Vehicle Re-identification By Extracting Discriminative Local Information

As a frontier subject in the field of computer vision,vehicle reidentification is currently mainly used in urban traffic,security tracking,and autonomous driving.The current mainstream vehicle re-identification algorithms are mainly based on deep learning methods.Although good results have been obtained on several public datasets,it still faces two major problems:First,different vehicles of the same type often have similar appearances,resulting in relatively small visible differences,which is more likely to be misidentified as the same vehicle;second,due to the different camera angles,the various postures presented by the car make it difficult to retrieve.In order to solve the above problems,this dissertation mainly proposes corresponding solutions from the following two aspects.1.This dissertation proposes a new representation learning method called Circle Pooling.Circle Pooling expands the feature maps extracted from the network in a circle from the center,extracts the feature blocks in the central area for average pooling,and concatenates the results to generate new features,thereby improving the network’s ability to learn more fine-grained local information.2.This dissertation proposes a new metric learning loss called Adaptive Margin of Triplet-Center Loss(AMTCL).According to the distribution of the training data,the loss function can adaptively adjust the hyper-parameters,so that the features between the classes are more separated;and in different stages of the network training,it is supervised to different degrees,so that the network quickly converges in the early stage of training,and enhances learning ability in the later stage of training.In order to verify the effectiveness of the functions of each module of the algorithm,this dissertation designs reasonable comparison experiments,and obtains better performance on two different source scenarios datasets.