Effectiveness for monocular SLAM about anms-codes HOT 2 CLOSED

Dear Pablo,

your observation is certainly valid. When detecting keypoints, there is a certain trade-off between their spatial distribution and their repeatability. If you only want to use a restricted number of points they might be difficult to match.
We recommend to use a keypoint tracker such as KLT instead of a matching approach (like KNN or BF). we actually use this strategy in our own SLAM system and it provides very stable tracks and compelling results.
For sparse stereo matching, you can detect keypoints on the left image and find correspondences by line scanning.

Best regards,

from anms-codes.

Thank you for you answer,

Using a KLT tracker could be useful for fixed focus cameras, in fact, we use it as a complement in our monocular keyframe-based SLAM. However, any time we get an occlusion we have to perform relocalization using feature matching, and there is when we need strong keypoints.

Anyway, I think we are exploring very different SLAM approaches. In our case we use it for Augmented Reality in mobile phones where only one camera exists and we have to deal with very limited computing resources. I guess you're probably using it with stereo cameras for autonomous driving so what is good for your SLAM system could be bad for ours and vice versa.

Warm regards

from anms-codes.