Multi-Resolution particle filter estimation, applied to object recognition in an office environment

This paper presents a multi-resolution particle filter, applied to the recognition and localization of objects in an office environment, using a laser scanner. A fast low level particle filter combines multiple measurements of the Sick laser scanner to localize the legs of tables and chairs in an office environment. This low level filter uses a simple leg model to distinguish the legs from other objects in the office environment. A high level particle filter then combines the estimated leg positions to recognize tables and chairs, and find their position and orientation. The high level filter is helped by the knowledge of the geometry of the tables and chairs. This geometric knowledge is modeled using a simple Bayesian network. The estimation of the high level filter is used to correct the low level estimation, to add missing legs that where not found by the low level filter, and to remove legs that are not part of a table or a chair. The combination of the fast low level tracking with the high level geometric knowledge, increases the performance of the estimation problem, and allows for realtime object recognition and tracking in a 50 [m] office environment. Experimental results show the effective recognition of two tables and two chairs in an unstructured office environment.