Research On Vehicle Re-identification Based On Domain Adaptatio

Vehicle re-identification can be considered as an image retrieval task.Vehicle reidentification refers to retrieving the vehicle images that are consistent with the identity of the queried vehicle in the gallery captured by the camera group.In the past five years,in order to accelerate the construction of smart city and smart transportation system,researchers and engineers have begun to pay extensive attention to vehicle re-identification and other artificial intelligence related technologies.Compared with the success of single-domain vehicle re-identification with supervised learning,cross-domain vehicle re-identification still faces great challenges.Due to the distribution difference between the source domain and the target domain,the vehicle reidentification model trained in the source domain will have a significant performance degradation when directly applied to the target domain.In this paper,the effectiveness of the pre-training model,the quality of pseudo label and the distribution differences among different domains are studied.The main research contents are as follows:(1)For the limitations of current convolutional neural networks in performing convolution and pooling operations with small receptive field and loss of detail information,this paper proposes a Transformer-based pre-training network.The part adaptive embedding module preserves the complete local part information through centroid offset and edge length scaling,which improves the network’s ability to capture local detail information;the Transformer layer expands the receptive field range through self-attention,which establishes correlation between long-range features,and promotes the network’s robustness to global feature extraction.In addition,the coupling relationship between the part adaptive embedding module and the Transformer layer promotes the global and local features to work together,which makes the extracted features more comprehensive and rich.(2)Based on Transformer’s pre-trained model,this paper builds a domain adaptive vehicle re-identification framework based on clustering generated pseudo-labels.In order to reduce the influence of noise in pseudo-labels,this paper proposes centroid similarity ranking to find the difficult and easy categories that always correspond to each other in soft pseudo labels by calculating the Euclidean distance between the cluster centriod of this category and the cluster centriods of other categories and ranking them.The difficult categories are retained and the easy categories are discarded in the subsequent iterations to enhance the learning of the network for difficult samples.(3)In order to alleviate the distribution difference between the source and target domains,this paper proposes domain-style adversarial learning.The domain style feature branch is introduced to treat different cameras capturing images as different domains,and the classification of domains is achieved by learning domain style features from vehicle images.In the adversarial learning process,vehicle features are forced to be classified by domain classifier,and domain style features are classified by vehicle IDs classifier,which improves the ability of feature extractor to capture domain invariant features in the branch of vehicle features.