High Speed Articulated Object Tracking Using GPUs: A Particle Filter Approach

This paper presents a novel application of the GPU processing power to a very computationally demanding articulated human body tracking problem in a view-based approach. This work includes some optimizations at the algorithmic level as well as some tricks at the implementation level using OpenGL and shader programming. An underlying particle filter framework is combined with a novel particle weight computation, where heterogeneous templates are considered for distribution mode recovering. Also, a form of elitism is taking into account to prevent flickering when the best candidate of the particle population is chosen. Impressive performance up to 317-713 frames per second is guaranteed for common configurations of about 1024-256 particles and 320×240 video resolutions.