As maps get bigger and changes to the environment will get more and more frequent we need to get over the current way of human detection which is done by using robot_tracking/dynamic_obstacles based on laser-beam classification by the localization.

Instead, we need a localization-independent general shape classifier that works on laser-data directly. This can then be used to

• perform human leg detection (there are various papers on that) in a pure local way (i.e. without dependencies to localization and map)
• perform furniture recognition to maybe update a given map on-the-fly given the new position of a known and recognized object.