Research And Implementation On Vehicle Re-identification Method Based On Deep Learning

With the development of society,the increase in the number of vehicles has made road traffic situation increasingly complex,which brings many challenges to traffic control and traffic safety.The vehicle re-identification technology can quickly and accurately retrieve the target vehicle in the massive vehicle picture library.It plays an important role in road traffic control and has important value for maintaining traffic safety and social public safety.Therefore,it has become an issue that more and more researchers are concerned about.The task of vehicle re-identification is to identify all images of the same vehicle under the cross-camera in the vehicle picture library.In recent years,deep learning has been widely used in various tasks of computer vision and has achieved breakthrough results.Facts have proved that the vehicle re-identification method based on deep learning has obvious advantages over manual feature extraction.However,due to the strong appearance similarity between different vehicles of the same vehicle model,that is,inter-class similarity,they can only be identified by subtle appearance differences,which bring great challenges to vehicle re-identification.In view of the above problems,this paper proposes a vehicle re-identification method based on multi-grain learning and a vehicle re-identification method based on locally discriminative features.In view of the inter-class similarity problem of the different vehicles of the same vehicle model,the method based on multi-grain learning is proposed.It uses vehicle model information to divide vehicle re-identification into two sub-tasks: coarse-grained re-identification of different vehicle models and fine-grained re-identification of different vehicles of the same vehicle model.For two sub-tasks,a coarse-grained ranking loss function and a fine-grained ranking loss function are proposed respectively,which make full use of vehicle model information to optimize the similarity distance of the feature space,so that the network can learn the feature difference of different vehicle models and the feature difference of different vehicles of the same vehicle model.In addition,for datasets with incomplete label of vehicle model information,such as Vehicle ID dataset,considering the cost and difficulty of manual annotation,a method for online generating vehicle model labels is proposed,and the vehicle model labels are continuously updated during the training process.In view of the problem that global features can not pay attention to local feature differences of different vehicles of the same vehicle model,the method based on locally discriminative features is proposed.It uses the network jointly optimized by softmax loss function and triplet loss function as the baseline network.On this basis,the effectiveness of the local feature method based on rule division is first verified.Secondly,for the fixedness and manual operation of rule division,this paper further proposes local feature method based on attention mechanism which automatically learns local discriminative features through the attention model.Finally,the network is optimized by combining the local features based on rule division,the local features based on the attention mechanism and the global features to obtain more discriminative and robust features.Experiments show that the performances of the above two methods on the two mainstream vehicle re-identification datasets Vehicle ID and Ve Ri-776 can achieve state-of-the-art performance.