Constantin Berzan and Stuart J. Russell

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2015-191

August 13, 2015

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-191.pdf

We work towards solving the Simultaneous Localization and Mapping (SLAM) problem using a Probabilistic Programming System (PPS). After surveying existing SLAM methods, we choose FastSLAM as the most promising candidate. FastSLAM uses ad-hoc methods for data association, and does not enforce mutual exclusion between observations arriving at the same timestep. This leads to poor accuracy on an example dataset. We propose a new probabilistic model for SLAM that handles association uncertainty and mutual exclusion. We then propose an algorithm for doing inference in this model: FastSLAM-DA (FastSLAM with Data Association), which uses a particle filter with a custom data-driven proposal. We show that FastSLAM-DA performs well on the example where FastSLAM previously failed. However, the new algorithm produces inaccurate maps when there is a high rate of false detections. To remedy this, we propose FastSLAM-DA-RM (FastSLAM with Data Association and Resample-Move), which adds MCMC moves on the recent association variables. We show that FastSLAM-DA-RM performs well where FastSLAM-DA previously failed. Our two new algorithms use no heuristics other than custom proposals, so they are suitable for implementation in a PPS. As a step in this direction, we implement a general-purpose resample-move particle filter in the BLOG PPS, and demonstrate it on a simplified SLAM problem.

Advisors: Stuart J. Russell

---

BibTeX citation:

@mastersthesis{Berzan:EECS-2015-191,
    Author= {Berzan, Constantin and Russell, Stuart J.},
    Title= {Monte Carlo Methods for SLAM with Data Association Uncertainty},
    School= {EECS Department, University of California, Berkeley},
    Year= {2015},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-191.html},
    Number= {UCB/EECS-2015-191},
    Abstract= {We work towards solving the Simultaneous Localization and Mapping (SLAM) problem using a Probabilistic Programming System (PPS). After surveying existing SLAM methods, we choose FastSLAM as the most promising candidate. FastSLAM uses ad-hoc methods for data association, and does not enforce mutual exclusion between observations arriving at the same timestep. This leads to poor accuracy on an example dataset. We propose a new probabilistic model for SLAM that handles association uncertainty and mutual exclusion. We then propose an algorithm for doing inference in this model: FastSLAM-DA (FastSLAM with Data Association), which uses a particle filter with a custom data-driven proposal. We show that FastSLAM-DA performs well on the example where FastSLAM previously failed. However, the new algorithm produces inaccurate maps when there is a high rate of false detections. To remedy this, we propose FastSLAM-DA-RM (FastSLAM with Data Association and Resample-Move), which adds MCMC moves on the recent association variables. We show that FastSLAM-DA-RM performs well where FastSLAM-DA previously failed. Our two new algorithms use no heuristics other than custom proposals, so they are suitable for implementation in a PPS. As a step in this direction, we implement a general-purpose resample-move particle filter in the BLOG PPS, and demonstrate it on a simplified SLAM problem.},
}

EndNote citation:

%0 Thesis
%A Berzan, Constantin 
%A Russell, Stuart J. 
%T Monte Carlo Methods for SLAM with Data Association Uncertainty
%I EECS Department, University of California, Berkeley
%D 2015
%8 August 13
%@ UCB/EECS-2015-191
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2015/EECS-2015-191.html
%F Berzan:EECS-2015-191