Waterfront surveillance and trackability

This paper presents a method for waterfront surveillance system. Unlike traditional approaches that model dynamic water background explicitly, we choose a relaxed background model to extract multiple object hypotheses. The hypotheses are then tracked with probablistic framework. Finally, the hypotheses are classified as positive objects or negative objects based on their trackability . Trackability is described by the stableness and the consistency of their trajectories and their appearances, and the properties of their accumulated templates.