VINS-MultiCam: An ASLfeat-Based MultiCam Visual-Inertial Odometry Framework

The visual-inertial odometry (VIO) with monocular or stereo cameras has been remaining an active topic in robotic research over the last decade. However, due to the limited field of view (FoV), the feature tracking in some complex conditions like agile flights cannot provide enough visual information for odometry system. The multicam VIO algorithm has become a promising solution and gained widespread attention. This paper proposes a generic visual-inertial navigation system framework, i.e., VINS-MultiCam, for multiple arbitrarily arranged cameras. First, multiple cameras are divided according to their FoV overlaps. Second, the ASLfeat-based feature extraction and tracking are designed to be its frontend. Then, a novel feature management strategy is introduced to construct special local loopclosure constraints within different cameras at different times of the sliding window. Finally, the experiments with datasets in both simulation and the real world demonstrate the effectiveness and enhanced accuracy over the state-of-the-art works.