Simultaneous Localization and Mapping (SLAM) is a well studied problem
in robotics since it is usually a prerequisite for autonomous
navigation in unknown environments. An interesting variant of the
general problem is the case of a robot equipped with bearing-only
sensors, most popular of which are monocular cameras.
There is a wide variety of algorithms that have been proposed to solve
the SLAM problem in general and Bearing-Only SLAM (BO-SLAM) in
particular. Our goal with this project is to implement the various
algorithms and compare their efficiency in accurately and robustly
solving the BO-SLAM problem. In our setup we assume that the sensor is
able to extract landmarks from its sensory input, so mapping the
environment corresponds to estimating the landmark coordinates.
The algorithms we are considering are the following:
- Extended Kalman Filter (EKF): Assuming the motion and odometry models
can be linearized and that the noise follows a white, Gaussian
model. We are testing the following EKF approaches:
- The most straightforward approach assumes that an Extended Kalman
Filter is used to simultaneously localize the robot and to locate the
- An improvement over the previous method can possibly be achieved
by using multiple Kalman filters to independently track each landmark.
Filters that do not agree with the measurements are pruned. When only
one Kalman filter remains then the landmark is added in the global
Kalman filter and the method continues as above.
- On-line Expectation Maximization: Iteratively solve the
Localization and the Mapping subproblems separately by assuming that
there is a solution to the sibling subproblem. Apply Particle Filters
(PFs) to solve each one of the subproblems. We are testing the
following variants of PFs:
- The Sampling Importance Sampling (SIS) approach with Resampling
according to the Effective Sampling Size (ESS) criterion.
- The Sampling Importance Resampling (SIR) aka. Monte Carlo
- The Dual-MCL and Mixture-MCL approaches.
- The SIR approach with Sensor Bias.
- The SIR approach with Sensor Resetting.
- The SIR approach with Roughing.
- Rao-Blackwellized Particle Filters: By applying an approach from
statistics called Rao-Blackwellization it is possible to show that we
can get an algorithm for the SLAM problem that converges (Online E/M
does not provably converge) and which has a better complexity than the
We will be soon reporting the results of our experimental comparison.
The above figures show an example of a simulated experiment. There are 20 landmarks
dispersed in the environment and a mobile robot. In the left figure we can see the
uncertainty in the landmark locations when the robot has just started moving. In the
right figure the robot has executed its exploration path and the landmark positions
have been estimated.