Markerless human motion capture using charting and manifold constrained particle swarm optimisation

We present a framework for markerless articulated human motion tracking in multi-view sequences. We learn motion models of common actions in a low-dimensional latent space using charting, a nonlinear dimensionality reduction tool which estimates automatically the dimension of the latent space and keeps similar poses close together in it. Tracking takes place in the latent space using an efficient modified version of particle swarm optimisation. Observed image data are multiview silhouettes, represented by vector-quantized shape contexts. Multi-variate relevance vector machines are used to learn the mapping from the action manifold to the shape space. Tracking results with walking, punching, posing and praying sequences demonstrate the good accuracy and performance of our approach.