Authors
SOUSA, E. V., LUCIO, D., FERNANDES, L. A. F., VELHO, L.
Abstract
The automatic detection of planes in depth images plays an important role in computer vision. Plane detection from unorganized point clouds usually requires complex data structures to pre-organize the points. On the other hand, existing detection approaches tailored to depth images use the structure of the image and the 2.5-D projection of the scene to simplify the task. However, they are sensitive to noise and to discontinuities caused by occlusion. We present a real-time deterministic technique for plane detection in depth images that uses an implicit quadtree to identify clusters of approximately coplanar points in the 2.5-D space. The detection is performed by an efficient Hough-transform voting scheme that models the uncertainty associated with the best-fitting plane with respect to each cluster as a Gaussian distribution. Experiments shows that our approach is fast, scalable, and robust even in the presence of noise, partial occlusion, and discontinuities.