Maximizing Measurement Visual-Inertial Odometry for Low-Light Environment through Online Image Pre-processing

Recently, the interest in exploring unknown areas such as caves and tunnels has increased, and Visual-Inertial Odometry (VIO) can provide a suitable localization solution thanks to its lightweight and low cost advantages for mobile robots or drones. However, in such environments, lighting conditions are often not guaranteed due to lack of infrastructure. Although, attached light sources, such as body-mounted lights on robots and headlights on autonomous vehicles, can be used to address this problem, it is difficult to obtain images of consistent quality since the direction of the light changes as the robot moves. This can adversely affect the extraction and tracking of feature points, resulting in insufficient measurements. To solve this issue, we propose Maximizing Measurement VIO, named MMVIO, to ensure consistent image quality and rich measurements, which enables prolonged feature tracking in low-light environments. We employ two image pre-processing techniques, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\gamma $ </tex-math></inline-formula>-correction and CLAHE, whose parameters are estimated online using Bayesian Optimization (BO). The objective function of BO is to maximize the number of feature points extracted by the FAST corner detector, effectively increasing the measurements. For a detailed analysis, we conducted experiments using public datasets to validate our approach. The results show that MMVIO achieves the improved tracking time of the feature points and superior localization performance compared to other approaches that use different image quality metrics as objective functions. Moreover, we present the results that the localization accuracy of MMVIO is not only improved by online image pre-processing, but also is comparable to or outperforms the state-of-the-art VIOs.