In the frontend visual odometry of SLAM, the traditional feature matching method has poor extraction effect and instability in the case of changes in viewpoint and illumination, while the feature matching method based on deep learning cannot meet the real-time requirements on embedded devices with low computing power. To solve the above problems, this paper improves the SuperPoint network using depthwise separable convolution and designs a lightweight feature extraction network named L_SuperPoint. Based on the L_SuperPoint network, this paper designs a visual SLAM system, and the system not only has a better mapping effect than ORB-SLAM2 but also can run in real-time on embedded devices. The results of dataset simulation experiments and real scene experiments show that the L_SuperPoint network has both the robust feature extraction of the SuperPoint network and the real-time mapping capability of ORB-SLAM2, which effectively improves the operating efficiency and accuracy of the SLAM.
Visual SLAM Based on Lightweight SuperPoint Network
2022-10-12
2603398 byte
Conference paper
Electronic Resource
English
British Library Online Contents | 2012
|ExplORB-SLAM: Active Visual SLAM Exploiting the Pose-graph Topology
Springer Verlag | 2022
|