high performance computing on graphics processing units: hgpu.org

In computational vision, visual tracking remains one of the most challenging problems due to noise, clutter, occlusion, and dynamic scenes. No one technique has yet managed to solve this problem completely, but those that employ control- theoretic filtering techniques have proven to be quite successful. In this work, we extend one such technique by Niethammer et al. in which implicitly represented dynamically evolving contours are filtered using a geometric observer framework. The effectiveness of the observer hangs upon the solution of two major problems: (1) the calculation of accurate curve velocities and (2) the determination of diffeomorphic correspondence maps between curves for geometric interpolation. We propose the use of novel image registration techniques such as image warping and optimal mass transport for the solution of these problems which increase the performance of the framework and reduce algorithmic complexity. One major drawback to the original scheme, as it relies on PDE solutions, is its computational burden restricting it from real time use. We show that the framework can, in fact, run in near real time by implementing our additions to the framework on the graphics processing unit (GPU) and show better execution times for these algorithms than reported in recent literature.