A Study On Person Re-identification Based On Salient Region Features

Person Re-Identification?Re-ID?have been applied in the monitoring systems of important areas such as airport,high-speed railway station,campus,mall and residential areas.Compared with the strict constraints of face recognition,people's overall appearance features have more abundant and convenient advantages during the process of obtaining and representation.However,due to the non-rigid features of the appearance of person,the work of person appearance based Re-ID is far from the practical requirements in terms of accuracy.Therefore,we are facing severe challenges to improve the accuracy of Re-ID.Highly distinguishable features of person appearance reflects the difference of the inter-class and intra-class of the person features,which directly affects the accuracy of person Re-ID.Therefore,extracting highly distinguishable features of person appearance has become one of the most important challenges in order to achieve person Re-ID.As one of the person appearance features,the salient region feature can both effectively enhance the inter-class feature differences and reduce the intra-class feature differences.In addition,the feature similarity measurement is also a key factor to improve the accuracy of person Re-ID.In order to address the above challenges,a person Re-ID based on the game matching of the single person image salient region features is proposed,which is suitable for single query person image and unclassified gallery person image dataset.For person Re-ID based on person video,two person Re-ID algorithms are proposed,which include Image-to-Video person Re-ID using the salient region features clustering and person video matching based on the salient region features.Furthermore,sufficient simulations are conduct to demonstrate the superior performance and validate the feasibility of the proposed algorithms.The major contributions of this dissertation are summarized as follows.1.Person Re-ID using the single person image salient region feature matching gameThe local person feature has a stronger ability to distinguish features than the global person features.Person salient features retain most salient features of person appearance in order to calculate the similarity,which enhances the robustness of person feature representation.However,existing person salient features mainly segment the whole image into sub-regions of the same size to compute each region's salience,which causes that the salient features usually have poor adaptabilities to person pose variation,occlusion and other factors.Besides,a large number of non-salient features are discarded in the process of extracting salient features.In order to address the above shortcomings,a person Re-ID algorithm is proposed using the salient region feature matching game.First,high distinguish semantic regions of person image are extracted based on an improved hypermetric contour map.Then,each semantic region's salience is obtained based on the global contrast-based salient feature detection in order to select the most salient region of person images.When computing the similarity of salient regional features,we obtain the preference lists from two aspects of each query image and each gallery image.To make the non-similar person Re-ID with similar salient features more accurate,a matching game is introduced to optimize the preference list of each query image.2.Image-to-Video Person Re-ID using the salient region features clusteringExisting person Re-ID algorithms are low efficiency on the extraction of the person salient features,and the problem of image classification in person video images database is not considered.In order to address the problem,firstly,we construct the person video classification image dataset by extracting rectangular image area which containing single person from each video segment.Then,using the mean shift algorithm and the global contrast-based salience computing strategy,we present an efficient person salient region extraction algorithm.For the purpose of taking full advantages of every dimension feature of person images and improving the accuracy of distance measurements in the Euclidean Space,the gallery person salient region features are projected into the tensor space through introducing the least-squares log-density gradient clustering.Furthermore,there are multiple images of the same person in the dataset during the process of calculating the feature similarity.The similarity between single person salient region feature and salient region feature set is used to learn the distance measurements from the salient region features of a single person to the set of salient region features.Thus,the similarity between the images and videos of the persons is obtained.3.Person video matching based on the salient region featuresExisting video feature based person Re-ID algorithms face the problem of low levels of salience of person video feature.In order to address the issue,we design a person Re-ID which can measure the similarity between a single query person video segment and each gallery person video segment.To solve the problem of person video segment feature representation,we extracts and classifies the person image sequence in the video segment.Then,represents the person video feature by constructing the affine Hull feature of the person image set.In order to improve the salience of person video features,the super-pixel regions of person image are obtained by a simple linear iterative cluster algorithm and the background regions are removed using the border region growth strategy.Then,a group of most salient super-pixel regions are determined by calculating the global contrast-based salience of all super-pixel regions so that the person salient regions with semantic features are generated based on the region growth strategy.The feature modelling of multiple person images has the advantages of effectively weakening the impact of illumination,occlusion and other adverse factors comparing with that of a single person image,which is able to better represent the person features.Furthermore,the person salient region feature sequence are modelled as an affine hall.The affine subspace features is accurately estimated by suppressing the influence of extreme features based on the R₁-PCA algorithm.Finally,the salient region feature based person video matching is obtained using the affine hull top-push distance learning model.