Técnica de compressão progressiva de nuvens de pontos e projeto de um scanner 3D de baixo custo

Abstract

Three-dimensional (3D) scanning systems can be used in a wide range of applications such as in cultural heritage, reverse engineering and computer graphics. In general, most of the existing 3D scanning systems are based on the analysis of light patterns captured as images of a specific surface using a charge-coupled device (CCD) sensor in order to perform the 3D mapping – operation used to convert two-dimensional (2D) pixels into 3D points –, whose result is represented as a point cloud. Most of the 3D scanners available in the market are designed for high resolution and precision. This yields 3D models with a large amount of information and an increase of the final product cost, since such characteristics require higher value-added equipments. Therefore, as first contribution of this master’s thesis, is shown the implementation of a low-cost 3D scanner based on the laser triangulation technique using only a camera and a line laser generator. A geometrical-based model is proposed to perform the 3D mapping, whose calibration is based on the use of an object with known real dimensions. Despite the increasing computer’s processing and storage capacity, the usage of 3D models can become a problem in applications for remote transmission of data, real time processing or in embedded systems with limited computational resources. In this way, as second contribution of this master’s thesis, a progressive compression technique of point clouds is proposed based on the JPEG image compression technique and cylindrical coordinates. The proposed compression technique allows to decrease the computational cost associated to the storage process of 3D data. As main results, the proposed 3D scanning system was implemented with a total cost equal to US$ 77.00 and it showed up capable of generating models of real objects with differences in the dimensions smaller than 2 millimeters. Additionally, the proposed compression technique generated computationally eficient results, overcoming other relevant techniques based on octree space decomposition when analyzing the rate-distortion (R-D) performance.