Research On Pedestrian Re-identification In Intelligent Monitoring System

In recent years,the popularity of monitoring systems presents great challenges to tradi-tional artificial monitoring analysis,therefore promote the development of visual-based object detection,recognition,scenario analysis,and related applications.The intelligent video analysis based on person re-identification is becoming investigative emphasis and hot spot,and playing increasing significant roles in criminal fighting,disaster warning and security assurance.It has become a mainstream to solve the re-identification problem by deep learning,metric learning or such machine learning methods.However,due to the issues of large viewpoints difference,complicated illumination changes and severe occlusions,stable and accurate re-identification is still an urgent problem.This research mainly focuses on robust person re-identification,and carry out researches on camera view matching,single shot re-identification,non-sequential multi-shot re-identification and sequential multi-shot re-identification according to the different experimental conditions.The main innovative results of this dissertation are summarized as follows:Firstly,in single shot person re-identification,low dimensional features tend to be dom-inated by high dimensional ones in traditional feature-level fusion,which accordingly causes performance degradation.In order to solve this issue,a new decision-level fusion method is proposed.Benefited from building joint feature map to features and structural ranking output-s,Structural Support Vector Machine could effectively balance between the distinctiveness and invariant properties to illumination changes,therefore obtaining a global optimization.In addi-tion,in order to solve the problem that the heterogeneous data under multi-modality illumina-tion changes can not simultaneously satisfy intra-class compactness and inter-class separability,multi-modality local metrics are exploited to better measure the sample distances under multi-modality distributions,taking advantage of shift invariant property in Log-Chromaticity space.This method can achieve higher matching accuracy under multi-modality illumination changes.Secondly,in non-sequential multi-shot re-identification,multi-instance Neural Network based re-identification method is proposed to solve the identity ambiguity problem by exploiting the interaction information shared by the multi-shot images.Discriminative features are extract-ed by Siamese Convolutional Neural Network,and similarity measures are obtained by Softmax regression,then joint optimization is made by aggregating the similarity information of instance pairs in 'collective principle' and obtaining bag-level similarities with multi-instance Neural Net-work exploiting the interaction information.In this method,M3P pooling layer is proposed to aggregate multiple instance features into bag-level feature,therefore reduce the influence of im-postor samples,and achieve more precise model learning by optimizing network parameters in feedback mechanism.Thirdly,in sequential multi-shot re-identification,multi-source feature extraction and met-ric learning method is proposed to solve the problem of matching desynchrony of appearance and motion information.For appearance and motion features,tree-structured model and op-tical flow are separately applied in video segmentation to build candidate sample pools,then multi-instance metric learning with imposter rejection framework is proposed to iteratively and alternatively select the most discriminative sample pairs and learn optimal distance metrics.Finally,in pixel-level camera view matching problem,matched pointed in different cameras present quite different external appearance,which makes it difficult to perform key-point based matching.In order to solve this problem,this paper proposed a robust camera view matching approach based on spacial-temporal activity feature,which is more discriminative,and helps in-crease matching accuracy.Loss function is formulated by both spacial-temporal activity features and intrinsic attributes,and solved by Graph Cuts.This method has stronger practicability by being independent of homography matrix.To sum up,by analyzing the person re-identification problem,this dissertation studies the robust feature mining and matching strategy,which,to some extent,enriches the data analysis methods of person re-identification and promotes researches in related fields.Studies in this dissertation has certain theoretical significance and practical value.