Efficient Object Localization and Scene Registration Using Contour-Enhanced Point Cloud Resampling Techniques

Abstract

In this thesis, I present an efficient implementation of contour-enhanced graph-based resampling of point clouds. I develop both color-blind and color-aware versions of this technique, which I then apply to the problems of object localization and scene registration. In object localization, use of contour-enhanced resampling prior to correspondence-based transformation estimation can reduce calculation runtime by more than 67% while maintaining a similar accuracy to a voxel-grid resampling method. When given a similar amount of time to run, contour-enhanced resampling can reduce translation error of localization by more than 50% over voxel-grid resampling. The contour-enhanced technique also has similar effects in scene recognition, and is able to alternatively either lower runtime while maintaining accuracy or improve accuracy while maintaining runtime, though these benefits tend to diminish as the amount of downsampling is increased. Ultimately, by being able to effectively downsample point clouds to their important contour points, the contour-enhanced graph-based resampling technique presented in this thesis is able to effectively reduce runtime and error in downstream processing, making it a powerful tool within the field of computer vision.