Central Feature Learning For Vehicle Re-identification

In recent years,with the country’s focus on new infrastructure and intelligent transportation systems(ITS),road traffic supervision facilities have become increasingly complete.In large and medium-sized cities,massive amounts of traffic monitoring data are generated every day,and such large-scale data has huge potential for use.With the maturity of artificial intelligence and cloud computing technologies,how to use algorithms to exploit the key information in massive amounts of data? How to improve the operating efficiency of ITS through data analysis? How to improve the efficiency of society through intelligent traffic supervision? These three tasks are very worthy of research.Vehicle re-identification technology is designed to distinguish whether the vehicle images taken in non-overlapping areas belong to the same vehicle.It is a significant task in the development of ITS,and has great research significance in improving social operation efficiency and maintaining public safety.At present,with the development of deep learning and person re-identification technology,many excellent vehicle re-identification methods have been proposed.However,the difference in vehicle appearance caused by the variation of viewpoint and the high similarity of the same model under the same viewpoint are still two major challenges that have not been resolved in the field of vehicle re-identification.Aiming at these two challenges,this paper proposes a vehicle re-identification method based on central feature learning,uses hierarchical density clustering algorithm to learn vehicle perspective grouping,and innovatively proposes intra-view triples and inter-view triplet loss functions.The triplet loss functions are used to reduce the feature distance between samples of the same vehicle,and finally feature learning is integrated to try to obtain a feature space with high recognition and stability.Besides,in order to further improve the robustness and recognizability of vehicle features,this paper proposes an orthogonal center learning method based on subspace masking,innovatively proposes a center constraint loss function to improve the compactness of similar features,and uses subspace masking mechanism to encourage the model to capture local discriminative features and enhances the model’s ability to extract discriminative information.This paper also uses the orthogonal regularization technology to improve the robustness of feature expression and avoid the learning process from falling into the local optimal solution,and at the same time ensure weak correlations between different classes of features.In this paper,a large number of ablation experiments have been carried out to prove the effectiveness and complementarity of each innovation.We also compared our method with state-of-the-art methods on four public benchmark datasets,the experimental results have proved that the central feature learning method proposed in this paper significantly outperforms other state-of-the-art methods,and fully demonstrates the accuracy and generalization ability of this method for vehicle re-identification tasks.