Research On Appearance Based Pedestrian Re-identification

In recent years, the large-scale deployment of city video surveillance system has promoted the development and application of the video investigation technology, which results in an urgent demand and higher requirements on relating technologies. As one of the most important technology, pedestrian re-identification plays an important role in the trajectory analysis of suspect objects. Unlike the traditional object track problem, pedestrian re-identification aims to recognize pedestrian objects in non-overlapping camera network. Because of the lighting variation, view and pose change, traditional object recognition methods, such as face recognition, gait recognition, are no longer suited to the accurate object recognition in the practical application. In contrast, appearance feature is robust to the influence of external factors and hence it is an ideal reference to recognize the identities of various pedestrians.Pedestrian re-identification is a new topic in the field of computer vision, there is not a unified theoretical framework. Recently, there are several methods:the method based on feature description and the method based on machine learning. The former focuses on finding good descriptor of the appearance; the latter is more concerned to find effective distance model of descriptor similarity measurement. From the practical point of view, the method based on machine learning faces several problems:(1) the high dimensional feature vector has high computational complexity, and requires a number of storage spaces;(2) the training process time is longer, and it is difficult to obtain required samples. Moreover, the method of feature description is based on visual information, considering image features that are relatively easy to extract, and generally do not need training process. For the practical applications, the method based on feature descrption has obvious advantages, and is the focus of this paper.In recent years, many appearance description methods were proposed, and experiments showed they were robust to illumination and viewpoint changes. The most representative method is Symmetry-Driven Accumulation of Local Features (SDALF). The SDALF method segments the human body into head, torso and legs, then extract three complementary features from each part, and combine three feature descriptors to represent the complete appearance. Although SDALF has a satisfactory performance on several standard test datasets, but there are still some problems in practical application:(1) the existing method's computational complexity is still high in the process of establishing descriptors and similarity calculation, and storage space is larger.(2) Pedestrian re-identification problem can be considered as ranking problem according to the similarity between the target and the candidate. Ranking results of recognition was not satisfactory after one search, and researchers proposed pedestrian re-identification methods using relevance feedback technique to refine the ranking results, trying to get the correct results as far as possible in the front positions, but high computational complexity leads to poor user experience.In view of the above problems, this paper is focuses on appearance based pedestrian re-identification technology for the video investigation application. Illumination invariant color descriptor and spatial information are combined to describe pedestrian appearance, considering the tradeoff of computing complexity and performance. Then, dissimilarity vectors between pedestrian appearance descriptor and a set of prototypes are used to represent the pedestrian appearance, reducing the feature descriptor storage cost and similarity computation complexity. Finally, a reranking process is introduced to refine pedestrian recognition results with good user experience. Experiment results show that this method can get performance close to SDALF, has the advantages of low computation cost and small storage overhead, and can be easily integrated into a practical application system.In this paper, the creative research achievements include the followings:(1) Appearance descriptor based on spacial color featureBecause color is affected by the illumination and viewpoint change, and basic color histogram does not retain the spatial information, which leads to the performance of pedestrian re-identification method based on color histogram is low. Aiming at this problem, this paper puts forward the spatial color characteristics of the pedestrian appearance based on the color descriptors invariant to light, to some extent, the influence of illumination change is overcome. On the other hand, in order to compensate for the lack of spatial information of basic color histogram, the appearance descriptor is calculated based on combining spatial information and color feature descriptor. The experiment results show this method gets performance close to the best, and has advantages of low computational complexity.(2)A fast appearance matching method based on dissimilarity representationBy aggregation of a variety of features, combining high dimentional vectors to compose the descriptor is bound to make feature extraction, feature description and matching complicated, which leads to the recognition system's low efficiency and the storage overhead. Aiming at this problem, this paper puts forward a fast matching method based on the dissimilarity representation. The method is inspired by dissimilarity representation for pattern recognition. A set of prototypes is selected from sample pedestrian database, and then difference between pedestrian appearance descriptors and the prototypes is calculated. Dissimilarity vector is new pedestrian appearance descriptor, it reduces the storage overhead of pedestrian appearance descriptors, and improve the similarity calculation speed. The experiment results show this method can improve efficiency without decreasing performance significantly.(3)Pedestrian re-identification method based on fast matching and rank-order distanceBecause of high complexity of pedestrian appearance descriptors calculation, the existing recognition methods using relevance feedback techniques to refine the ranking result get poor user experience, but also increase the burden of user to browse results and mark relevant ones. Aiming at this problem, this paper proposes a pedestrian re-identification method based on fast matching and rank-order distance. The method consists in adding a result refining process, namely after a search, the target pedestrian is put into the candidate base, then every candidate is used as a query to perform a search, and calculated rank-order distance for each pair of pedestrians. Finally, a new sorting result is returned according to the rank distance. The experiment results show performance improvement is attained.To sum up, in order to make the pedestrian re-identification technology for video analysis application system for investigation, the research objective of this paper is to design a pedestrian re-identification method with the tradeoff of performance and efficiency. Focusing research on describing appearance, matching method and refining the recognition result, three corresponding methods are put forward. The effectiveness of these methods is proved by experiments, and this work has important theoretical significance and application value.