Real-time photo-realistic 3D mapping for micro aerial vehicles

We present a real-time photo-realistic 3D mapping framework for micro aerial vehicles (MAVs). RGBD images are generated from either stereo or structured light cameras, and fed into the processing pipeline. A visual odometry algorithm runs on-board the MAV. We improve the computational performance of the visual odometry by using the IMU readings to establish a 1-point RANSAC instead of using the standard 3-point RANSAC to estimate the relative motion between consecutive frames. We use local bundle adjustment to refine the pose estimates. At the same time, the MAV builds a 3D occupancy grid from range data, and transmits this grid together with images and pose estimates over a wireless network to a ground station. We propose a view-dependent projective texture mapping method that is used by the ground station to incrementally build a 3D textured occupancy grid over time. This map is both geometrically accurate and photo-realistic; the map provides real-time visual updates on the ground to a remote operator, and is used for path planning as well.