Deep VSLAM

The paper analyzes the possibility of using deep learning methods in the problem of visual localization and mapping by video stream (vSLAM). One of the main stages of the vSLAM methods is selection and comparison of descriptors (i.e., features describing the scene) on the frames of the video stream. Usage of classical methods for solving this problem, based on the usage of ORB, SIFT, etc. descriptors, does not allow to take into account the global context of the scene, which can negatively affect the results of vSLAM. Trainable neural network methods of identification and comparison of features proposed to eliminate this drawback. In particular, the SuperPoint method is proposed for the selection of features and the SuperGlue method for comparison. The latter was also refined and improved using better initial approximation. The scheme proposed in this paper allows to improve the stability of the selection of keypoints by 2 times, as well as speed up the entire localization and mapping method by 13.6%, which can be useful in tasks with high stability requirements, such as analyzing terrain with robots.