A System for Reconstruction from Point Clouds in 3D: Simplification and Mesh Representation Lyuba Alboul and Georgios Chliveros Centre for Automation and Robotics Research, MERI Sheffield Hallam University Sheffield S1 1WB, UK e-mail: l.alboul@shu.ac.uk Abstract—In this paper we present a complete system for acquisition of fused (textured) point clouds in 3D, from a Laser Range Finder (LRF) and a CCD camera. Furthermore, we describe an approach to build and process the resulting models, including their pre-processing and mesh simplification. This approach allows manipulating the resulting data structure into consistent geometric representations, which can be further adapted based on user requirements. The advantage of our system is that of low computational cost, ease of use and accuracy in the representation of the environment, even without prior data smoothing. Index Terms—Point cloud acquisition, mesh reconstruction, curvature criteria I. I NTRODUCTION We present a system for object reconstruction from point clouds in 3D developed within the remits of the VIEWFINDER project 1 for indoor search and rescue operations. However, our system can be used for other applications; where an object needs to be acquired and economically represented while maintaining its main geometric features. In search and rescue robotics user requirement documents refer to an inherent need to include remote exploration of wreckage or altered interiors of partially collapsed buildings. In VIEWFINDER, the user requirements are similar to those detailed at NIST [1], with respect to the overall solution implementation [2]. The project deploys one outdoor and one indoor robot in order to explore scenes after an incident has occurred. The aim is to assess whether the environment is safe for human Rescue Teams’ operations. Range image analysis is a term used to describe the fusion of unregistered camera images (array of RGB values) and laser range finder (LRF; also referred to as LiDAR for large distances) point clouds in 3D. The resulting data are range values combined with visual information (e.g. RGB texture) in order to enhance perception. An important part of such a fusion procedure is the extrinsic calibration of LRF with respect to camera, which has been addressed in recent years albeit by only a few researchers. There are schemes that require user input (for feature correspondence, [3]); automatic, by adapting calibration using standard checkerboard patterns calibration [4], [5]; or via additional sensory information [6]. The approach described in this paper can be considered as 1 VF official page: https://view-finder-project.eu merging of extrinsic and intrinsic calibration, as only one acquisition of the scene is sufficient for calibration. However, the focus of the paper is on the acquired data processing, which begins with a structured data representation. This representation allows for the identification of points of interest, outliers, ‘holes’ and can be further simplified by preserving the major features of the scene. Finally, using our method computationally fast mesh construction and processing are possible, without the use of Graphical Processing Unit hardware / software interplay [7]. A. Notes on 3D scene reconstruction One of the common problems in surface and object recon- struction, is that the resulting mesh (polygonal surface or poly- hedral object) often has unexpected holes, self-intersections and other artifacts, which are very difficult to track and correct. If for a human eye such a mesh might still be acceptable, in many applications, such as in robotics, where a machine needs to work with the mesh without human intervention, it is desirable or even a must that a mesh has easily recognizable features. It is also desirable to reduce such features to mini- mum, in order to facilitate further processing, but so that the most important characteristics of the object are preserved. Any object can be described by four parameters: shape, size, texture and colour. Colour and texture are basic features used in many visual processing applications whereas spatial data obtained by measurements are processed with the use of geometric concepts. Shape is the most crucial characteristic of any three-dimensional object. The intuitive notion of shape is formalized by means of geometry, namely by curvature. The 3D scene reconstruction process in our system consists of the following steps: • Fusion of the measurement data obtained by LRF and camera image, and formatting the data in a specific type. • Constructing the mesh (triangulated surface) that span the data. As the data are structured, the resulting mesh allows identification of ‘holes’ in the mesh, and outliers. • The mesh is then further processed. Among the methods that can be used are smoothing and decimation based on curvature criteria. In general, the mesh is without self-intersections, by con- struction, but if these occur, they can be easily identified. In general their presence indicates that the acquisition or LRF 978-1-4244-7815-6/10/$26.00 c 2010 IEEE ICARCV2010 2010 11th Int. Conf. Control, Automation, Robotics and Vision Singapore, 7-10th December 2010 2301