Computer Vision Onramp - Part 5 - Object Tracking and Kalman Filters

Object Tracking

Why Kalman Filters?

Tracking by detection uses a detector and a Kalman filter to follow objects. The filter predicts where an object will be, which is crucial for handling occlusions (when an object is temporarily hidden).

Initialize a Kalman Filter

To set up a Kalman filter, you define assumptions about motion:

1. Motion Model: “ConstantVelocity” for the turtle.
2. Initial Location: Centroid of the first detection: [xmin + width/2, ymin + height/2].
3. Measurement Noise: Variance in centroid measurement (e.g., 100).
4. Motion Noise: Deviation from the model (e.g., [1 25] for location/velocity).
5. Initial Estimate Error: Allows the filter to adjust quickly initially (e.g., [1 25]).

kf = configureKalmanFilter(motionModel, initialLoc, initialError, motionNoise, measurementNoise);

Track a Single Object: Predict, Detect, Correct

The workflow for each frame:

1. Predict: Where should the object be now?
2. Detect: Run the detector.
3. Correct: If detected, refine the filter with the actual location.

while hasFrame(turtleVideo)
    trackedLoc = predict(kf);
    frame = readFrame(turtleVideo);
    [bbox, score] = detect(detector, frame);

    if ~isempty(bbox)
        strongestBbox = selectStrongestBbox(bbox, score, NumStrongest=1);
        centroid = [strongestBbox(1)+strongestBbox(3)/2 strongestBbox(2)+strongestBbox(4)/2];
        trackedLoc = correct(kf, centroid);
    end
    
    % Red circle for tracked location, Green for detection
    frame = insertShape(frame, "filled-circle", [trackedLoc 20], Color="red");
    imshow(frame)
    drawnow
end

---

Achievement Certificate

Congratulations on completing the Computer Vision Onramp!

View Course Completion Certificate

Click to view certificate