Research On Person Re-identification

Person re-identification is an important research point in the field of video analysis and understanding. It has great value in intelligent surveillance, motion sensing game and so on. The performance of person re-identification has been improved by researchers' hard work, which made proposed algorithms capable for the real scenarios. Due to the human pose variety, background clutter, illumination change, viewpoint transformation and so on, the appearance of human changed a lot, which brings big challenges in solving the problem of person re-identification.To solve the problems, two useful methods are proposed in this thesis. Such as, extract Gaussian local features based on human components analysis. Specifically, the main contributions of this thesis are summarized as follows:Proposed a Gaussian local descriptor based on human components analysis. Aimed at feature extraction in Person Re-identification, we propose a Gaussian local descriptor based on human components analysis(Ga LF). Through the human component analysis and color correction, GaLF extract each part of the picture by the local descriptor of Gaussian model. Then by measuring the Gaussian distance of each part, GaLF get the measurement of human body based on multiple parts, namely the similarity between two samples. The experiments show that, GaLF use the body parts information effectively and extract color, texture and spatial information robustly, achieves much higher performance than that state of the art method on VIPeR and i-LIDS pedestrian re-identification dataset.Proposed an object and attribute classification system for person re-identification. This system uses the semi-supervised method for screening a large number of network data. By using the deep neural network model to annotate the network data, for extending annotated dataset. In order to enhance the accuracy of this system, the deep neural network is retrained through the expanded training dataset. After long time running, the system can extract the high order semantic features of pedestrian that can be used for person re-identification. The system is capable of continuous operation, combined the never-ending image learning system with deep neural network, leveraging their comparative advantages. Never-ending image learning system performs poor in classification, which is the strong point of R-CNN method; But R-CNN method needs lots of labeled data for training, and the annotated data is less and less, that can be supported by the ELDA for extending dataset.