A Novel Object Segmentation Method for Silhouette Tracker in Video Surveillance Application

In recent years, surveillance cameras are deployed almost everywhere. More and more video analytics features have been developed and incorporated with video surveillance system for conducting intelligence tasks, such as motion detection, human identification, etc. One typical requirement is to track suspicious humans or vehicles in the cameras' live or recorded footages, and over the years researchers have proposed different tracking methods, such as point tracking, kernel tracking and silhouette tracking to support this requirement. In particular, silhouette tracker has received considerable attention because it works well for objects with a large variety of shape, provided that reasonably good object masks or contours are initialized properly for the silhouette tracker. A properly initialized object mask and contour, however, cannot be obtained easily. On one hand, a simple bounding box contains too much irrelevant background objects, while a manually specified mask could provide accurate silhouette but this also requires lots of interactive which greatly limits its practicality. In this paper, we present a novel block based object mask segmentation method for silhouette tracker initialization. Essentially, the proposed method re-uses the motion information extracted during the video encoding phase, which provides approximated object masks for silhouette tracker. Experimental results confirm that such a block-based object masks is sufficient for a robust silhouette tracker to reliably track moving objects.