Non-Linearly Embedded Visual Tracking

Abstract. Many difficult visual problems like monocular human tracking require complex heuristic generative models defined over high-dimensional parameter spaces. Despite their successes, optimization with such models remains notoriously complex due to the difficulty of flexibly using prior knowledge in order to reshape an initially designed representation space. Non-linearities, inherent sparsity of high-dimensional training sets and lack of global continuity makes dimensionality reduction challenging and low-dimensional search inefficient. To address these problems, we present a sampling-based optimization framework that restricts tracking to low-dimensional spaces via non-linear embedding. The formulation leads to a layered generative model where global continuous optimization over the embedded manifold is made possible. Our prior flattening method allows a simple analytic treatment of boundary and manifold intrinsic curvature constraints and allows consistent iterative and closed-form solutions for embedded geodesic and sequence smoothing calculations. We analyze the structure of reduced manifold representations for a variety of human interaction activities and demonstrate that the approach gives accurate tracking and reconstruction of fast self-occluded motion in cluttered monocular video.