Depth image based featureless SLAM for small UAVs

Abstract

This master’s thesis presents a keyframe-based featureless SLAM utilizing a single depth image stream. The real-time performance is obviously guaranteed on a small embedded computer for small UAVs, and the system is robust to UAV motions which are mainly composed by forward and rotational motions. The motion of the camera is estimated by the depth image alignment in contrast to previous point-cloud approaches. A Jacobian matrix appropriate for the depth image is analytically derived for the error optimization method. The speed of the error convergence is greatly improved by the projective matching scheme of the image alignment approach, the reduction of the image resolution, and the proposed intensity gradient-based pixel sampling method. The 3D consistent map is constructed from depth-maps of the keyframes through the pose-graph optimization. The system is integrated into a small UAV platform, and several experiments with a motion capture system are conducted to validate the performance of the system.