Visual localization, mapping and reconstruction using edges

Visual navigation is one of the fundamental problems in robotics. The last decade specifically has seen many important contributions in this field. As of today, feature point based approaches are by far the most popular. While successful in a host of applications, untextured environments can be highly problematic for these methods, since the number of reliable feature points is often low in these scenarios. Nevertheless, edges may still be abundantly available, however, remain unused. In this dissertation, we propose complementary edge-based methods for visual localization, mapping and dense reconstruction that can still operate in theoretically minimal scene configurations. Starting from sparse stereo edge matching, we propose two techniques with different performance/efficiency trade-offs that are both targeted at real-time operation. Besides a comparison to popular dense stereo techniques, we also compare the algorithms to our efficient adaptation of a line segment based stereo approach. Moving on to stereo visual odometry, we propose a line segment based reprojection optimization that is able to prevail in untextured environments where a proven state-of-the-art feature point based method fails. We argue that our approach can even cope with the theoretically minimal case, consisting merely of two nonparallel line segments. We then extend this approach to a full line segment based simultaneous localization and mapping solution. Using bundle adjustment we are able to build consistent line segment maps that have a high geometric expressiveness with respect to the underlying scene geometry. Especially our long line segment tracks are notable. These are made possible by being completely independent of photometric influences, and additionally our line segment end point estimation approach. We show that we are even able to close loops with viewpoint changes of 180◦. Finally, based on our line segment maps, we propose an efficient method for dense surface reconstruction. Without using restricting assumptions about the scene geometry, we show real-time suitable processing times that make our reconstruction approach highly applicable to robotic exploration use cases in structured environments.