People Matching In Multiple Surveillance Videos Based On VSD

Traditional methods of people tracking in videos are mostly through analyzing pixel-level features to detect, recognize and even tracking pedestrian. The successful people matching rate is relatively low. In this thesis, a method of people matching based on Video Structured Description(VSD) is proposed. Through matching of people’s static features supplementing by dynamic features analysis, not only people tracking accuracy could be increased, but also a text-based search of videos could be made possible. A description scheme designed specific for people tracking in multiple surveillance videos is proposed, including static features and their standard descriptions, dynamic features and event descriptions. A set of matching rules for static features are also proposed, including major factors such as error thresholds and weights. Experiment is carried out using actual surveillance videos from a library. The videos are described with the proposed standard static feature descriptions and, statistical data are collected for the recognition quality of different features in different backgrounds. Based on the analysis of the experimental results, references are proposed for setting the weights and error thresholds for static features matching.