Unsupervised Pedestrian Re-identification for Loitering Detection

This paper presents a framework of detecting loitering pedestrians in a video surveillance system. First, to represent pedestrians an appearance feature which contains geometric information and color structure is proposed. After feature extraction, pedestrians are tracked by a proposed Bayesian-based appearance tracker. The tracker takes the advantage of Bayesian decision to associate the detected pedestrians according to their color appearances and spatial location among consecutive frames. The pedestrian’s appearance is modeled as a multivariate normal distribution and recorded in a pedestrian database. The database also records time stamps when the pedestrian appears as an appearing history. Therefore, even though the pedestrian leaves and returns to the scene, he/she can still be re-identified as a loitering suspect. However, a critical threshold which determines whether two appearances are associated or not is needed to be set. Thus we propose a method to learn the associating threshold by observing two specific events from on-line video. A 10-minute video about three loitering pedestrians is used to test the proposed system. They are successfully detected and recognized from other passing-by pedestrians.