

The vehicle re-identification is to retrieve the target images from the massive images collected by the urban intelligent traffic monitoring system for re-identification.The research on vehicle re-identification has far-reaching significance for the tracking and identification of illegal vehicles in traffic management,the statistics and analysis of traffic flow distribution rules,and the search and tracking of suspect vehicles in public safety departments.At present,most of the vehicle re-identification methods are based on the appearance characteristics of the vehicle.The re-identification accuracy cannot meet the actual needs due to the high similarity among vehicles and the complex background noise.This thesis proposes a vehicle re-recognition method based on Joint Deep Learning and Re-ranking(JDLR).The deep learning method is used to extract the high-level semantic features,detailed features and low-level color features of the vehicle and re-ranking the coarse recognition results of the vehicle for re-identification.The method is divided into two stages: vehicle coarse identification stage and re-identification stage with re-ranking.(1)In the vehicle coarse identification stage,a coarse identification method based on Joint Deep Learning(JDL)is proposed.The Siamese-Res Net50 network is constructed to extract the high-level semantic features and details features of the vehicles.At the same time,the VGG16 network is used to extract the low-level color features of the vehicles.Finally,the above two networks are combined to form a JDL network model,and the coarse identification sorting list of the vehicle is obtained.(2)In the reordering and re-identification stage,multi-feature fusion and K-Reciprocal Nearest Neighbor method are used to calculate the distance of the feature pairs among the vehicle images in the sorted list obtained in the vehicle coarse recognition stage,and the sorting result is optimized.The semantic features,detail features,and color features are extracted by the trained JDL model and then concatenated.The Mahalanobis distance of the feature pairs and the Jaccard distance of the K-Reciprocal Nearest Neighbor are taken as the distance metrics and linearly normalize them.Then the coarse identification results are further re-ranked to obtain the final vehicle re-identification result.The proposed algorithm is compared with the state-of-the-art vehicle re-identification methods on the Ve Ri-776 and Vehicle ID public data sets.The experimental results show that the proposed joint deep learning model can extract more robust vehicle features and the re-ranking tactics can improve the accuracy of re-recognition furthermore.