Eligible Features Segregation for Real-time Visual Odometry

Abstract

Stereo vision aided inertial navigation enables accurate self-localization and navigation in unknown environment without the need of positioning systems (e.g. GNSS). The feature extractor is an elementary part of this technology because it extracts the landmarks from the camera images, which are then used for optical navigation. This is why the feature extractor is usually the first module of data processing chain. In most of the cases, the features from the feature extractor are filtered by a non-maximum suppression algorithm. An ideal non-maximum suppression algorithm suppresses "weak" features while keeping "strong" and well-distributed features. Only if the feature extractor is combined with an appropriate non-maximum suppression module, the computer vision system can get reasonably good results. In this paper, we propose a novel non-maximum suppression algorithm. The algorithm does not only provide well-distributed features over the whole image but is also be able to control the maximum number of required features in output, which is very important for real-time system. We apply our framework to the AGAST feature extraction algorithm, and it is very easy to incorporate with other feature extractors. Finally, we combine our algorithm with the Integrated Positioning System (IPS) which is developed by the German Aerospace Center (DLR). The comparison of testing results is illustrated.

Available Versions of this Item

• Eligible Features Segregation for Real-time Visual Odometry. (deposited 12 Dec 2017 10:14) [Currently Displayed]
  • Eligible Features Segregation for Real-time Visual Odometry. (deposited 28 Jan 2020 08:42)