helperTrackLocalMap error with Monocular SLAM

Hi everyone,

I have run the Monocular SLAM (https://www.mathworks.com/help/vision/examples/monocular-visual-simultaneous-localization-and-mapping.html.) with Matlab 2020a, and I have tried some datasets, e.g., the TUM dataset in the tutorial and the Kitti dataset. How, I have got the following error.

Error using worldToImage
Expected worldPoints to be nonempty.

Error in worldToImage (line 83)
validateattributes(worldPoints, {'double', 'single'}, ...

Error in helperTrackLocalMap (line 63)
projectedPoints = worldToImage(intrinsics, R, t, localMapPoints);

Error in main (line 218)
        helperTrackLocalMap(mapPointSet, vSetKeyFrames, mapPointsIdx, ...

I have found that this error is due to few keypoints. I did not change anything of the tutorial except changing the name of the helper function helperSURFFeatureExtractorFunction to helperExtractorFunction (helperExtractorFunction ). I was wondering if anyone else can run the tutorial successfully. Any help would be appreciated.

[Qu Cao]

The error indicates that not enough features can be tracked in the current frame.

Could you attach an image of the point cloud? It will help to understand if the tracking failure occurs even before this frame.

[Sashamaru John]

I think the tracking is success for previous frames, but there are no inliers for current frame (1051).

[Qu Cao]

I agree with your that the tracking for previous frames are succsessful.

Please try tunning the parameters to see if it helps improve the robustness:

• In helperIsKeyFrame, set the minimum number of tracked map points to a larger value, say 90;
• In helperCreateNewMapPoints, change 'MaxRatio' to a smaller value, say 0.7.

[Sashamaru John]

Thanks, it works.

[Qu Cao]

Sashamaru, great to hear that tuning the parameter works!

Could you let me know which parameter of the two you changed?

[Sashamaru John]

Both work for me. Changing the first seems to be better than the second.

Answer 1

[Qu Cao]

Answer pasted from comments:

Please try tunning the parameters to see if it helps improve the robustness:

• In helperIsKeyFrame, set the minimum number of tracked map points to a larger value, say 90;
• In helperCreateNewMapPoints, change 'MaxRatio' to a smaller value, say 0.7

[jose daniel hoyos giraldo]

hello. Can you give me a hand improving the robustness, because I'm using a low resolution camera. Do you suggest change the line (in helperIsKeyFrame):

tooFewMapPoints     = numel(mapPointsIndices) < 100;

for

tooFewMapPoints     = numel(mapPointsIndices) < 500;

and the line:

tooFewTrackedPoints = numel(mapPointsIndices) < 0.9 * numPointsRefKeyFrame;

for

tooFewTrackedPoints = numel(mapPointsIndices) < 0.95 * numPointsRefKeyFrame;

Or similar?

On the other hand, in helperCreateNewMapPoints, u suggest use a smaller value of 'MaxRatio', but at mathworks it says "Use the max ratio for rejecting ambiguous matches. Increase this value to return more matches." (https://la.mathworks.com/help/vision/ref/matchfeatures.html).Then, if we have poor image quality and issues with a low number of points, why do you suggest decrease this value?.

Thanks for your help!

[cui,xingxing]

I think this is more of an empirical threshold that is dynamically adjusted to your own images @jose daniel hoyos giraldo ，You can refer to the thresholds designed in the latest version of R2022b for the exact TUM dataset