This project seeks to exploit the computational resources and parallelism offered by modern programmable GPUs in the context of computer vision.

We have been implementing a robust model-based three-dimensional tracking system on the GPU, operating at frame-rate on a humanoid robot during walking. Leveraging the computational resources of the GPU for vision has enabled us to ensure our tracker’s robustness to the large camera displacements, camera shake and severe aspect changes typically encountered during humanoid navigation.