Getting Robots to Recognize Places

Abstract

If a robot relies only on its odometer and compass to navigate, its position estimate can become erroneous to a point where the estimate is not useful.

The robot can correct the error by visually identifying its current location. This is place recognition.

Our robot uses visual panoramas of places for recognition. The panoramas look like this:

Here the robot identified the door as a place it has seen before:

How well does this work?

The current implementation recognizes places while allowing for

Image translations of upto 1/3 of the image,

Scale changes and non-linear wapring of upto +- 20%

Intensity changes of upto +- 50%

Presense of outliers such as people standing in the image

We are collecting the statistics of the false positive and false negative rates by testing the algorithm on real mobile robot.

References

Hemant's Home Page