Raymond's Homepage
As a member of the 2003 RoboCup Four-Legged League soccer team, I was responsible for, amongst other things, feeding geometry-based localisation into the existing infrastructure, an Extended Kalman filter. Unfortunately, the information I was getting, such as "You are 30cm away from a line", really doesn't linearise well when that line could be anywhere on the field!  I decided to see if something else, such as a Monte-Carlo Localisation Filter, would do any better. So, when I went back to Curtin University to finish my degree (at the time, I was at the University of New South Wales as a cross-institutional student), I decided to implement the same sensor model but as an input to an MCL filter. The results were, I think, quite promising! The environment was an indoor, domestic office-style room with a regular pattern on the floor. This is really hard to linearise as any one observation of the pattern (crosses on the floor) could mean you were in one of 48 distinct positions (if you wanted to do this with a multimodal Kalman filter you'd need at least 48 distributions). We also had overhead cameras but they weren't terribly high resolution and there were many opportunities for occlusion, combined with other moving objects. For more details, please see our paper! This was published at the 2004 RoboCup Symposium. |
