Research On Vehicle Re-identification Based On Multi-feature Joint Learnin

The purpose of vehicle re-identification based on deep learning is to retrieve target vehicles in a large database through vehicle appearance features to achieve vehicle tracking,which is a core technology of intelligent transportation systems.With the development of artificial intelligence and big data technology,vehicle re-identification has a wide range of applications in suspect tracking,unmanned parking lot management,smart logistics and automatic driving.Especially if license plates are blocked,removed,destroyed,this technology will play a huge role.Since vehicle images are captured by different cameras,the illumination,viewing angle,and resolution are different,vehicle re-identification faces the challenge of small inter-class differences and large intra-class differences.Therefore,how to extract robust features of vehicle images is the fundamental task of vehicle re-identification.At present,the most commonly used method is to extract global and local features of vehicles.The method of extracting global features is simple and the calculation is small,but the instability of global features leads to low recognition rate.Local features are more robust to different environments,but using local features has the problem of extra consumption.Therefore,this paper proposed two models to solve the above-mentioned defects for global and local feature methods respectively.The research to be carried out in this paper includes:(1)Aiming at the inefficiency of recognition using global features,a global feature extraction model based on improved Dense Net and joint loss was proposed.SE module is inserted into Dense Net121 to learn the importance of each channel in the training process.Each channel is assigned the corresponding weights,which reduces the transmission of redundant information in the process of feature reuse in Dense Net121.At the same time,the proposed model leverages the complementary expression advantages of middle and deep features of the convolutional neural network to fuse the features of middle layer and last layer to extract more effective global features.In addition,a joint loss with focal loss and triplet loss was proposed.The joint loss increases the proportion of the loss of difficult-to-separate samples to pays more attention to the indistinguishable vehicle samples and enhance the model’s ability to discriminate difficult-to-separate samples.(2)Aiming at the problem that the use of local features requires huge annotations and complex models,a multi-feature learning model to enhance local area perception was proposed.The model includes two branches: global and local feature enhancement.The global branch obtains intermediate and high-level semantic features at the same time and integrates multi-scale pooling to enhance the global representation of the vehicle.The local feature enhancement branch proposes an enhanced local area perception model.First,the feature maps are divided into multiple non-overlapping local blocks to strengthen the learning of local area,and then the same areas in the same local blocks of training images are randomly discarded in batches to strengthen the learning of the remaining regions of each local block.By combining the global and local branches,the model simultaneously learns the overall structure of the vehicle and fine-grained multi-feature information,which enhances the model’s discriminative ability.