Abstract

In this paper a new robust method to estimate motion from a sequence of range images is developed. This approach uses the correspondences between planar surfaces in a sequence of images instead of using the usual method where point and line correspondences are used. Rotation and translation are both determined using the properties of the planar surfaces such as the surface normals and surface intersections. Because global features are used the method is less sensitive to noise, quantization errors and partial occlusion of the surfaces. Also, the method may be applied to a scene containing several objects, each with a different motion. The algorithm has been tested on sequences of synthetic and real laser range data.

© 1989 Optical Society of America