Stereo Vision Head Vergence using GPU Cepstral Filtering

Vergence ability is an important visual behavior observed on living creatures when they use vision to interact with the environment. The notion of active observer is equally useful for robotic vision systems on tasks like object tracking, fixation and 3D environment structure recovery. Humanoid robotics are a potential playground for such behaviors. This paper describes the implementation of a real time binocular vergence behavior using cepstral filtering to estimate stereo disparities. By implementing the cepstral filter on a graphics processing unit (GPU) using Compute Unified Device Architecture (CUDA) we demonstrate that robust parallel algorithms that used to require dedicated hardware are now available on common computers. The cepstral filtering algorithm speed up is more than sixteen times than on a current CPU. The overall system is implemented in the binocular vision system IMPEP (IMPEP Integrated Multimodal Perception Experimental Platform) to illustrate the system performance experimentally.