Multi-level Pixel Detection Algorithm Based on Monocular Visualinertial System

Abstract:
To enhance the localization accuracy of the conventional monocular visual-inertial localization system, a multi-level pixel feature detection algorithm combining pixel-level and sub-pixel-level is proposed. The feature detection firstly performs initial feature detection by pixel-level feature algorithm to obtain the number of feature points to be optimized, then iterates the initial values by sub-pixel detection algorithm to improve the extraction accuracy of the corner, and finally applies edge constraints to the results to prevent the detection of sub-pixel corner edge from crossing the boundary. According to the comparison experiments on the EuRoC dataset, the algorithm improves the accuracy of system localization compared to the VINS-mono algorithm.