Collision detection on point clouds using a 2.5+D image-based approach

Abstract

This work explores an alternative approach to the problem of collision detection using images instead of geometry to represent complex polygonal environments and buildings derived from laser scan data, used in an interactive navigation scenario. In a preprocessing step, models that are not point clouds, are sampled to create representative point clouds. Our algorithm then creates several 2.5+D maps in a given volume that stacked together form a 3D section of the world. We show that our new representation allows for realistic and fast collision queries with complex geometry such as stairs and that the algorithm is insensitive to the size of the input point cloud at run-time.