lfcooh's profile - activity

I use this code to track object from findContour method.

public class Tracker extends JTracker {

public Tracker(float _dt, float _Accel_noise_mag, double _dist_thres,
        int _maximum_allowed_skipped_frames, int _max_trace_length) {
    tracks = new Vector<>();
    dt = _dt;
    Accel_noise_mag = _Accel_noise_mag;
    dist_thres = _dist_thres;
    maximum_allowed_skipped_frames = _maximum_allowed_skipped_frames;
    max_trace_length = _max_trace_length;
    track_removed = 0;
}

double euclideanDist(Point p, Point q) {
    Point diff = new Point(p.x - q.x, p.y - q.y);
    return Math.sqrt(diff.x * diff.x + diff.y * diff.y);
}

public void update(Vector<Rect> rectArray, Vector<Point> detections, Mat imag) {
    this.rectArray = rectArray;
    this.imag = imag;

    if (tracks.size() == 0) {
        // If no tracks yet
        for (int i = 0; i < detections.size(); i++) {
            Track tr = new Track(detections.get(i), dt,
                    Accel_noise_mag, nextTractID++);
            tracks.add(tr);
        }
    }

    int N = tracks.size();
    int M = detections.size();

    double[][] Cost = new double[N][M]; // size: N, M
    assignment.clear();

    for (int i = 0; i < tracks.size(); i++) {
        for (int j = 0; j < detections.size(); j++) {
            Cost[i][j] = euclideanDist(tracks.get(i).prediction, detections.get(j));
        }
    }

    AssignmentOptimal APS = new AssignmentOptimal();
    APS.Solve(Cost, assignment);

    Vector<Integer> not_assigned_tracks = new Vector<>();

    for (int i = 0; i < assignment.size(); i++) {
        if (assignment.get(i) != -1) {
            if (Cost[i][assignment.get(i)] > dist_thres) {
                assignment.set(i, -1);
                not_assigned_tracks.add(i);
            }
        } else {
            tracks.get(i).skipped_frames++;
        }
    }

    for (int i = 0; i < tracks.size(); i++) {
        if (tracks.get(i).skipped_frames > maximum_allowed_skipped_frames) {
            tracks.remove(i);
            assignment.remove(i);
            track_removed++;
            i--;
        }
    }

    Vector<Integer> not_assigned_detections = new Vector<>();
    Vector<Integer> assigned_id = new Vector<>();
    if (tracks.size() <= 23) {
        for (int i = 0; i < detections.size(); i++) {
            if (!assignment.contains(i)) {
                not_assigned_detections.add(i);
            } else {
                assigned_id.add(i);
            }
        }

    }
    if (not_assigned_detections.size() > 0) {
        for (Integer not_assigned_detection : not_assigned_detections) {
            Track tr = new Track(detections.get(not_assigned_detection), dt, Accel_noise_mag, nextTractID++);
            tracks.add(tr);
            trackIds.add(nextTractID);
        }
    }

    // Update Kalman Filters state
    updateKalman(imag, detections);

    for (int j = 0; j < assignment.size(); j++) {
        if (assignment.get(j) != -1) {
            pt1 = new Point(
                    (int) ((rectArray.get(assignment.get(j)).tl().x + rectArray
                    .get(assignment.get(j)).br().x) / 2), rectArray
                    .get(assignment.get(j)).br().y);
            pt2 = new Point(
                    (int) ((rectArray.get(assignment.get(j)).tl().x + rectArray
                    .get(assignment.get(j)).br().x) / 2), rectArray
                    .get(assignment.get(j)).tl().y);
            //show ID
            Imgproc.putText(imag, tracks.get(j).track_id + "", pt2,
                    2 * Core.FONT_HERSHEY_PLAIN, 1, new Scalar(255, 255,
                            255), 1);

            if (tracks.get(j).history.size() < 100) {
                tracks.get(j).history.add(pt1);
            } else {
                tracks.get(j).history.remove(0);
                tracks.get(j).history.add(pt1);
            }

            double x = pt2.x;
            double y = pt2.y;

            if (tracks.get(j).track_id == 0) { //& ((pt2.x - p0.x)<=5 | (pt2.x - p0.x)>= -5) & ((pt2.y - p0.y)<=5 | (pt2.y - p0.y)>= -5)){
                p0 = pt2;
                double dist = euclideanDist(tracks.get(j).trace.get(tracks.get(j).trace.size() - 1), p0);
                pointArray0.add(p0);
                                    //System.out.println("Point1:"+p0);
                //System.out.println(tracks.get(j).trace.get(tracks.get(j).trace.size()-1));
                //System.out.println("Dist:"+dist);//Insert(0, pt2.x, pt2.y,dist); 
            }
            if (tracks.get(j).track_id == 1) {
                p1 = pt2;
                double dist ...

