Remarkable performance has been achieved using the state-of-the-art monocular Simultaneous Localization and Mapping (SLAM) algorithms. However, due to the scale ambiguity limitation of monocular vision, the existing monocular SLAM systems can not directly restore the absolute scale in unknown environments. Given the amazing results in the field of depth estimation from Convolutional Neural Networks (CNNs), we propose a CNN-based monocular SLAM, where we naturally combine the CNN-predicted depth maps together with the monocular ORB-SLAM, overcoming the scale ambiguity limitation of the monocular SLAM. We test our method using the popular KITTI odometry benchmark, and the experimental results show that the overall performance of average translational and rotational error can reach 2.00% and 0.0051°/m. In addition, our approach can work well under the pure rotation motion, which shows the robustness and high accuracy of the proposed algorithm.
Paper
Full text
Scale-aware Monocular SLAM Based on Convolutional Neural Network
Semantic Scholar · Computer Science · 2018
Abstract
Remarkable performance has been achieved using the state-of-the-art monocular Simultaneous Localization and Mapping (SLAM) algorithms. However, due to the scale ambiguity limitation of monocular vision, the existing monocular SLAM systems can not directly restore the absolute scale in unknown environments. Given the amazing results in the field of depth estimation from Convolutional Neural Networks (CNNs), we propose a CNN-based monocular SLAM, where we naturally combine the CNN-predicted depth maps together with the monocular ORB-SLAM, overcoming the scale ambiguity limitation of the monocular SLAM. We test our method using the popular KITTI odometry benchmark, and the experimental results show that the overall performance of average translational and rotational error can reach 2.00% and 0.0051°/m. In addition, our approach can work well under the pure rotation motion, which shows the robustness and high accuracy of the proposed algorithm.