Combining View-Based and Model-Based Tracking of Articulated Human Movements

Many existing systems for 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. Recognition of body postures can be based on simple image descriptors, like the moments of body silhouettes. We present a system that combines these two approaches within a common closed-loop architecture. Central characteristics of our system are: (1) Mapping of image features into a posture space with reduced dimensionality by learning one-to-many mappings from training data by a set of parallel SVM regressions. (2) Selection of the relevant regression hypotheses by a competitive particle filter that is defined over a low-dimensional hidden state space. (3) The recognized postures are used as priors to initialize and support classical model-based tracking using a flexible articulated 2D model that is driven by local image features using a vector field approach. We present pose tracking and reconstruction results based on a combination of view-based and model-based tracking. Increased robustness and improved generalization properties are achieved even for small amounts of training data.