MAT4015: Fundamentals of applied probability and random processes

curriculum

OpenCV를 활용한 Mean Shift 기반 Object Tracking 구현

Object tracking

Object tracking based on color histogram and real-time back projection

Process

  1. Modeling color histogram of selected object
  2. Calculate histogram similarity
  3. Back projection
  4. Mean shift process (clustering of pixels with back projected weights)
  5. Update object position and parameters after tracking

Additional functions

  • Dynamic navigation area
  • Dynamic change of object color histogram
  • Dynamic change of object size

Codes

  • Calculate color histogram
double * hists = (double*)calloc(param.hist_bins, sizeof(double)), hist_size = 256 / param.hist_bins;
int eWidth = min(rc.x + rc.width, img.size().width),
    eHeight = min(rc.y + rc.height, img.size().height);
for (int i = rc.x; i < eWidth; ++i)
    for (int j = rc.y; j < eHeight; ++j)
        hists[matrixAt(img, i, j) / (int)hist_size]++;


double total = 0;
for (int i = 0; i < param.hist_bins; ++i)
    total += hists[i] * hists[i];
total = sqrt(total);
for (int i = 0; i < param.hist_bins; ++i)
    hists[i] /= total;


return hists;
  • Back projection
Mat imx = hsv.clone();
for (int i = 0; i < hsv.cols; ++i)
    for (int j = 0; j < hsv.rows; ++j)
    {
        w = this->objectHists[matrixAt(imx, i, j) / param.hist_bins];
        double pixel[] = { 255 * w, 255 * w, 255 * w };
        matrixSet(imx, i, j, pixel);
    }
  • Mean shift
Mat hsv;
cvtColor(img, hsv, CV_BGR2HSV);
Rect nRect = myRect, temp, bRect = myRect;
double nX = 0, nY = 0, tW = 0, w;
do {
    tW = 0, nX = 0; nY = 0; myRect = nRect;


    // set searching area
    temp = Rect(max(nRect.x - (int)param.search_range, 0), max(nRect.y - (int)param.search_range, 0),
        min(nRect.width + (int)param.search_range * 2, hsv.rows), min(nRect.height + (int)param.search_range * 2, hsv.cols));
    double sRatioWidth = temp.width / param.sampling, sRatioHeight = temp.height / param.sampling;
    for (int i = 0; i <= temp.width / 2; i+=sRatioWidth)
        for (int j = 0; j <= temp.height / 2; j+=sRatioHeight)
        {
            w = mySimilarity(hsv, Rect(temp.x + (temp.width / 2) + i, temp.y + (temp.height / 2) + j, nRect.width, nRect.height), this->objectHists);
            tW += w;
            nX += w * (temp.x + (temp.width / 2) + i);
            nY += w * (temp.y + (temp.height / 2) + j);
            if (i != 0 || j != 0)
            {
                w = mySimilarity(hsv, Rect(temp.x + (temp.width / 2) - i, temp.y + (temp.height / 2) - j, nRect.width, nRect.height), this->objectHists);
                tW += w;
                nX += w * (temp.x + (temp.width / 2) - i);
                nY += w * (temp.y + (temp.height / 2) - j);
            }
        }
    nX /= tW;
    nY /= tW;


    // get mean of w, if w > 0.25 make range narrow, or not make extend;
    double twRatio = tW / ((temp.width / sRatioWidth) * (temp.height / sRatioHeight));
    param.search_range *= (twRatio > (EXTEND_LIMIT) ? 1 - twRatio : (1 - EXTEND_LIMIT) +twRatio);


    nRect = Rect(max((int)nX - nRect.width / 2, 0), max((int)nY - nRect.height / 2, 0), myRect.width, myRect.height);
} while (sqrt(pow(myRect.x - nRect.x, 2) + pow(myRect.y - nRect.y, 2)) > param.search_range);