Real-time stereo tracking of multiple moving heads

Abstract

Tracking a number of persons moving in a cluttered scene with occlusions is an important issue in computer vision. In this paper we present RealTrack, a system for simultaneously tracking of multiple moving heads in real time for real world applications. It leverages on depth information to alleviate the influence of shadows and to disambiguate occlusions by depth ordering. Augmented by rotation insensitive head-shoulder contour models together with an adaptive tracking algorithm, our system can robustly track people in various conditions such as crossing, gathering, scattering, and re-appearing in a direction different from the moving direction before occlusion. We also employ dynamic background update method, which makes our system suitable for long time surveillance under slow lighting changes and disturbances such as objects going in and out of the scene.