Matching keypoints between images showing the same scene under different conditions is a fundamental step for a variety of applications. Recent approaches based on convolutional neural networks show superior results in terms of discriminability compared to well established descriptors like SIFT or ORB. However, there is less previous work which brings the CNNs to automated driving applications like SLAM and analyze the performance in terms of accuracy and runtime. In this work, we take state-of-the-art patch comparison CNNs, train them from scratch and analyze the performance on the KITTI odometry benchmark. For that, we replace the ORBfrontend within the publicly available ORB-SLAM2 framework through our trained CNN variants and compare both. We show that it is necessary to downsize the complexity of the original architectures to achieve real-time capability. Furthermore, our evaluation shows that the downsized models achieve significantly higher matching performance than the ORB descriptor. Moreover, we achieve slightly better results on the KITTI odometry benchmark compared to ORB-SLAM2 while using a CNN-based feature descriptor, which can easily be adapted to different environments.
An Approach for CNN-Based Feature Matching Towards Real-Time SLAM
2019-10-01
806216 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Parsimonious Real Time Monocular SLAM
British Library Conference Proceedings | 2012
|Parsimonious real time monocular SLAM
IEEE | 2012
|Real-Time Pose Graph SLAM based on Radar
IEEE | 2019
|Scan matching SLAM in underwater environments
British Library Online Contents | 2014
|