SURF 特征检测 - OpenCV

Question

我正在开发一个 android 应用程序，其主要目的是检测场景中的询问对象。为此，我使用 OpenCV 的 SURF 算法。我对检测没有“好运”，因为我不知道什么时候“找到”了一个物体。

我用我的设备相机获取一个帧，然后按照以下步骤获取对象的关键点和描述符：

Java 代码

public void onSnapClick(View v) {
    Imgproc.GaussianBlur(frameGray, frameGray, new Size(3, 3), 2);
    Imgproc.Canny(frameGray, frameGray, 40, 120);
    Imgproc.resize(frameGray, frameGray, new Size(320, 240));
    FindFeatures(frameGray.getNativeObjAddr()); //JNI call
    //Some code to store data in DB...
}

JNI 调用

double hessianThreshold=600;
int nOctaves=4;
int nOctaveLayers=2;
bool extended=true;
bool upright=false;

JNIEXPORT void JNICALL Java_es_ugr_reconocimiento_Juego_FindFeatures(JNIEnv* env, jobject, jlong addrGray) {
    Mat& frameGray= *(Mat*) addrGray;
    vector<KeyPoint> keyPoints;
    Mat descriptores;
    SurfFeatureDetector detector_Surf(hessianThreshold, nOctaves, nOctaveLayers, extended, upright);
    SurfDescriptorExtractor extractor_Surf;
    detector_Surf.detect(frameGray, keyPoints);
    if (keyPoints.size() > 0)
        extractor_Surf.compute(frameGray, keyPoints, descriptores);
}

现在我选择要查找的对象，然后按照以下步骤进行操作：

Java 代码

public void onSearchClick(View v) {
    Imgproc.GaussianBlur(frameGray, frameGray, new Size(3, 3), 2);
    Imgproc.Canny(frameGray, frameGray, 40, 120);
    Imgproc.resize(frameGray, frameGray, new Size(320, 240));
    nObject = FindObjects(frameGray.getNativeObjAddr()); //JNI call
    if (nObject = searchObject) 
        //draw frame with a rectangle around the found object in the scenario....
}

JNI 调用

double hessianThreshold=600;
int nOctaves=4;
int nOctaveLayers=2;
bool extended=true;
bool upright=false;

JNIEXPORT jint JNICALL Java_es_ugr_reconocimiento_Juego_FindObjects(JNIEnv* env, jobject, jlong addrGray) {
    Mat& frameGray = *(Mat*) addrGray;
    vector<KeyPoint> keyPoints_esc;
    Mat descriptores_esc;
    SurfFeatureDetector detector_Surf(hessianThreshold, nOctaves, nOctaveLayers, extended, upright);
    SurfDescriptorExtractor extractor_Surf;
    detector_Surf.detect(frameGray , keyPoints_esc);
    if (keyPoints_esc.size() == 0) return -1;
    extractor_Surf.compute(frameGray , keyPoints_esc, descriptores_esc);
    if (descriptores_esc.rows() == 0) return -1;

    for(int i=0;i<lstObjects.size();i++){
        Mat descriptores_obj = lstDescriptors.at(i);
        vector<KeyPoint> keyPoints_obj = lstKeyPoints.at(i);

        FlannBasedMatcher matcher;
        vector<vector<DMatch> > matches;
        matcher.knnMatch(descriptores_obj, descriptores_esc, matches, 2);
        // ----------------------------------------------------------------------
        // Draw only "good" matches (i.e. whose distance is less than 2*min_dist,
        // or a small arbitary value ( 0.02 ) in the event that min_dist is very
        // small)
        // PS.- radiusMatch can also be used here.
        // ----------------------------------------------------------------------
        vector<DMatch> good_matches;
        //THIS LOOP IS SENSITIVE TO SEGFAULTS
        for (int i = 0; i < min(descriptores_obj.rows - 1, (int) matches.size());i++){
            if ( (matches[i][0].distance < 0.6 * (matches[i][1].distance)) && 
                 ((int) matches[i].size() <= 2 && (int) matches[i].size() > 0) ) {
                    good_matches.push_back(matches[i][0]);
            }
        }

        if (good_matches.size() >= nThreshold) {
            vector < Point2f > obj;
            vector < Point2f > scene;

            for (int i = 0; i < good_matches.size(); i++) {
                //-- Get the keypoints from the good matches
                obj.push_back(keyPoints_obj[good_matches[i].queryIdx].pt);
                scene.push_back(keyPoints_esc[good_matches[i].trainIdx].pt);
            }

            Mat H = findHomography(obj, scene, CV_RANSAC);

            vector<Point2f> obj_corners(4);
            obj_corners[0] = cvPoint(0, 0);
            obj_corners[1] = cvPoint(240, 0);
            obj_corners[2] = cvPoint(240, 320);
            obj_corners[3] = cvPoint(0, 320);
            vector<Point2f> scene_corners(4);

            perspectiveTransform(obj_corners, scene_corners, H);

            line(frameGray, scene_corners[0], scene_corners[1], Scalar(255, 0, 0), 4);
            line(frameGray, scene_corners[1], scene_corners[2], Scalar(255, 0, 0), 4);
            line(frameGray, scene_corners[2], scene_corners[3], Scalar(255, 0, 0), 4);
            line(frameGray, scene_corners[3], scene_corners[0], Scalar(255, 0, 0), 4);

            for (unsigned int i = 0; i < scene.size(); i++) {
                const Point2f& kp = scene[i];
                circle(frameGray, Point(kp.x, kp.y), 10, Scalar(255, 255, 255, 255));
            }

            return i; //position of the matched object

        }

    }
}

我不知道在这个比较中哪个阈值可能是最好的

if (good_matches.size() >= nThreshold) // do findHomography...

我一直在搜索，几乎我发现的每个代码都包含数字 4 作为 nThreshold，但对我来说它效果不佳。我的代码几乎每次都“找到”一个对象。

还有其他更好的方法吗？就像使用不同的匹配器或其他阈值，或者试图弄清楚进行单应性是否会创建类似于矩形的东西（我之所以这么说是因为有时它“找到”了一些东西，但绘制了四条线而不是构建一个矩形）。

Answer 1

请在您的代码中进行以下更改

int nThreshold= 100;
       if (good_matches.size() >= nThreshold) 
        {
        continue; // This line is to prevent further steps of matching if there are too many good matches (Lot of ambiguous points results in false match)
        }
        vector < Point2f > obj;
        vector < Point2f > scene;

        for (int i = 0; i < good_matches.size(); i++) {
            //-- Get the keypoints from the good matches
            obj.push_back(keyPoints_obj[good_matches[i].queryIdx].pt);
            scene.push_back(keyPoints_esc[good_matches[i].trainIdx].pt);
               }

// Skip doing homography if the object and scene contains less than four points(cant draw a rectangle if less than 4 points, hence your program will crash here if you do not handle the exception)
      if(obj.size() < 4 || scene.size() < 4)
       {
       continue;
       }

       Mat H = findHomography(obj, scene, CV_RANSAC);