Sampling in image space for vision based SLAM
Olaf Booij, Zoran Zivkovic and Ben Kröse
In Proceedings of the Inside Data Association Workshop during the Robotics: Science and Systems Conference (RSS), June 2008.
Abstract
Loop closing in vision based SLAM applications is a difficult task. Comparing new image data with all previous image data acquired for the map is practically impossible because of the high computational costs. This problem is part of the bigger problem to acquire local geometric constraints from sensor data for geometric map building termed data association. Commonly the computational costs are kept small by sampling the image data uniformly over time or using a position estimate from a mapping and localization algorithm. In this paper we propose a more natural sampling approach, by determining a subset that best describes the complete image data in the space of all previously seen images. The actual problem of finding such a subset is called the Connected Dominating Set problem which is well studied in field of graph theory. The proposed method is particularly beneficial for realistic mapping scenarios including moving objects and persons, motion blur and changing light conditions. Evaluation on multiple large indoor datasets show that the method performance is very close to that of an exhaustive data association scheme and outperforms other sampling approaches.
Downloads
pdf (2.2 mb)
