Research On Key Matching Techniques For Person Re-identification

Due to the importance in automatic video surveillance,person re-identification(re-id)has attracted a lot of research interest in the computer vision and machine learning communities.Although the existing person re-id methods have tackled a number of difficulties and achieved interesting re-identification results,there still exist many practical problems to be studied in the process of pedestrian matching.(1)Distance learning is an effective matching technique in person re-id.The existing distance learning based person re-id methods learn the distance metric by exploiting the discriminative information contained in negative samples.However,these methods either treat all negative samples equally,or only utilize the negative samples that contain relatively more discriminative information,leading to the fact that they cannot make full use of the discriminative information contained in all negative samples.(2)Due to poor device quality or large distance between pedestrian and camera,the captured pedestrian images usually suffer from low resolution.In practice,low resolution will result in the loss of visual appearance information contained in pedestrian image.Since person re-id mainly relies on the visual appearance information of pedestrian,it is necessary to investigate how to reduce the influence of low resolution to the following re-identification.(3)In practice,large variations usually exist between different pedestrian videos,as well as within each video.These variations will directly influence the matching between pedestrian videos.However,the existing video-based person re-id methods can not deal with these two kinds of variations simultaneously.(4)The existing person re-id works mainly focus on the matching between pedestrian images,or that between pedestrian videos.In many practical scenarios,person re-id needs to be conducted between image and video.However,the problem of person re-id between image and video has not been well studied.In this thesis,we studied the above four problems existed in the matching process,and achieved some valuable results in terms of pedestrian matching techniques:(1)To make use of the different discriminative information conveyed by all negative samples more effectively and promote the discriminability of learned distance metric,we propose a novel distance learning approach for person re-id,which treats negative samples differently in the distance learning process.Specifically,for each target sample,we divide its negative samples into impostors and well separable negative samples(WSN-samples).Then we learn the distance metric by utilizing impostors and WSN-samples differently.For impostors,we design a symmetric triplet constraint,which requires the impostor to be far away from both samples of its corresponding positive sample pair;for WSN-samples,we require them to keep their favorable separability.Experimental results on three public benchmark pedestrian image datasets demonstrate that the distance metric learned by our approach owns better discriminability.(2)To solve the problem of person re-id under low-resolution scenario,which we call super-resolution(SR)person re-id,we propose a semi-coupled low-rank discriminant dictionary learning(SLD2L)approach.With the high-resolution(HR)and low-resolution(LR)dictionary pair,as well as the mapping matrices learned from the features of HR and LR training images,SLD2L can convert the features of Li probe images into HR features,such that the influence of low resolution can be decreased.To ensure that the converted features have favorable discriminative capability and the learned dictionaries can well characterize intrinsic feature spaces of HR and LR images,we design a discriminant term and a low-rank regularization term for SLD2L.Moreover,considering that low resolution results in different degrees of loss for different types of visual appearance features,we propose a mult1-view SLD'L(MVSLD2L)approach,which can learn the type-specific dictionary pair and mappings for each type of feature.Experimental results on multiple publicly available datasets demonstrate the effectiveness of our proposed approaches for the SR person re-id task.(3)To deal with the large variation between different pedestrian videos as well as that within each video,we propose a simultaneous intra-video and inter-video distance learning(SI2DL)approach for video-based person re-id.By regarding a pedestrian video as a set of spatial-temporal features extracted from each walking cycle of the video,SILDL simultaneously learns an intra-video distance metric and an inter-video distance metric from the training videos.The intra-video distance metric aims to make each video more compact,and the inter-video one aims to make the distance between two truly matching videos smaller than that between two wrong matching videos.To enhance the discriminability of learned metrics,we design a video relationship model,i.e.,video triplet,for SILDL.Experiments on the two public image sequence datasets show the effectiveness of our approach.(4)To solve the problem of matching between pedestrian image and video,we propose a joint feature projection matrix and heterogeneous dictionary pair learning(PHDL)approach.Specifically,PHDL jointly learns an intra-video projection matrix and a pair of heterogeneous image and video dictionaries.With the learned projection matrix,the influence of variations within each video to the matching can be reduced.With the learned dictionary pair,the heterogeneous image and video features can be transformed into coding coefficients with the same dimension,such that the matching can be conducted using coding coefficients.Furthermore,to ensure that the obtained coding coefficients own favorable discriminability,PHDL designs a point-to-set coefficient discriminant term.Experiments on two publicly available datasets demonstrate the effectiveness of the proposed approach.