New result
motters
system. This time the images were taken within a small bedroom
approximately 3.5 metres square. The stereo images together with the
robot design file (which includes camera calibration information) and
a simulation file which describes the path of the robot can be found
here:
http://sentience.googlegroups.com/web/sequence2.zip (3MB)
This new data has been useful, because it highlighted a number of bugs
in the system which I subsequently fixed. You can observe the mapping
process in the following animation.
http://sentience.googlegroups.com/web/bedroom1map.gif
Here you can really see the DP-SLAM algorithm doing its job as it
decides between multiple possible interpretations of the incoming
data. The evolution of the ancestry tree can also be seen in this
animation:
http://sentience.googlegroups.com/web/bedroom1.gif
Once all the hypotheses collapse to a single decision that part of the
path turns green and is permanently added (distilled) to the occupancy
grid. Although the simulation file contains the exact pose
information from which the stereo images were taken, the simulated
robot does not have access to this information. Instead it only knows
its forward and angular velocity, with some amount of random noise
added. If it were relying upon velocity information alone huge dead
reconing errors would accumulate, turning the map into an
unrecognisable jumble.
Although the occupancy grid produced here looks somewhat crude this is
quite a pleasing result, because it really demonstrates that the DP-
SLAM algorithm is able to cope well even when faced with substantial
uncertainty both in the sensor readings and the velocity data
available to the robot.