OpenCV (C# Emgu) 物件追蹤程式，使用 Camshift 演算法

這學期擔任影像處理助教，所以要出作業給學弟妹練習一下 OpenCV
會使用 Emgu 不外乎是環境好建置，只要把將範例專案做好基本上下載回去能馬上使用
可以專心在撰寫功能不需要去建置環境，雖然說建置環境對一個工程師也是一項基本功啦…

首先先來搞清楚一下整個物件追蹤的來龍去脈…

整個流程

1. 畫面上選取要追蹤的區域這裡我稱為 物件 (Object)，也就是要被追蹤的東東
2. 計算出物件的 直方圖 (Histogram)
3. 將物件的直方圖與輸入畫面進行 反向投影(Backproject) 計算
4. 透過 物件追蹤演算法(Camshift) 快速地框出反向投影計算出的結果
5. 回傳追蹤區域
於是一直重複 3~5 的過程，我們的物件追蹤程式就大功告成了！

 

範例專案 ObjectTracking.cs

先宣告所有變數

public Image<hsv, byte=""> hsv;
public Image<gray, byte=""> hue;
public Image<gray, byte=""> mask;
public Image<gray, byte=""> backproject;
public DenseHistogram hist;
private Rectangle trackingWindow;
private MCvConnectedComp trackcomp;
private MCvBox2D trackbox;

建構子

public ObjectTracking(Image<Bgr, Byte> image, Rectangle ROI)
{
    // Initialize parameters
    trackbox = new MCvBox2D();
    trackcomp = new MCvConnectedComp();
    hue = new Image<Gray, byte>(image.Width, image.Height);
    hue._EqualizeHist();
    mask = new Image<Gray, byte>(image.Width, image.Height);
    hist = new DenseHistogram(30, new RangeF(0, 180));
    backproject = new Image<Gray, byte>(image.Width, image.Height);

    // Assign Object's ROI from source image.
    trackingWindow = ROI;

    // Producing Object's hist
    CalObjectHist(image);
}

更新 HUE 畫面

private void UpdateHue(Image<Bgr, Byte> image)
{
    // release previous image memory
    if (hsv != null)    hsv.Dispose();
    hsv = image.Convert<Hsv, Byte>();

    // Drop low saturation pixels
    mask = hsv.Split()[1].ThresholdBinary(new Gray(60), new Gray(255));
    CvInvoke.cvInRangeS(hsv, new MCvScalar(0, 30, Math.Min(10, 255), 0),
        new MCvScalar(180, 256, Math.Max(10, 255), 0), mask);

    // Get Hue
    hue = hsv.Split()[0];
}

計算物件直方圖

private void CalObjectHist(Image<Bgr, Byte> image)
{
    UpdateHue(image);

    // Set tracking object's ROI
    hue.ROI = trackingWindow;
    mask.ROI = trackingWindow;
    hist.Calculate(new Image<Gray, Byte>[] { hue }, false, mask);

    // Scale Historgram
    float max=0, min=0, scale=0;
    int[] minLocations, maxLocations;
    hist.MinMax(out min, out max, out minLocations, out maxLocations);
    if (max != 0)
    {
        scale = 255 / max;
    }
    CvInvoke.cvConvertScale(hist.MCvHistogram.bins, hist.MCvHistogram.bins, scale, 0);

    // Clear ROI
    hue.ROI = System.Drawing.Rectangle.Empty;
    mask.ROI = System.Drawing.Rectangle.Empty;

    // Now we have Object's Histogram, called hist.
}

執行追蹤

public Rectangle Tracking(Image<Bgr, Byte> image)
{
    UpdateHue(image);

    // Calucate BackProject
    backproject = hist.BackProject(new Image<Gray, Byte>[] { hue });

    // Apply mask
    backproject._And(mask);

    // Tracking windows empty means camshift lost bounding-box last time
    // here we give camshift a new start window from 0,0 (you could change it)
    if (trackingWindow.IsEmpty || trackingWindow.Width==0 || trackingWindow.Height==0)
    {
        trackingWindow = new Rectangle(0, 0, 100, 100);
    }
    CvInvoke.cvCamShift(backproject, trackingWindow,
        new MCvTermCriteria(10, 1), out trackcomp, out trackbox);

    // update tracking window
    trackingWindow = trackcomp.rect;

    return trackingWindow;
}

完整範例專案放置於 Github