Grant: Discovery Grants Program - Individual

John Barron (and his colleagues and students) are principally interested in measuring 2D and 3D image motion (the variants are called optical flow, scene flow and range flow) in 2D and 3D monocular and binocular image sequences. Some issues they are addressing include:(1) Image noise is a continuing source of error in any flow calculation. We are devising theoretical and practical ways to measure local (Gaussian) image noise and using this information in flow calculations to increase the accuracy of computed flow while maintaining its density. (2) Discontinuities in 2D and 3D optical flow fields caused by occlusion are currently handled poorly by state-of-the-art optical flow algorithms and yet this type of flow is fundamental in human vision. We are designing new occlusion boundary-based optical flow algorithms, which will yield good quantitative and qualitative flow at such boundaries.(3) We are measuring scene and range flow in the synthetic Auckland and real Daimler stereo driving sequences using scene flow and range flow. Scene flow uses stereo disparity maps (and its gradient field) and left and right image optical flow while range flow uses stereo depth maps and their spatio-temporal derivative maps. Both provide estimates of 3D camera motion at each visible scene point. Which is best and why? The answer will lead to better (hybrid) scene/range flow algorithms. Scene/range flow will also greatly benefit from our image noise and occlusion work.Our ultimate goal is to produce robust, accurate and fast scene/range flow algorithms. We plan to validate them using the real-time stereo depth maps produced from Steven Beauchemin's ``Roadlab'' car (outfitted with stereo cameras and a state-of-the-art multiprocessor computer). Our work will greatly enhance Computer Vision based driving aids.