Intelligent Point cloud for 3D data registration

If we can not recognize 3D objects from 3D point clouds, those point clouds are just a bunch of xyz coordinates. Today I found that for complicated environments like inside the vision group laboratory, even though there is only minor changes of the environment (walking of a student, movement of a small board), two scans can not be registered together accurately.

That's too bad, because we can never expect construction site will be clean and static. If we can not automatically recognize object from point clouds and separate noisy information from useful information, we can not register the 3D point clouds using traditional algorithms. Another thing is that if we want to detect the environment change, we can not directly register two model and get the differences between them, we must segment the point clouds into patches and do "patch-based matching" between two 3D models and find a transformation which can figure out "outlier patches", which is the actual differences between two scenarios.

Anyway, if we know in advance which point belongs to which object, computer and operate on a "object" level and utilize domain knowledge to reason about the scene and detect changes or find a data registration results making more sense(this is a wall, it will not change, that is a table, it could be moved), so we will avoid the problem caused by non-semantic approach: all points are viewed as the same thing(a point, nothing else!), and the algorithm simply trying to minimize the distance between points so that small changes in the scenario will influence the whole registration results and result in failure of 3D registration.