Mobile AR Point Cloud Matching

Abstract

This thesis explains the process of point cloud matching, to assess its viability in mobile markerless augmented reality solutions. Traditional point cloud matching algorithms like 4-Point Congruent Systems are described and considered. 2 Point Normal Sets was chosen for real-world experimentation due to being faster and easier to implement. Additional ideas to improve performance were implemented in addition to the chosen algorithm: A pair limiter that only uses up to 100 pairs each iteration for the matching process, changing the algorithm to only perform rotations across one axis and a point sampler to reduce the amount of points used. The best values for rotation error delta were also analyzed. The first experiment done was with a point cloud in the shape of a rectangular box, to confirm if the additional ideas did improve performance. For the second experiment, various point cloud scans were made of a real-world room in Gallery Pallas. The algorithm was tested using these scans and the experiment results were documented. For the third experiment, an additional limiter was added to the point cloud scanner, so that it would only accept points closer than 2 meters from the scanner. This was done for more accurate point clouds. A second set of tests were done with this modification and the experiment results written. The fourth and final experiment was done on a mobile device. While the previous experiments showed promise and improvements with each advancement, the algorithm had problems on a mobile device when matching against a unique point cloud for each run. Since there were still good matches, the possibility of using point cloud matching for markerless augmented reality is there, but would require additional work before it can be considered usable in applications.