Detecting Loop Closure using Enhanced Image for Underwater VINS-Mono

In this paper, we adapt VINS-Mono to an underwater scene. It is found that the original method in VINS-Mono cannot detect enough loop closures in some underwater environments, the reasons are twofold. Firstly, the visual dictionary using FAST corners in underwater VINS-Mono may result in lots of false loop candidates. Secondly, though the right loop candidate may be found, too few feature matches or even mismatches may also exist in the feature matching based on FAST and Shi- Tomasi corners between the current frame and the loop candidate when using VINS-Mono in an underwater scene. To handle this, we propose a modified loop detection method using enhanced images. Firstly, through lots of field experiments, we find the visual dictionary with Oriented FAST and Rotated BRIEF (ORB) features performs better in loop candidate detection than using the original FAST features. Secondly, to deal with the matching problem, we use Dark Channel Prior (DCP) to enhance the images, and the geometric constraints are considered and then followed by the RANSAC iteration to enhance the robustness to outliers. Therefore, the transform matrix between the two frames can be obtained by computing the fundamental matrix and the scale ambiguity problem for translation can be solved by using the metric scale estimated in initialization. Compared with the original VINS-Mono, the proposed method can detect more loops and hence reduce the total error of the vehicle's trajectory. Experiments in an outdoor pool illustrate the feasibility of the proposed method.