Deep Learning Based Person Re-identification

Person re-identification,which aims at matching pedestrian images from nonoverlapping cameras,is a research hotspot in the field of computer vision in recent years.The application scenarios of person re-identification technology include finding missing persons,public safety and assisting police in finding criminal.However,the practical application of pedestrian re-identification is still limited,mainly because the technology still faces difficulties in the differences in lighting conditions,the impact of camera resolution,the presence of background interference and occlusion,and changes in pedestrian poses.These factors make images from different cameras look different from each other,even they belong to the same person,leading to misdetection and false alarms at pedestrian images match.Bases on deep learning,this thesis optimize these issues,including deep metric learning,pedestrian's viewpoint variations and camera style problems.First,this thesis addresses the current deep metric learning methods that pay attention on samples and ignore class.The current methods,such as verification loss and triplet loss,only focus on the constrains for samples.However,the sample-level constrain is easily influenced by wrong sample or extremely hard samples.The loss may arise when selecting these samples and farther cause shock for training.This thesis focuses on this problem and proposes the center-level verification loss,which designs the loss function based on centerlevel instead of sample-level.Each class of training set will get a center by iteration.As for the design of constrain,the intra-class constraint limits the distances feature and the corresponding center.As for the inter-class constraint,the distance of different centers should be larger than a margin,so as to make the gap among different pedestrian images larger.This thesis takes the center into consideration and optimize the intra-class distance and inter-class distance at the same time,achieve competitive results.Secondly,this thesis considers the viewpoint problem,which exists excessively large differences in viewpoint,making it difficult to correctly match images.In order to successfully model the differences between different viewpoints,according to the direction of the pedestrian towards the camera,the pedestrian images are divided into front,side and back.For pedestrian images between different viewpoints,it is considered that there are certain differences in the features proposed by the same network from them.In order to eliminate this difference between features due to viewpoint problem,we construct a viewpoint transform matching network to transform the features of source viewpoint to target viewpoint.Besides,we propose viewpoint-aware classifier and viewpoint transform loss.The former keeps the features of each viewpoint discriminative for ID,and the latter forces the transformed feature closer to the feature of target viewpoint.At the testing stage,all features of different viewpoints are transformed to the same viewpoint.We conduct experiments on Market-1501,Duke MTMC-re ID and CUHK03 and achieve competitive performance.Finally,this thesis considers the impact of camera style(camstyle)factors on person re-identification.Camstyle factors are defined as interference factors that are jointly affected by the camera built-in parameters and the external deployment environment.Due to the complexity of the source,it is hard to accurately model the camstyle to solve it by prior information.This thesis proposes a method based on deep learning to try to eliminate the influence of camera style factors on person re-identification.We call this method camstyleidentity disentangling network.This method first constructs a disentangling network to disentangle the camstyle factors from the feature of the pedestrian image,and then generates more samples with different camera styles through data augmentation.More camera style samples will participate in training and help the network eliminate the effects of camera style.We conduct experiments on Market-1501 and Duke MTMC-re ID,and obtain state-ofart results.