Automating point cloud acquisition methods for scene understanding

Abstract

Advances in technology during the years have allowed three-dimensional data acquisition systems to become more accessible increasing interest in possible applications like object recognition. However, obtaining such datasets can prove to be costly and time consuming. This dissertation proposes a solution by simulating the acquisition process which could alleviate some of the drawbacks of 3D scanning. The simulation will allow any user to generate and use three dimensional data from virtual scenes. No scanning equipment would be necessary to acquire point clouds. The simulation allows users to sample a given scene autonomously. Different parameters can be adjusted in order to create the desired data set. The point clouds generated by the simulation resemble those acquired by real 3D scanners. A set of point clouds generated by the simulation were compared to point clouds obtained from a real scanner. The noise function used for the simulation proved to be reasonably effective in comparison. A heuristic based approach was used to traverse a scene and sample the objects present. A measure is provided to check the amount of coverage that a particular scan obtained from a scene. A set of point clouds were generated using the simulation and processed using algorithms commonly found in scene understanding such as normal estimation. The generated data sets could be used to identify strengths and weakness of these algorithms.