Research On Vehicle Re-identification Based On Deep Metric Learning

The public transportation system generates massive amounts of data through surveillance cameras each day.Using artificial intelligence technology to analyze these data is beneficial for the construction of intelligent transportation systems and safe cities.Vehicle re-identification aims at retrieving vehicles in non-overlapping cameras,which plays an important role in improving transportation efficiency and maintaining public safety.There are two main challenges in this task as follows: 1)The appearances of different vehicles with the same type and color are similar under the same view? 2)The vehicle appearance of the same vehicle is various in different views.Aiming at addressing the first issue,this thesis proposes a novel vehicle re-identification method based on global-identity-center feature learning,which learns an identity center for each vehicle identity and encourages the model to decrease the discrepancy between vehicle feature and the corresponding identity center.Moreover,this method can also increase the discrepancy of vehicle feature to other identity centers to achieve both intra-class compactness and interclass separation in the embedding space.Considering that the identity centers are crucial,this thesis updates them with both global and local strategies,which can maintain accurate global identity centers with affordable computational cost.Aiming at handling the second problem,this thesis introduces view information on the basis of global-identity-center feature learning and proposes a vehicle re-identification method based on multi-view global center feature learning.The proposed method can effectively learn one global view center for each vehicle identity under each view,and encourage the model to reduce the distance between the vehicle feature and the corresponding view center while also reduce the distance between the view center and the identity center,which can further reduce the intra-class differences.Besides,based on the relationship between vehicle identity and view,the sample pairs are divided into same-view pairs,different-view pairs,and cross-view pairs to further constrain the distance between samples.Extensive experiments and ablation experiments are conducted on two well-known public datasets,and the experimental results prove the rationality and effectiveness of each innovation.