3D object tracking with a high-resolution GPU based real-time stereo
Ratheesh Kalarot, Georgy Gimel’farb, John M. Morris · 2012
Stereo correspondence algorithms, which are fast enough for real-time use, require hardware assistance and inevitably trade some matching accuracy for speed. A cloud of 3D points thus produced by our previously reported GPU accelerated implementation of a dynamic programming correspondence algorithm is noisy and contains artifacts, which hinder tracking accuracy. We have augmented this implementation with modules for re-projection and filtering. A fast clustering procedure based upon a set of simple volume rules identifies candidate objects. An opportunistic tagging system tracks objects through occlusions. Kalman filtering predicts positions in the next frame. These steps reduce the effects of dynamic programming streaks in the depth maps. Experiments with synthetic and real-world video sequences confirmed the accuracy in tracking multiple objects (e.g. humans) in various environments.