Visual SLAM (part IV)

Visualization of correcting a robot's odometric position with visual information. The robot (blue triangular shape) observes objects with a camera (upper right). These observed objects will be used as landmarks and are put in a map (green crosses).

When the robot observes a landmark for the second time, it corrects its own position relative to the observed landmark postion; 'loop closure'.

The blue arrow is the 'odometric' robot position (where the robot 'thinks' it is). The red arrow is the 'corrected' robot position.

The ellipses represent the uncertainties of the positions of the landmarks. The smaller the ellipse, the more certain the robot is about the position of the landmark.

The robot (or actually the corrected position) has its own uncertainty ellipse. This ellipse grows during driving. When the robot observes a previously seen landmark, both the uncertainties of the landmark and the robot are adapted.

The small orange dots represent the trail of corrected positions. The small gray dots represent the (uncorrected) odometric trail.

Category:

Howto & Style

Tags:

• visual
• slam
• buffer
• simultaneous
• localisation
• mapping
• robot
• SIFT
• feature
• monocular

[aisjoerd]

I used a visual buffer to select stable visual interest points. I didn't use ANNs, in the final version of my implementation a decision (to use just a part of the visual field to select interest points) was made based on information from the created map. A full description can be found in my master thesis, accessible via my website

[miguelzarth]

So, what u use for adaptation and search for objects? ANNs?