Three-dimensional scene analysis using multiple range finders — Data capture, coordinate transformations and initial segmentation

Abstract

A system for acquiring a complete three-dimensional surface description of a complex scene using multiple rangefinders is described. These scenes may contain multiple free-form objects with both convex and concave surface elements. A unique approach for registering each of the views is also proposed. A diamond shaped model with known geometric properties is used to register each of the six views. Range data acquired from each of the views is transformed to match the model using affine transforms that are calculated using a least squares minimisation approach. Once the transformation is known, it can be used for all successive range data acquisition. This method is accurate, requires very little computation time, and is invariant to initial conditions. An efficient method for calculating surface normals at each data point is also described. It makes use of the rectangular grid of points in the image from which each three-dimensional data set is extracted. Low level processing of the registered data using the shared nearest neighbour clustering technique is also described. This provides a very rich set of semantic free partitions upon which to base higher level processing. This will ultimately lead to complete scene descriptions, including object identity, pose and juxtaposition, suitable for supporting intelligent robotic manipulation.