VINS-FEN: Monocular Visual-Inertial SLAM Based on Feature Extraction Network

Monocular visual-inertial simultaneous localization and mapping (SLAM) technology is able to be widely used to provide pose for unmanned aerial vehicles. It usually uses artificially designed feature points and descriptors as the feature and basis for image matching. However, it is easy to cause the problem of difficult feature extraction and feature matching error under uneven illumination and weak texture environment. In order to solve the above problems, this paper adopts the deep convolutional neural network (CNN) instead of traditional artificial design features to replace the traditional front end of visual-inertial system (VINS). My main work includes designing deep convolutional neural Network–Feature Extraction Network (FEN), for feature extraction, proposing a two-stage matching strategy, and porting the above improvements to the front end of VINS to form a complete system. Finally, verification is conducted on HPatches dataset and EuRoc dataset. The experimental results show that FEN is 3%~23% higher than the traditional method in repeatability and accuracy of extracting feature points. The VINS with FEN as the front end has stronger robustness and improves localization accuracy by 17.3% under uneven illumination and weak texture conditions.