| Abstract: |  Point cloud registration is an important process in mobile robotics, serving as the cornerstone of simultaneous localization and mapping. The contribution of our work is twofold: firstly, we compare local registration methods using high-quality datasets and a custom protocol. In terms of precision and robustness to initial pose displacement, the capabilities of the methods are explored in an unprecedented detail, overcoming any previous work that we know of. Secondly, we propose enhancements to a global, feature-based registration method that take advantage of visual information, specifically camera imagery. Proposed changes include an extension of the feature descriptor, and a modification of reference frame determination. To investigate the modified methods, a dataset containing visual data is created. Experimental results indicate a significant improvement over the original method.
|
|---|