Person Re-identification In Complex Scenes Based On Local Features

Person re-identification(Re ID)is a technology for retrieving corresponding pedestrians in the datasets.Because of the rapid development of deep learning technology and the increasing demand for social security,coupled with the installation of a large number of cameras in public places,pedestrian can be re-identified.The development provides the conditions.Person re-identification technology solves the problem that a clear face cannot be photographed or cannot be photographed under the surveillance camera,and it relies on the overall characteristics of the pedestrian’s body to match pedestrians.However,in complex scenes,pedestrians in most public places will be blocked by other buildings or other pedestrians.In addition,the problem of relatively few pedestrian identities in the current pedestrian dataset also brings person re-identification many challenges.For these issues,the main work of this thesis is as follows:(1)A person re-identification network based on part features is proposed for most occluded pedestrian scenes.In order to simulate the pedestrian occlusion scene,the pedestrian image in the dataset is randomly and proportionally cropped with a certain probability.In order to solve the problem of incorrect matching of various parts of the pedestrian image,the algorithm of pedestrian local relocation is improved to further improve the performance of the model.Then add a global branch to preserve the global information of pedestrians feature in the model,the two network branches share the sample pairs selected from the hard sample mining triplet loss for joint training.(2)Considering that there are relatively few pedestrian identities in a single dataset,and the network is easy to overfit,a pedestrian data generation model that can maintain the consistency of pedestrians is proposed.Using the characteristics of Cycle GAN,the generated pedestrian images can maintain the identity of the original sample.The generated pedestrian images will have a certain amount of noise.In order to solve the problem of noise in the generated images and prevent the model from being affected by noise,label smoothing regularization is added to the generated samples.Experiments show that the proposed method has certain effectiveness in improving the performance of person re-identification network.(3)At last,Aiming at the issue that most of the current person Re ID models are complex and difficult to apply to actual scenes,a lightweight siamese network is proposed,and the idea of hard sample mining is introduced in the contrastive loss to select the more difficult ones.The identified pedestrian positive and negative sample pairs,because the model backbone network adopts Mobile Net V2,is relatively simple and efficient,and is not limited by hardware conditions.And this thesis constructs a person re-identification prototype system to visualize the person re-identification task and use it to retrieve the corresponding pedestrians in the databases.