Research On Person Re-identification Based On Local Matching Fusion

Person re-identification technology refers to make correlation analysis of person images captured by different cameras,thereby determining a particular individual through those different images.Nowadays,person re-identification is a research hotspot in video processing.But this is a rather challenging task because of the unconstrained conditions,including the poses,viewpoint illumination,clothing changes,the partial occlusion of body,the background interference and other factors.In order to tackle there problems,researchers have proposed lots of algorithms to improve accuracy for person re-identification.The traditional methods extracted the whole features of person images,however,the overall appearance of a same person in different cameras' perspective usually have significant change.To fill this gap,some scholars divided person images into multiple local parts,and improve the recognition accuracy with the local matching fusion method.Based on this idea,this paper focuses on the segmentation of local parts,the description of local parts,and the fusion of local parts.The main work and research results are as follows:1.In order to tackle the problem of how to segment the local parts,this context proposes a method of human body parts segmentation based on adaptive clustering.Firstly,we make use of the best of the heuristic idea to determine the initial clustering center,and then determine the various regions based on the distribution of the pixel energy values.Finally,the improved squared error is used as the criterion for the end of clustering.This method can automatically decide the size and number of local parts according to the image content,and tackle the problem of rules in local parts segmentation are lack of guidance.The experimental results show that this method can improve the recognition accuracy of person re-identification compared with other segmentation methods.2.In order to improve the distinguishing performance of local parts in person re-identification,a person local feature extraction method based on improved color distribution field and Weber Local Descriptor is proposed.Firstly,we improve the distribution field model in the process of describing the color characteristics of person appearance by color distribution field,and analyze the distribution of pixel values of different persons with K-means,and then the level of adaptive color distribution is constructed.Secondly,the improved Weber Local Descriptor presents the texture features of person appearance.At first we make use of the differential excitation of the circular neighborhood to present the texture of the image,then use the directional component encoded by Local Binary Pattern to present the direction of the image.Finally,the person appearance is modeled by combining the above-mentioned color distribution field characteristics and Weber Local Descriptor.The experimental results show that this feature has better robustness than other features.3.In order to realize the person re-identification based on the recognition results of each local part,a local matching fusion method based on the saliency between images and the intra-images prominence is proposed.The saliency of intra-images describes the importance of local parts in all areas of the database,which is related to the composition and quantity of the training set with instability.And the saliency of the images indicates the importance of the local part in the image,regardless of the composition and quantity of the training set image.In this paper,we take advantage of the area under the normalized partial curve to calculate the images saliency between the local parts and the structured support vector machine to learn the intra-images saliency,and combine the two kinds of saliency to obtain the weight of each local part.The experimental results show that the fusion method can effectively improve the recognition accuracy of person re-recognition.