Research And Implementation Of Progressive Vehicle Re-identification Algorithm Based On Multi-attribute Fusion

Vehicle automatic identification technology is an important research topic in intelligent transportation system.It aims to deal with the problem of automatic identification and retrieval of specific vehicles in actual traffic scenarios for improving the efficiency of traffic management and information retrieval.This paper mainly studies the automatic identification of vehicle images in actual road monitoring,including vehicle color identification,vehicle model identification and vehicle re-identification.The main research work is as follows:(1)A vehicle color identification method based on superpixel features is proposed.First,superpixel segmentation is performed on vehicle image.And then superpixels representing the background area of vehicle are determined according to the spatial position relationship between superpixels and the outermost pixels of the image.Due to the relatively fixed color characteristics of vehicle windows,3200 vehicle window images are collected so as to obtain the general color feature of vehicle windows by k-means clustering.Furthermore,the superpixels remained after vehicle background removal are k-means clustered into two categories.Similarity measurement between color features of the two categories of superpixels and general color feature of vehicle window is performed with the aim of detecting the superpixels of vehicle window areas and obtaining the vehicle color interest area.Finally,color classification is done using a linear SVM classifier.Experiments show that the accuracy of the proposed color recognition algorithm in color recognition of various vehicles exceeds 92%.(2)A vehicle re-identification algorithm that combines global and RNN dual-channel attention features is proposed.It uses the pre-trained Inception?Resnet?v2 as the basis for the vehicle feature extraction network.And then for one thing,the global attention featue channel learns the instance feature of vehicle derectly from the output of the pre-trained network.For another,the feature relationship between vehicle model feature and instance feature is mined through the hierarchical RNN attention channel.In addition,both channels are embedded with special attention mechanisms to guide the network for feature learning.Experiments on Vehicle ID and Ve Ri datasets indicate that the proposed algorithm shows its superiority in both rank-k and m AP indicators.(3)A vehicle color identification method and a vehicle model identification method based on the transfer learning model are proposed.To save computing resources and reduce the difficulties of model training,the models are based on the convolutional neural network fine-tuned on vehicle re-identification dataset.Furthermore,due to the differences of the abilities to mine image features between different network layers,the structure of specific layers of the network is improved accordingly to adapt to the task of color and vehicle identification.The experiments of the proposed algorithms are carried out on the Vehicle ID dataset.Experimental results show that the proposed algorithm not only reduces the difficulty of network training,but also ensures a high identification accuracy.(4)A progressive vehicle re-identification system based on multi-attribute fusion is implemented.The system combines the three modules of vehicle color identification,model identification and vehicle re-identification to meet the user's needs for vehicle information retrieval at different granularities.