Initial pose estimation using cross-section contours

This paper presents a means to approximate an object's pose, suitable for initialization of the Iterative Closest Point algorithm. The class of problems considered is objects lying stably on a planar surface, for which a relatively small number of pose types are possible. Within each pose type, the registration problem is reduced to 3 dimensions. Using contours computed from horizontal slices, it is shown that relatively noisy point-cloud samples can yield good estimates of pose. Experimental results using an Atlas robot are presented. The proposed method offers an efficient means to initialize point-cloud fitting, resulting in faster, more reliable convergence.