I want to detect and track player on video but when a detected object lost, it get a new tracking ID. I want to get old tracking id with old object. How can I improve it? This code owner tell me " You should config Kalman Filter " . I'm very confuse about it.

public class Kalman extends KalmanFilter {
private KalmanFilter kalman;
private Point LastResult;
private double deltatime;

public void init() {

}

public Kalman(Point pt){
    kalman = new KalmanFilter(4, 2, 0, CvType.CV_32F);

    Mat transitionMatrix = new Mat(4, 4, CvType.CV_32F, new Scalar(0));
    float[] tM = { 
            1, 0, 1, 0, 
            0, 1, 0, 1,
            0, 0, 1, 0,
            0, 0, 0, 1 } ;
    transitionMatrix.put(0,0,tM);

    kalman.set_transitionMatrix(transitionMatrix);

    LastResult = pt;
    Mat statePre = new Mat(4, 1, CvType.CV_32F, new Scalar(0)); // Toa do (x,y), van toc (0,0)
    statePre.put(0, 0, pt.x);
    statePre.put(1, 0, pt.y);
    kalman.set_statePre(statePre);

    kalman.set_measurementMatrix(Mat.eye(2,4, CvType.CV_32F));

    Mat processNoiseCov = Mat.eye(4, 4, CvType.CV_32F);
    processNoiseCov = processNoiseCov.mul(processNoiseCov, 1e-4); //1e-4
    kalman.set_processNoiseCov(processNoiseCov);

    Mat id1 = Mat.eye(2,2, CvType.CV_32F);
    id1 = id1.mul(id1,1e-1);   
    kalman.set_measurementNoiseCov(id1);

    Mat id2 = Mat.eye(4,4, CvType.CV_32F);
    //id2 = id2.mul(id2,0.1);
    kalman.set_errorCovPost(id2);
}

public Kalman(Point pt, double dt, double Accel_noise_mag) {
    kalman = new KalmanFilter(4, 2, 0, CvType.CV_32F);
    deltatime = dt;

    Mat transitionMatrix = new Mat(4, 4, CvType.CV_32F, new Scalar(0));
    float[] tM = { 
            1, 0, 1, 0, 
            0, 1, 0, 1,
            0, 0, 1, 0,
            0, 0, 0, 1 } ;
    transitionMatrix.put(0,0,tM);

    kalman.set_transitionMatrix(transitionMatrix);

    // init
    LastResult = pt;
    Mat statePre = new Mat(4, 1, CvType.CV_32F, new Scalar(0)); // Toa do (x,y), van toc (0,0)
    statePre.put(0, 0, pt.x);
    statePre.put(1, 0, pt.y);
    statePre.put(2, 0, 0);
    statePre.put(3, 0, 0);
    kalman.set_statePre(statePre);

    Mat statePost = new Mat(4, 1, CvType.CV_32F, new Scalar(0));
    statePost.put(0, 0, pt.x);
    statePost.put(1, 0, pt.y);
    statePost.put(2, 0, 0);
    statePost.put(3, 0, 0);
    kalman.set_statePost(statePost);

    kalman.set_measurementMatrix(Mat.eye(2,4, CvType.CV_32F));

    //Mat processNoiseCov = Mat.eye(4, 4, CvType.CV_32F);
    Mat processNoiseCov = new Mat(4, 4, CvType.CV_32F, new Scalar(0));
    float[] dTime = { (float) (Math.pow(deltatime, 4.0) / 4.0), 0,
            (float) (Math.pow(deltatime, 3.0) / 2.0), 0, 0,
            (float) (Math.pow(deltatime, 4.0) / 4.0), 0,
            (float) (Math.pow(deltatime, 3.0) / 2.0),
            (float) (Math.pow(deltatime, 3.0) / 2.0), 0,
            (float) Math.pow(deltatime, 2.0), 0, 0,
            (float) (Math.pow(deltatime, 3.0) / 2.0), 0,
            (float) Math.pow(deltatime, 2.0) };
    processNoiseCov.put(0, 0, dTime);

    processNoiseCov = processNoiseCov.mul(processNoiseCov, Accel_noise_mag); // Accel_noise_mag = 0.5
    kalman.set_processNoiseCov(processNoiseCov);

    Mat id1 = Mat.eye(2,2, CvType.CV_32F);
    id1 = id1.mul(id1,1e-1);
    kalman.set_measurementNoiseCov(id1);

    Mat id2 = Mat.eye(4,4, CvType.CV_32F);
    id2 = id2.mul(id2,.1);
    kalman.set_errorCovPost(id2);
}

public Point getPrediction() {
    Mat prediction ...