Probabilistic classification between foreground objects and background

Abstract

Tracking of deformable objects like humans is a basic operation in many surveillance applications. Objects are detected as they enter the field of view of the camera and they are then tracked during the time they are visible. A problem with tracking deformable objects is that the shape of the object should be re-estimated for each frame. We propose a probabilistic framework combining object detection, tracking and shape deformation. We make use of the probabilities that a pixel belongs to the background, a new object or any of the known objects. Instead of using arbitrary thresholds for deciding to which class the pixel should be assigned we assign the pixel based on the Bayes criterion. Preliminary experiments show the classification error drops to about half the error of traditional approaches.