Scene Correlation Learning by Event Co-occurrence Modeling for Camera Network

Learning the scene correlation of uncalibrated static cameras is in increasing demand for intelligent surveillance system, such as making a inference of topology, or allocating computation resources for video retrieval in multi-robot system. However, many existing approaches learn the scene correlation among camera views by camera calibration or tracking targets across cameras. They seldom learn the scene correlation among cameras with co-occurrence analysis. In this paper, we propose a novel approach based on event co-occurrence modeling to learn scene correlation between camera views, which automatically forms the visual attention cross a number of camera views in case that the cameras are not calibrated. Firstly, motion based co-occurrence modeling applies spatio-temporal motion frequency representation (STMFR) to analyze correlation of motion patterns between two cameras. Then, the content based appearance modeling is put forward to represent high level appearance co-occurrence, which can be combined with low level feature to make a correlation inference. The method shows its effectiveness in the PKU-SES intelligent system, which is a multi-camera system including two sites in the campus, totally 10 cameras in realtime video surveillance.