Research On Vehicle Re-identification Algorithm Based On Graph Convolutional Neural Network

At present,the massive traffic data generated by cameras at urban traffic intersections can be used in fields such as traffic management and building smart security.Therefore,the retrieval of specific vehicles,namely vehicle re-identification(vehicle ReID),becomes very important.Given a query image,the task of vehicle re-identification aims to find out images of the same vehicle taken across different cameras.Since 2012,with the substantial increase in computing power and massive data generation,deep learning methods have continuously refreshed the highest performance records in various fields,including the vehicle re-identification task.In recent years,although there are various methods for vehicle re-identification,few methods considered the structural relationships among local features and between local features and global features.The problem makes it difficult for the model to obtain better results due to the lack of information in the extracted features.In addition,these methods rarely consider the background redundancy problem that is common in the dataset when extracting features,making the representation ability of the feature being weak.In response to the above problems,this paper proposes a vehicle re-identification algorithm based on the graph convolutional neural network.On the basis of this algorithm,this paper further proposes a vehicle re-identification method based on the spatial transformer network.The research content and results of this paper are as follows:(1)Vehicle re-identification algorithm based on the graph convolutional neural network.For the problem that most of the existing work ignores the spatial structural relationship between the local and global areas,this paper proposes a vehicle reidentification algorithm based on the graph convolutional neural networks.The algorithm first obtains multiple local feature maps by cropping the global feature map.Then the algorithm utilizes the graph theory to builds a graph structure with global and local features as vertices.The graph convolutional neural network is used to further learn the structural feature.Finally,the proposed algorithm concatenates the global feature and the structural feature.And the softmax loss function and the triple loss function are combined for joint learning to further boost the performance of vehicle re-identification.(2)Vehicle re-identification algorithm based on the spatial transformer networks.For the problem that the background redundancy in complex scenes makes the representational ability of the learned feature being weak,this paper proposes a vehicle re-identification algorithm based on the spatial transformer networks.The algorithm uses a spatial transformer network to transform the local features obtained by our proposed vehicle re-identification algorithm based on the graph convolutional neural network to remove a large amount of background redundancy.To solve the problem of the uneven spatial distribution of background redundancy,this paper further proposes a multi-spatial transformer network that can learn the local features adaptively.Through ablation experiments and comparative experiments on two public datasets Vehicle ID and Ve Ri-776,the effectiveness and superiority of the vehicle re-identification algorithm based on the multi-spatial transformer network are verified.