VINS-Mask: A ROI-mask Feature Tracker for Monocular Visual-inertial System

Abstract

Feature tracker is usually believed to be one of the most important components to the performance influence on a Visual-inertial System (VINS). This paper proposes the VINS-Mask scheme, a more robust feature tracker for monocular VINS through Region of Interest (ROI) masks. It could achieve real-time feature tracking with high accuracy and robustness. Firstly, we propose an edge mask to generate the edge-sensitive feature candidate regions from the incoming image frame. Next, we design an interest point sensitive SuperPoint mask with deep learning framework to obtain repeatable and reliable feature candidate regions. We also dynamically adjust the inflation radius by monitoring the initial status from VINS Initialization module to obtain more accurate ROI masks. Notably, compared with the best baseline approach (i.e., VINS-Mono), our VINS-Mask scheme achieves an average improvement accuracy of 0.068m on the dataset of EuRoc drone. After paper publication, our source codes will be available at https://github.com/sunjia-yuanro/VINS-Mask.git.