A multi-staged system for efficient visual person reidentication

An important field in today's computer vision is person centric video analysis. The basis of this person centric analysis is the detection and tracking of people in video data. In many cases it is not sufficient to track people when they continuously appear in the camera's field of view, but to also reacquire a track after a person has left a field of view and reenters it. In this paper, we introduce a technique that conducts this person reidentification based on SIFT features only. This approach fits into an Implicit Shape Model (ISM) based person tracking approach by employing the SIFT features collected during tracking for reidentification. The ISM characteristics of a person are used to perform reidentification in an efficient 3-staged approach which combines computation efficiency with high distinctiveness. The evaluation is performed in an open-set classification approach on a public dataset of 60 persons which was acquired with a thermal camera. Despite the challenges of person reidentification in thermal imagery, the approach shows nearly perfect performance and outperforms other reidentification approaches on this dataset.