Reseach On Person Re-identification Under Surveillance Video

With the widespread using of video surveillance, video monitoring system is playing an increasing role in keeping public safety and maintaining stability. For most of the monitoring systems, it is the human and their behavior on which the researchers concern and study focus. In the application of video detection, we use the detected information for further analysis which concluded by people detection, recognition, tracking and analysis of their specific behavior in the video monitoring image. In video surveillance, the key problem is the pedestrian re-identification technology. It refers to the judgment of the pedestrians under different camera whether they belong to a pedestrian. As the increasing demand for social application and development of computer vision, the problem of pedestrian re-identification is gradually developing into a hotspot. Based on the pedestrian characteristics under surveillance cameras, this thesis proposes several aspects of re-identification algorithms based on appearance characteristics. The main work is as follows:First of all, this thesis proposes a pedestrian re-identification algorithm based on the super-pixel segmentation. After dealing the detected images, the images of pedestrians are divided into super-pixels. The color features of each super-pixels will be extracted. According to the clustering center of the prior training pedestrian color features, the nearest clustering center to each color feature among these super-pixels can be calculated. Counting the clustering centers' histogram of each super-pixels as the feature of the superpixel. In order to match different pedestrian, this thesis presents a algorithm of superpixels projection matching, which focuses on solving the matching problem among pedestrians.Secondly, this thesis establishes a pedestrian re-identification system framework under a single video monitoring in different time and different distance and proposes a recognition algorithm based on pedestrian's components segmentation. After training the detectors of the pedestrian's different components, the color and texture features of each component are extracted, cascading up the multiple color space as the color feature, making LBP(local binary patterns) local texture feature into a fusion. As a result, the matching problem between the pedestrian is converted into a matching problem between parts of a pedestrian.Eventually, this thesis establishes a pedestrian re-identification system framework under different cameras in the same period but different angles and puts forward a pedestrians' matching method based on different pedestrian area and feature clustering. The different pedestrians' regions are divided into small pieces of same size. After extracting the mean color of each piece, clusters all of the mean color in corresponding areas between the pedestrians and the target pedestrians. Combined the extracted color histogram with the characteristics of SIFT(scale invariant feature transform), the distance of the corresponding pedestrian area could be calculated to achieve pedestrian matching.