Fast structure recognition in point clouds
Point clouds are an up-and-coming approach for coding 3D scenes. They can be created from laser scanners or from finding matching points in several frames recorded by a camera.
While some people consider meshes created from such point clouds still point clouds, this is not correct in the strict sense. For all point clouds taken directly from a recording of the real world, there are too many inaccuracies to seriously do this.
Instead, we want to pursue another idea in this thesis, which works roughly as follows:
- Randomly define a flat plane that intersects the point cloud
- Project all point cloud points into this plane, with the closest point defining the result. The plane encodes heights, not colours.
- Smooth the plane and compute the gradients in the plane.
- Try to identify regions in the plane that may represent geometric objects such as planes, cylinders or balls
- Repeat until you have an understanding of the space
It is recommended to use CUDA for programming done in this thesis.