MATLAB: Multi-target tracking “assignDetectionsToTracks” function has unexpected results

Computer Vision ToolboxMATLABmulti-object trackSensor Fusion and Tracking Toolboxtrack;

When I use the "assignDetectionsToTracks" function, the 5th (row) tracking is associated with the (1)th col detection, and the threshold(costOfNonAssignment=1) I set is 1, and there should be no correlation. Why?

cost = [2.62407,1.93346,4.28524,0.0126375;
    1.34658,3.33612,0.769382,3.78107;
    2.04141,0.012594,3.98619,1.92012;
    3.35982,2.04044,5.14111,1.01901;
    1.59666,3.05409,2.13898,2.69844;
    4.54989,6.35886,3.17712,6.65899];
costOfNonAssignment = 1;
[assignment,unassignedTracks,unassignedDetections] = ...
            assignDetectionsToTracks(cost,costOfNonAssignment)

assignment =

4×2 uint32 matrix

5 1

3 2

2 3

1 4

unassignedTracks =

2×1 uint32 column vector

4

6

unassignedDetections =

0×1 empty uint32 column vector

As can be seen from the result assignment, the fifth track is associated with the first detection data, and the cost between them is 1.59666, and the threshold I set is 1, the cost is greater than the threshold, why the fifth track Related to the first test? Shouldn’t it belong to unassignedTracks?

Best Answer

The fifth row / first column is associated because if these were not associated you would incure a cost of unassignment for both row and column, meaning 1+1 = 2. 2 is greater than 1.59666 so the assignment is favored instead of unasignment.

Related Solutions

MATLAB: How to remove shadow in our program ,

If your video is RGB, you can try converting the video frame to YCbCr using rgb2ycbr function, and pass only the Cb and the Cr components to the foreground detector. This should reduce the false detections caused by shadows, but you may also start missing real objects that only differ from the background in intensity.

MATLAB: How would i track the speed of the centroid of the moving objects

If it is acceptable to compute velocities post hoc, I suggest the following edits to this code:

1. Add tracks as an output so you can process them from another script afterwards:

function tracks = multiObjectTracking()
...

2. In the initializeTracks() function, add a field in the tracks structure to store position over time (I suggest using an animatedline object). Also add a field to keep track of which tracks are active or not:

tracks = struct(...
    ...
    'active',{}, ...
    'position',{});

3. In the createNewTracks() function, instantiate an animated line when you create a new track and set the track to active:

newTrack = struct(...
    ...
    'active',true, ...
    'position',animatedline()); %you can add optional parameters to the animatedline here if you want

4. In the deleteLostTracks() function, instead of deleting tracks, simply set them to inactive:

Replace

lostInds = ...;
tracks = tracks(~lostInds);

with

lostInds = find(...);
for ii=1:length(lostInds)
    tracks(lostInds(ii)).active = false;
end

5. In the displayTrackingResults() function, add a condition to the reliableTrackInds to check active state:

reliableTrackInds = ... & [tracks(:).active];

Then, add the centroid location to the animatedline object for each active track (I suggest after "mask = insertObjectAnnotation..."):

x = double(bboxes(:,1)+bboxes(:,3)/2);
y = double(bboxes(:,2)+bboxes(:,4)/2);
for ii=1:length(reliableTracks)
    addpoints(tracks(reliableTracks(ii).id).position,x(ii),y(ii));
end

6. The last step is to make sure the function predictNewLocationsOfTracks(), detectionToTrackAssignment(), updateAssignedTracks(), and updateUnassignedTracks() ignore the inactive tracks. There are many potential ways to do this; I'll leave this part to you.

7. After the code has been edited, you should be able to run:

tracks = multiObjectTracking();
[x,y] = getpoints(tracks(4).position);

to get the centroid path for the 4th track, for example. The velocity can be computed easily enough with this information.

Hope this helps.

Related Question