How to eliminate erroneous match(es) in matchFeatures example?

Hi all,

I have a set of images that are taken from the dash cam of a car. I am trying to find a set of corresponding points in two successive frames to estimate the direction of the car. (I will use this information to obtain the visual odometry of the car). I used the match features code given here: https://www.mathworks.com/help/vision/ref/matchfeatures.html

But sometimes I am getting wrong matches. For example:

My function:

function [matchedPoints1, matchedPoints2] = Corresponding_Points(Im1,Im2)
    %%Find Corresponding Interest Points Between Pair of Images

      % Find the corners.
      points1 = detectHarrisFeatures(Im1);
      points2 = detectHarrisFeatures(Im2);

      % Extract the neighborhood features.
      [features1,valid_points1] = extractFeatures(Im1,points1);
      [features2,valid_points2] = extractFeatures(Im2,points2);  

      % Match the features.
      indexPairs = matchFeatures(features1,features2);   

      % Retrieve the locations of the corresponding points for each image.
      matchedPoints1 = valid_points1(indexPairs(:,1),:);
      matchedPoints2 = valid_points2(indexPairs(:,2),:); 

end

My code:

    [matchedPoints1, matchedPoints2] = Corresponding_Points(image_1_gray,image_2_gray);

    x1 = matchedPoints1.Location';
    x2 = matchedPoints2.Location';

Note: the returned matchedPoints1 and matchedPoints2 belong to class cornerPoints . Using matchedPoints1.Location returns a matrix type single. (I am using Matlab 2016a.)

Thanks!

Answer 1

[Image Analyst]

Do you know the (x,y) coordinates of each? You could just throw out any pairs where the distance between them is more than, say 10% of the image width.

[Canberk Suat Gurel]

Thank you for your answer. Yes, I know their (x,y) coordinates. I can do something like this:

for i = 1:length(x1)
    d = sqrt((x1(1,i)-x2(1,i))^2+(x1(2,i)-x2(2,i))^2)
    if(d>100)
        x1(:,i) = [];
        x2(:,i) = [];
    end
end

Can you think of a more efficient method? Getting rid of the for loop would be great.

[Image Analyst]

Not sure if shrinking the array at every iteration is efficient. Maybe try just keeping track of columns to keep and vectorize it and extract:

distances = sqrt((x1(1,:)-x2(1,:)).^2+(x1(2,:)-x2(2,:)).^2)
columnsToKeep = distances < 100;
x1 = x1(:, columnsToKeep);
x2 = x2(:, columnsToKeep);