Abstract:

We present a Visual Inertial Odometry (VIO) framework which uses 3D points, lines and planes from RGB-D data to address the challenge of rigid registration between successive camera poses in feature-poor environments. Previous VIO approaches have incorporated 2D features, but robust correspondences cannot be determined in low-texture and low-light situations. We show that by directly exploiting 3D geometric primitives we can achieve improved registration. We demonstrate our approach on different environments and compare the addition of different 3D geometric primitives to a ground truth trajectory obtained by a motion capture system. We consider computationally efficient methods for detecting 3D points, lines and planes, since our goal is to implement our approach on mobile robots, such as drones.