rtabmap outdoor

When running rtabmap in localization mode (using a map that was made e.g. a few hours before) we're having problems finding loop closures.

I can see that many of the extracted keypoints are features of shadows, and because the sun moves, or it gets clouded temporarily, the loop closures don't happen anymore.

The problem is a lot less when it's clouded.

Is there any way to tune the key point extractor to ignore shadows? Or any other parameters that are useful in these situations?

Asked by rembert on 2018-10-18 03:11:08 UTC

Comments

Answers

Shadows can be indeed very challenging for visual loop closure detection. There is no way in rtabmap to know if some parts of the image are shadows or not.

With its default parameters, rtabmap extracts features only if corresponding depth pixel is valid. The general approach is to filter the depth image where we don't want features extracted. If you could know from another approach (e.g., deep learning) which parts of the image are shadows, the corresponding mask could be applied to depth image so that rtabmap doesn't extract features from shadows.

cheers,
Mathieu

Asked by matlabbe on 2018-10-19 10:24:27 UTC