In this talk, I present methods for fitting local planar surface elements, that we call patchlets, to 3D data obtained from correlation stereo images. We use these patchlets to robustly extract and estimate bounded planar surfaces from complex and noisy stereo scenes. The patchlet element is a small planar surface the size of a pixel projected onto a world surface. It has a position, a normal direction, a size, and confidence metrics on its position and orientation. The confidence metrics are generated from the accuracies of the stereo vision system, which are propagated to 3D point data and then to the patchlet parameters. The patchlets are used to extract larger bounded planar surfaces that are useful for environment modeling. We use a region-growing approach to identify how many surfaces exist in a stereo image and an initial estimate of the surface parameters. We then use Expectation-Maximisation (EM) to refine these surface parameters to an optimal estimate using a probability maximisation approach. The confidence metrics of the patchlet parameters allow proper weighting of patchlet contributions to the probability maximisation solution.
We verify by experimental means the accuracy of correlation stereo matching and demonstrate that the patchlet confidence metrics obtained from that accuracy fit expected normal distributions. We compare the results of the patchlet-based surface segmentation to manually constructed ground-truth segmentation and find the segmentation accuracy for a scene ranged from 82% to 93%. We also present a method for filtering noise from correlation stereo disparity images that is highly successful.
Back to the LCI Forum page