Abstract
Computer vision-based markerless human body tracking plays an important role in e.g. surveillance and driver assistance applications, and in computer graphis, e.g. for retargeting in computer animation.
Many existing systems for computer vision-based human body tracking are based on dynamic model-based tracking that is driven by local image features. Alternatively, within a view-based approach, tracking of humans can be accomplished by the learning-based recognition of characteristic body postures which define the spatial positions of interesting points on the human body.
I will present a learning-based computer vision system that combines these two approaches into a common closed-loop tracking architecture. Central system characteristics are
1. one-to-many mappings between image features and body postures
2. a flexible 2D graph model
3. automatic 2D model-based registration and tracking using a vector field approach based on partial differential equations with symmetry constraints
4.a dynamic particle filter that integrates image features and prior information from a synthesis model
It is demonstrated that the combination of view-based and model-based tracking results in increased robustness and improved generalization properties. Specifically, the system generalizes over different individuals for similar action categories at similar view points from small amounts of training data.
