Next-Best-View Planning

下一最佳视角NBVAdvanced

Deciding where a sensor should look next, given what it has already seen, to gain the most useful information.

Next-best-view (NBV) planning addresses this problem: given a partial set of observations so far, choose the next camera or scanner pose that will add the most new information for the least cost. It traces back to a 1985 ICRA paper by Connolly, who evaluated candidate viewpoints using a partial octree model. A common modern approach scores each candidate viewpoint by how many unknown voxels it would newly reveal in an occupancy map, combined with the cost of moving there. It is used for 3D-scanning objects, exploring a scene, and grasping occluded objects, and is one of the central problems in active perception.

ExampleA robot arm with a wrist-mounted camera scans an unknown object, moving at each step to the pose that reveals the most previously unobserved area, until the 3D model is complete.

Also called: NBV, NBV Planning, Best Next View
Related: Active Perception · Active Exploration · Occupancy Grid Map · OctoMap · Occlusion · Wrist Camera
Sources: The determination of next best views (Connolly, ICRA 1985) - Semantic Scholar