Online Image Pre-processing Using Bayesian Optimization for Visual-inertial Odometry

In this study, we propose an online image pre-processing method to improve the accuracy of visual-inertial odometry (VIO) in a low-light environment. Vehicles operating in low-light environments mostly attach light source mounted on their body, so their motion causes rapid local brightness changes in the image of a vision sensor. Since brightness change impedes extraction and tracking of feature points for VIO, it can degrade the localization performance. To solve this problem, γ-correction and contrast limited adaptive histogram equalization techniques for an online image pre-processing are utilized to improve the image quality. The techniques find optimal values of their key parameters based on Bayesian optimization. The objective function of the optimization is to maximize the number of extracted feature points so that more measurements are obtained despite changes in brightness. To validate our method, we construct a multi-state constraint Kalman filter-based VIO algorithm with the proposed method and present the results of an enhancement of the feature extraction and a reduction of localization error in the open dataset.