Research On Key Technologies Of Progressive Person Re-identification

Person re-identification aims to find the same target person in non-overlapping multi-camera monitoring network with the given image sequence of a target person.It can provide the locations and corresponding times of the target per-son in the city,then his activity trajectories and behaviors can be acquired and analyzed through utilizing the information.It has gradually become a re-search hotspot at home and abroad.However,due to the diverse human be-haviors,complex surveillance environments,jagged device configuration and massive surveillance videos,person re-identification faces the following four challenges:1)How to obtain human appearance features with strong discrim-inative ability to instantly filter out irrelevant persons;2)How to mitigate the impact of viewpoint change on gait recognition performance;3)How to auto-matically learn the attentive periodic spatial-temporal information from irregu-lar gait sequences;4)How to fuse human appearance and gait features to ensure the efficient recognition of massive persons.This thesis proposes a series of novel models and methods from three aspects:human multi-level appearance feature representation,view-invariant feature extraction of human gait,and attentive periodic spatial-temporal infor-mation learning of irregular gait.Then a progressive person re-identification framework is constructed for large-scale surveillance scenes.Our proposed models and methods are verified by extensive experiments on large-scale pub-lic benchmark datasets.The main contributions of this thesis are as follows:(1)Human multi-level appearance feature representation.This thesis pro-poses a Siamese inception architecture network model to automatically learn the high-level semantic features of human appearance.For low-level and mid-level features,we extract the visual features such as color,shape,and texture of hu-man appearance.For high-level semantic features,Siamese inception structure is adopted to deal with the large intra-class variations and ambiguous inter-class differences of human appearance.Finally,the multi-level appearance features are combined into a comprehensive feature representation through exploiting the null space based metric learning method,which improves the discrimina-tive ability of human appearance.(2)Arbitrary view transformation method for human gait images.In prac-tical applications,viewpoint changes occur most frequently and have a great impact on gait recognition performance,so we propose cycle-consistent atten-tive generative adversarial networks model to translate gait images from source views to target view.The synthesized gait images of target view are exploited to improve the performance of cross-view gait recognition.Firstly,xwo-branch generative network is constructed to simultaneously perceive the global con-texts and local body details of human gait images.Secondly,a novel atten-tive adversarial network is designed to adaptively learn different weights to the discriminator's receptive fields with attention mechanism.It helps to mine the most attentive information to improve the quality of the synthesized gait images.Finally,the gait image reconstruction network and forward cycle-consistency loss are introduced to effectively exploit the data distribution prior knowledge of source domain and target domain in training process,which can preserve the identity of the synthesized gait image.(3)Attentive periodic spatial-temporal information learning of irregular gait.Due to the complex surveillance environments and diverse human behav-iors,human gait is always irregular,which seriously decreases the perfonnance of gait recognition.Therefore,we propose attentive spatial-temporal summary networks model to automatically learn the discriminative and attentive gait cy-cle features from irregular gait sequences for irregular gait recognition.First of all,general attention and residual attention units are designed and explored to discover the identity-related salient semantic regions from the spatial feature maps of gait images.Furthermore,LSTM unit is adopted to model the periodic motion of irregular human gait.Finally,an attentive temporal summary unit is designed to adaptively learn the attentive periodic motion cues for improving the performance of irregular gait recognition.(4)Progressive person re-identification fr-amework.This thesis designs a progressive person re-identification framework to instantly and accurately iden-tify the target person from large-scale surveillance videos.It is considered as a progressive search process in the feature space,which consists of two stages:appearance-based coarse filtering and human gait-based fine search.In the first stage,human multi-level appearance features such as color,shape,texture and high-level semantics are utilized to quickly filter out most of the irrelevant per-sons from coarse to fine.The search space is significantly reduced with little time cost to form the relevant person subspace.In the second stage,attentive periodic features of human gait are exploited to precisely search the target per-son in the relevant person subspace.The proposed framework can simultane-ously improve the timeliness and accuracy of person re-identification through adopting progressive strategy.In summary,we build a progressive person re-identification prototype sys-tem of fusing the human appearance and gait features.The effectiveness of this prototype system is verified by extensive experiments on large-scale pub-lic benchmark datasets.