Person Re-identification Based On Appearance Features

Person re-identification is a challenging problem in both computer vision and machine intelligence in recent years. The goal of person re-identification is to detect the target objects using knowledge learned from human recognition model and vision schemes. In order to recognize and match the objects from different scenes, we have to represent and combine visual information from different views. The study of person re-identification involves the knowledge of cognition psychology, computer vision, pattern recognition, machine learning and so forth. The problem has a wide range of potential applications including video surveillance, human behavior analysis, judicial investigation, human-robot interaction, personalized healthcare, domotics and online shopping.Solving the person re-identification problem has gained a rapid increase in attention of researchers from both academic communities and industrial laboratories in recent years. However, due to the complexity of human body structures and the surrounding environments, the research of person re-identification is still in the stage of exploration. Most of the standard datasets reflect a "closed world" senario, e.g. exactly two camera views and 1:1 identity correspondence between the cameras. We conducted a deep research on appearance based person re-identification. The main contributions of our work are summarized as follows:1. Proposed person re-identification based on reference image feature learning. Based on human experiences, using image set can extract more reliable and comprehensive information for re-identification. Existing feature learning methods cannot fully exploit the visual information contained in the image set. In this paper, we propose to learn the most discriminative and reliable features from the image sets using max-margin criteria and reference image sets. The approach exploits region covariance matrix in the log-Euclidean space, and combines it with color histograms. The combined feature captures the statistical properties and chromatic information of each object, which is robust to low resolution, viewpoint changes and pose variations. Inspired by sparse encoding and prototype based feature representation, we propose to represent the given image set by explicitly modeling its relationship with the reference image set. The difference is treated as the discriminative features for the given object. These features are extracted from the decision boundary between the given image set and a reference image set supervised by max-margin criteria. Experiments conducted on benchmark datasets demonstrate promising performance of the approach.2. Proposed person re-identification method based on free energy score space feature mapping. Existing person re-identification methods mainly count on local and global image features to capture the visual information of human appearance. Local features can provide raw and basic cues with the body parts, while global features can provide the overall configuration of different body parts. After extracting local features, different strategies can be employed to combine them to capture different appearance information. In typical object recognition, Bag of Visual Words (BOV) model is an effective way to encode local features and has gained widely applications in image retrival and object recognition. The BOV model only encodes appearance frequencies of the visual words but not includes other statistics. Here in this paper we propose a re-identification method based on Free Energy Score Space (FESS) feature mapping. We use the Gaussian Mixture Model (GMM) to model the distribution of image features, and get the feature mapping by calculating FESS of the GMM. FESS feature maps of different image regions are encoded into a fixed-length feature vector for re-identification. This feature vector contains the average value, covariance and second order statistics of image features. Experimental results demonstrate promising performance of our method on challenging datasets.3. Proposed person re-identification method based on multi-view body part detection. In video surveillance scenarios, facial details and other reliable biometric cues may not be available. Human operators usually treat each human figure as an articulated body, and re-identify a previously seen object by focusing on similarities between the appearances of body parts. Based on this experience, we propose to exploit the human body structural constraints in re-identification. The Multi-view Pictorial Structures (MPS) model is used to estimate the pose and view-point of the object. After getting the body part configuration, we can extract appearance features in each part and applied these features in re-identification. We train 8 view-specific PS models using different kinematic tree priors and appearance models. Then we rely on the bank of view-specific models and train 8 viewpoint linear SVM classifiers on the outputs of the validation set. Using the MPS model, we can estimate the human body pose and its viewpoint simultaneously. The full object image is segmented into separate body part regions relying on the position and orientation of each body part. We encode such information in the form of separate body part masks and a combined full body mask. Having the masks, we extract gray-color histogram and maximally stable color region as the human signature. Experiments carried out on different datasets verify that MPS model can effectively estimate body pose and viewpoints, and show some robustness to pose variation and low resolution.