Robust real-time stereo-based markerless human motion capture

Abstract

The main problem of markerless human motion capture is the high-dimensional search space. Tracking approaches therefore utilize temporal information and rely on the pose differences between consecutive frames being small. Typically, systems using a pure tracking approach are sensitive to fast movements or require high frame rates, respectively. However, on the other hand, the complexity of the problem does not allow real-time processing at such high frame rates. Furthermore, pure tracking approaches often only recover by chance once tracking has got lost. In this paper, we present a novel approach building on top of a particle filtering framework that combines an edge cue and 3D hand/head tracking in a distance cue for human upper body tracking, as proposed in our earlier work. To overcome the mentioned deficiencies, the solutions of an inverse kinematics problem for a - in the context of the problem - redundant arm model are incorporated into the sampling of particles in a simplified annealed particle filter. Furthermore, a prioritized fusion method and adaptive shoulder positions are introduced in order to allow proper model alignment and therefore smooth tracking. Results of real-world experiments show that the proposed system is capable of robust online tracking of 3D human motion at a frame rate of 15 Hz. Initialization is accomplished automatically.