MONOCULAR 3D SLAM USING A VISUAL LANDMARK DATABASE FOR AUTONOMOUS NAVIGATION NEAR SMALL CELESTIAL BODIES

Close proximity operations near small celestial bodies adds a new range of problems requiring more autonomy, more precision and more flexibility from the navigation software. Addressing these problems, this paper focuses on a 3D monocular SLAM based on the Rao-Blackwellized Particle Filter approach using a hybrid mapping module combining binary search trees acting as a visual landmark signature catalog, and an octree occupancy grid meant to offer a spatial representation of the landmark uncertainty distribution. SLAM motion estimation is performed at the beginning of each time step according to a probabilistic scheme based on fast linear visual pose estimation algorithms, sampling probable motion estimates from subgroups of visual features matched across pairs of subsequent navigation camera images. The autonomous SLAM-based navigation scheme estimates the spacecraft position and attitude or pose at each time step by picking the most probable pose from a sample population of particles representing different hypotheses on the spacecraft path and the landmark distribution conditioned on it. A low-pass filter processes the raw output of the SLAM to provide a smoother response that can be used by a conventional trajectory tracking controller.