Solving partial occlusion with range images and 3D histograms

In this paper, a complete experimental set-up for range finding and the software developed for its control is presented. The results obtained are used not only to provide numerical information of the objects´ pose in a scene to a high level planner of a robotic system, but also to identify them. For the special case of object partial occlusion, a method for identification has been developed based on the range image obtained from the range finder: a complete scan of the scene is performed at its highest resolution, followed by several levels of image filtering, in order to obtain an image of distances-the range image. The method consists primarily of segmenting a range image once, and searching the special case of partial occlusion. Then, for those range blobs found to contain more than one object, a second segmentation is performed based on the 3D histogram (number of pixels versus height Z) which is built during the first segmentation. As long as the overlapping objects meet certain conditions, separation of occluded and occluding objects is obtained. The characteristic parameters defining the latter are used to identify it after matching them to a pre-existing models database