3D Scene Reconstruction using a PMD range camera

Abstract

This internship thesis proposes an algorithm to solve a 3D indoor scene reconstruction problem using a PMD range camera. It explains in detail the image registration step using the Iterative Closest Method (ICP) method. Extensive literatures have been reviewed on different variations and implementations of the ICP algorithms. The Singular Value Decomposition (SVD) based ICP is selected to perform the registration task. To construct the indoor scene, the PMD camera first scans through the indoor environment continuously and outputs hundreds of frames. These frames are then made to go through data calibration, a step to fill in the invalid z coordinate values of the frames. After this, these frames are input into an integrated algorithm. The integrated algorithm first selects the frames at a regular interval, then performs pair-wise registration to stitch the images into a complete scene. A sub-sampling step is followed to reduce the number of data points in the total image. As the last step, a Delaunay triangulation is performed on the point cloud. It joins the points using triangles and constructs the surface. The last step compensates the effect of limited data points and restores the shape of the original object. Tests using both the simulated data and the real data have been performed to evaluate the effectiveness of the algorithm. Some of the test results are presented in the report. Last but not least, recommendations of future work are listed.