Research On Discriminative Feature Extraction Method In Person Re-identification

Person re-identification(Re-ID)is a challenging task of matching a specific pedestrian image in non-overlapping camera views with visual features of pedestrians.It can be regarded as a pedestrian retrieval task,which aims to find a specific person in a gallery of cross-view images.The key to the person Re-ID is learning and extracting the discriminative features that can identify the pedestrian identity from noticeable differences images obtained under different shooting angles.Therefore,extracting the discriminative features beneficial to judgment pedestrian identity by analyzing pedestrian characteristics from the images,has important research significance.However,there are some challenges in the practical application of person Re-ID caused by the changes in lighting,viewpoint,occlusion,human pose,etc.Thus,there are still some problems that need to be addressed:(1)The extracting of pedestrian discriminative features mainly according to the pedestrian information from a single perspective,and it is difficult to consider the discriminative features from other views at the same time,which leads to the limitation in the discriminative feature extraction;(2)In the process of the CNN learning the deep features for person Re-ID,the deep features are usually interfered with by the features that cannot distinguish particular pedestrian and are caused by the background and viewing angle changes.These interference features affect the performance of the person Re-ID model during the training phase.(3)During the retrieving phase,a large number of discriminative features extracted from the CNN always have high dimensions,which consumes a lot of computing and storage resources.Therefore,this dissertation conducted in-depth research on the discriminative feature extraction methods in person Re-ID,and proposed the corresponding solutions to meet the requirements mentioned above for person Re-ID.The main contributions of this dissertation are summarized as follows:1.Research on the pedestrian discriminative feature extraction method based on the cross-image fusion feature learning.This dissertation proposed a cross-image features fusion strategy(CFFS)with the neural network parallel training structure for the problem that the discriminative feature extraction mostly arrived from a single perspective.The method learns the single-image representation(SIR)and the crossimage representation(CIR)jointly for the discriminative feature extraction based on the metric learning method.This dissertation proposed a modularized cross-image features fusion strategy(M-CFFS)based on the CFFS which is more flexible for application.The CFFS method can be applied to many end-to-end neural network architectures for person Re-ID,obtaining the pedestrian discriminative features from multiple perspective images.The method improves the efficiency of learning the discriminative features of the same pedestrian from cross-view images.2.Research on the pedestrian discriminative feature extraction method based on the pedestrian discriminative feature dynamically selection.This dissertation introduced the swarm intelligence algorithms to select deep features.It proposed the region selection with discrete fireworks algorithm(RS-DFWA)for person Re-ID to solve the problem that the non-discriminative features caused by the original image noise information interfere with the performance of person Re-ID architecture.A heterogeneous network structure for both person Re-ID backbone network and RS-DFWA algorithm collaborative training--Region Selection-Re-Identification: Heterogeneous Co-Training(RSRI: HCT)is proposed further in the dissertation.While optimizing the backbone network for person Re-ID,the RSRI: HCT network structure focuses on the learning and extracting of pedestrian discriminative features.The RSRI: HCT dynamically reduces the interference of noise regions such as the background in the original image in the final pedestrian depth features3.Research on the pedestrian discriminative feature extraction method based on the pedestrian discriminative feature dimensionality reduction.This dissertation proposed a Guided Autoencoder(GAE)to reduce the high dimensionality of deep pedestrian features at the retrieval stage,for the problem that the deep discriminative features extracted from the CNN with huge amounts always have high dimensions.The GAE learns the liner relationship and nonlinear relationship among different dimensions by the proposed weight initialization,meanwhile,the GAE pays more attention to the identity of each feature while reducing the dimensionality in order to reserve discriminative information.The GAE we proposed has a high performance in dimensionality reduction of deep pedestrian features at the retrieval stage.The GAE could retain the pedestrian discriminative features that have a greater impact on identifying pedestrian identities.Meanwhile,the GAE could reduce the consumption of calculation and storage caused by the high dimensionality of deep features.