Adaptive image segmentation based on color clustering for person re-identification

Person re-identification refers to identifying a particular person who has occurred in the monitoring network automatically by computer in the surveillance video, which is significantly important for the improvement of intelligence of video monitoring. But the research of person re-identification is not mature and facing many challenges. The following factors may lead to a certain difference for the same person in different monitoring video images; for example: the illumination changing in the monitoring environment, the shooting angle difference, and the posture difference. These may lead to low recognition accuracy. In this paper, a new appearance-based person re-identification method was proposed and an investigation was launched on the following topics for the improvement of recognition accuracy. First, a simple and feasible method for color invariants was proposed, so that the affection of color by change of illumination and shooting angle could be eliminated. Then, a highly adaptive image segmentation method based on color clustering and a color feature representation scheme for specific color characteristics were designed, which could help to extract color features in more reasonable points. Finally, an effective similarity measure criterion was obtained through QSF measure learning, which could ensure that the different pedestrians can be distinguished and was better able to capture the visual change of the same person. In addition, the traditional evolutionary algorithm was improved and applied to the process of iterative computation for QSF. The experimental results show that our method is an effective way for the person re-identification problem.