Automatic Extraction of Planar Projections from Panoramic Range Images Abstract This paper presents a segmentation technique to decompose automatically a panoramic range image into a set of planar projections. It consists of three stages. Firstly, two orthogonal surface orientation histograms are generated. Secondly, from these histograms the major surfaces’ orientations are extracted. Finally, a histogram of distances is computed for each one of these orienta- tions; it will be used to define the position of the projection planes as well as the corresponding clipping planes. The original panoramic range image is divided into as many planar projections as main directions in the orientation histograms are extracted. This technique can be used with both indoor and outdoor scenes. Experimen- tal result with a panoramic range image is presented. 1. Introduction During the past years 3D computer vision has experi- enced a fast growth. The appearance of new sensors, which allow the obtainment of a huge amount of three- dimensional information in a short time, and the need to process and represent these images efficiently, has given rise to new research topics in the 3D computer vision community. One of these topics is the 3D digital repre- sentation that has gained an important place in different fields (e.g. architectural [1], [2], [3], automotive engi- neering [4], robotics [5], computer animations [6], to mention a few). However, the gap existing between 3D digital representations and the classical planar drawings need to be saved—considering that planar drawings are the “de facto” representations for many applications. Planar drawings are still nowadays used because of their portability. In industrial environments or building works is not easy to find the appropriate tools to visualize or show 3D digital representations. In this sense, this paper presents a technique to generate automatically pla- nar projections from 3D panoramic images. Up to our knowledge, [7] is one of the first works taking the prob- lem of panoramic range image decomposition. In [7], panoramic range images of forest scenes have been inves- tigated. In the current work indoor and outdoor scenes, scanned by using the Imager 530 scanner, developed by the Z+F company, have been tested. This sensor allows a scanning rate up to 625,000 points per second with a pan- oramic field of view in the horizontal direction and a field of view of 135 degrees in the vertical direction (more technical details about the Imager 530 scanner are given at the company’s Web page: www.zf-uk.com). The Imager 530 scanner allows capturing the full geometry of big environments, with a high fidelity, in a short time. The required space to store all this information (images bigger than 500MB) or the CPU power to process all these data are not a problem for the current technol- ogy. The only constrain for these panoramic range images appears when it is necessary to print or represent all these 3D data in a single snapshot. The envisaged solution is the segmentation of the original panoramic range image into a set of easy to understand planar representations. Planar drawings can be used not only as a final repre- sentation but they can also be used to define the next position of the sensor. This problem, known in the litera- ture as the next-best-view problem [5][8], consists in computing the positions where the range sensor should be placed in order to acquire the surfaces of the objects present in a scene minimizing the total amount of scans. The computed planar representations can be used to detect occluded areas or low resolution areas, defining thus the next position of the sensor. The proposed technique consists of three stages. Firstly, two surface orientation histograms are computed; they will unveil the major surfaces’ orientations of the given panoramic scene. Secondly, main directions are extracted by combining local maximum from the previous The author has been supported by The Ramón y Cajal Program Angel D. Sappa Computer Vison Center Edifici O, Campus UAB 08193 Bellaterra - Barcelona, Spain angel.sappa@cvc.uab.es Proceedings of the 2nd Intl Symp 3D Data Processing, Visualization, and Transmission (3DPVT 2004) 0-7695-2223-8/04 $20.00 © 2004 IEEE