creating a bag of features for new image set for monocular SLAM

Hi,

I am interested in visual SLAM. I have taken the example project from "Monocular Visual Simultaneous Localization and Mapping" and have been able to run it with the specified dataset. If I want to run with a different dataset (KITTI), what is the process for setting up the data? I believe I need to create a bag of features off-line? Is that correct? How exactly is that done. It is not explained well in the example.

In the example, a mat is loaded here: I was able to use the bagOfFeaturesDataSLAM.mat file from the example.

% Load the bag of features data created offline

bofData = load("bagOfFeaturesDataSLAM.mat");

For a new data set, you would need to create a new file, correct? That is not obvious to me what that procedure is.

thank you

Answer 1

[Qu Cao]

The bag-of-features data may not work for the KITTI dataset because it was trained using a small amount of image data. You may want to built your own bag-of-features bag following this section of the example.

[Louis]

Qu,

Thank you for your response. I did build a new bag -of-features for the KITTI set. I entered the correct focal length and principle points for KITTI as follows (I found it in the calib.txt file with the data).

The problem I am having now is in the helperTrackLastKeyFrame.m function.

At some point in the dataset (when the car makes the first right hand turn), the call to estWorldPose returns a status of 2 (not enough inliers found, minimum of 4). If that happens, the code just sets the mats to empty, which causes failure later.

if status
       currPose=[];
       mapPointIdx=[];
       featureIdx=[];
       return;
end

What do you think the proper error handling should be if status = 2?

It doesn't seem what emptying the arrays is the proper answer, since shortly after matlab throws an error.

If you have any suggestions, that would be helpful. It does work fine with the freiburg office scenery.

I am not a SLAM expert, but I would like to use Matlab, since the sparse point cloud visualization is good.

thx.

[Qu Cao]

If you want to run the pipeline on a different dataset, you may want to tune some of the parameters. For example, you may want to increase the numbers of feature points extracted from each image because the image resolution in the KITTI dataset is much larger than the one used in the example (480x640). Also, the frame rate of KITTI dataset is 10Hz, so you want to decrase the maximum skipped frame numSkipFrames to a smaller value, say 5.

For your convenience, I've attached a updated main example file for you. Note that you need to define your path to the image data and the camera intrinsic parameters.

[Louis]

Qu - this looks like it is working now. Thank you for your help!