Research On Person Re-identification Algorithm Under Occluded Scenarios

With the rapid development of information technology and economics,the quantity of surveillance videos is growing rapidly.It's necessary to detect and analyze the behavior of the target automatically in the video for large quantity of surveillance videos through intelligent surveillance technology.Person Re-identification plays an important role in intelligent surveillance.Person Re-identification aims to identify query person across different cameras.However,in real-world scene,as a result of the lack of person's information caused by occlusion,the person's feature is disturbed.This thesis researches on Person Re-identification algorithm under occluded scenarios.In this thesis,two algorithms are proposed in similarity measurement and feature learning.The main contributions of this thesis are concluded as follow:(1)For the Person Re-identification in occlucded scenarios,this thesis proposes a ranking optimization method via graph matching from the perspective of similarity measurement.The method fuses the local structure information of pedestrian image to remove the disturbance of invisible region in the similarity measurement stage.Most of the existing methods focus on the global structure information among samples and look for the neighbors in the feature embedding space.But under occluded scenarios,person is partially occluded.So the person image includes invisible regions.The noise of invisible region will disturb the feature representation,so that it disturbs searching for neighbors in the feature embedding space.In this thesis,the local features of pedestrian image are modeled as an attributed graph.The graph matching algorithm solves the correspondence between local features by searching the largest subgraph with similar graph structure to locate the shared visible region between the query image and gallery image.Experiments on occluded datasets prove that the method can improve the robustness of Person Re-identification under occluded scenarios.(2)For the Person Re-identification in occlucded scenarios,this thesis proposes a person re-identification method via spatial invariance exploitation and saliency analysis from the perspective of feature learning.Under occluded scenarios,the noise of invisible region will disturb the feature representation,so existing Person Re-identification methods under non-occluded scenarios can't be directly applied in occluded scenarios.Related research works of occluded scenarios need pose information or segmentation information as prior knowledge,which depend on a large quantity of data annotation and recognition model accuracy.This thesis proposes an efficient end-to-end framework.It automaticly aligns the feature between two person images via the Spatial Transformer Network,which is benificial to locate the shared visible region in the next step.Otherwise,this thesis proposes a Salience Guide Module,which prompts network to focus the share visible region automatically.This module learns the visibility scores of each horizontal parts and enhances local features with high visibility scores.In the similarity measurement phase,the visibility scores is used as the supervision information of similarity measurement to optimize the ranking list.Experiments on occluded datasets and non-occluded datasets prove that the generalization of the proposed method.