Person Re-identification Based On Apparent Features

With the development of network technology and the enhancement of image acquisition technology, video monitoring has been widely used in many fields. The increasing videos bring serious challenges to video analysis, so the intelligent video analysis is needed. Person re-identification is one of the important researches in intelligent video surveillance, which aims to associate people across camera views at different locations and time in a distributed multi-camera surveillance system. And it underpins many crucial applications such as long-term multi-camera tracking and suspects search. Because of the complex environment and different view angles of cameras, person re-identification has been one of the most challenging problems. In order to deal with the challenges, features based on the appearance are designed to improve the robustness of target character description.To deal with the difference of view angles change between the cameras and the posture change, patches matching method is considered to reduce the effects of these two changes. A method which based on superpixel segmentation is proposed, and combined with the bag of words model to form superpixels features. Specifically, the dense SIFT features from training dates are clustered to build the dictionary of bag of word at first. For a new target, it is divided into patches by using superpixel segmentation method, then dense SIFT features are mapped to the dictionary. Then, the histogram which is related to the words frequency of each superpixel is obtained and these histograms gather together to form the superpixels feature. Because the superpixels feature is lack of color information, the average color information of pixels of each superpixel is added to improve the stability of the superpixels feature. Finally the EMD is used to calculate the similarity between superpixels features.Considering that the superpixels feature describes the local information of targets and ignores the overall characteristics, a method that fusing global color features and superpixels features is put forward. The weighted hue histogram feature is chosen as a global description for its insensitiveness to light change. Then, Bhattacharyya distance and EMD distance are used to determine the similarity between targets. Experiments on VIPeR dataset and PRID450 s dataset verify the validity of the fusion features.