The Algorithmic Research On Person Re-identification Based On Manifold Learning

Due to the rapid development of video surveillance network,it is a great challenge to process the huge video data with traditional artificial video analysis method.The intelligent video analysis method based on person re-identification has become the focus and hot spot in the field of computer vision.Person ReIdentification is a method which verifies whether the pedestrian captured by the camera at different location and times is the same person.At present,the research of Person Re-identification can be divided into two categories: one is based on feature representation,and the other is based on model learning.Person Re-identification based on the visual descriptors designs the feature descriptors to extract the distinguishing and stable features of pedestrian,instead of automatically learning features from the data.On model learning method,heterogeneous data with complex distribution is not strong and robustness.Due to all above,an advanced method is used to extract the color and textural features of pictures,and under the condition that in real scene data from target domain is not easily acquired,semi-supervised Fisher is used to reduce dimensions and then more robust manifold alignment method is used to do metric learning about heterogeneous data from two different cameras.Take advantage of the learned projection vector to project feature vectors from two different cameras from feature space to public latent space,which gets a result that feature vectors of one person from different cameras are closer while feature vectors of different persons from different cameras are farther.Finally,sort the distances in the public latent space to construct cross-view Person Re-identification system based on manifold learning.Inspired by manifold alignment,the author uses deep learning to do researches towards Person Reidentification algorithm and proposed a multi-channel deep convolutional network,which can automatically and concurrently extract partial features and global features in image segmentation,and then advances the previous feature fusion method and proposes a self-adaptive feature fusion module.Finally,the author uses triplet loss function with constraint to optimize the whole network parameters.This paper analyzes the contribution to the entire network accuracy and the function towards each partial and global features in Person Re-identification made by the self-adaptive feature fusion module.In the paper,firstly Person Re-identification based on manifold learning is experimentally verified in the Person Re-identification database like VIPeR,PRID and iLIDS.And then,Person Re-identification algorithm based on manifold learning is experimentally verified in the Person Re-identification database like Market1501,DukeMTMC-reID,CUHK01 and CUHK03.And,the experimental result shows that the proposed two algorithms can both make better achievement in Person Re-identification.