pt: extractor algorithm based on OpenCV's SimpleBlobDetector

Tunables are probably wrong, we'll get to that later.

1 files changed, 144 insertions, 22 deletions

diff --git a/ftnoir_tracker_pt/point_extractor.cpp b/ftnoir_tracker_pt/point_extractor.cpp
index dab1d967..86454729 100644
--- a/ftnoir_tracker_pt/point_extractor.cpp
+++ b/ftnoir_tracker_pt/point_extractor.cpp
@@ -28,11 +28,6 @@ std::vector<Vec2f> PointExtractor::extract_points(Mat& frame)
 	// convert to grayscale
 	Mat frame_gray;
     cvtColor(frame, frame_gray, cv::COLOR_RGB2GRAY);
-
-	// mask for everything that passes the threshold (or: the upper threshold of the hysteresis)
-	Mat frame_bin;
-
-    threshold(frame_gray, frame_bin, s.threshold, 255, THRESH_BINARY);
    
     int min_size = s.min_point_size;
     int max_size = s.max_point_size;
@@ -40,30 +35,157 @@ std::vector<Vec2f> PointExtractor::extract_points(Mat& frame)
 	unsigned int region_size_min = 3.14*min_size*min_size/4.0;
 	unsigned int region_size_max = 3.14*max_size*max_size/4.0;
     
-    std::vector<std::vector<cv::Point>> contours;
-    cv::findContours(frame_bin, contours, CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE);
+    // testing indicates threshold difference of 45 from lowest to highest
+    // that's applicable to poor lighting conditions.
+    
+    static constexpr int diff = 20;
+    static constexpr int steps = 10;
+    static constexpr int successes = 8;
+    
+    int thres = s.threshold;
+    
+    struct blob {
+        double max_radius;
+        std::vector<cv::Vec2d> pos;
+        std::vector<double> confids;
+        
+        cv::Vec2d effective_pos() const
+        {
+            double x = 0;
+            double y = 0;
+            double norm = 0;
+            for (unsigned i = 0; i < pos.size(); i++)
+            {
+                x += pos[i][0] * confids[i];
+                y += pos[i][1] * confids[i];
+                norm += confids[i];
+            }
+            cv::Vec2d ret(x, y);
+            ret *= 1./norm;
+            //qDebug() << "ret" << ret[0] << ret[1] << "norm" << norm << "count" << pos.size();
+            return ret;
+        }
+    };
+    
+    struct simple_blob
+    {
+        double radius;
+        cv::Vec2d pos;
+        double confid;
+        bool taken;
+        simple_blob(double radius, const cv::Vec2d& pos, double confid) : radius(radius), pos(pos), confid(confid), taken(false)
+        {
+            //qDebug() << "radius" << radius << "pos" << pos[0] << pos[1] << "confid" << confid;
+        }
+        bool inside(const simple_blob& other)
+        {
+            cv::Vec2d tmp = pos - other.pos;
+            double p = sqrt(1e-4 + tmp.dot(tmp));
+            return p < radius;
+        }
+        static std::vector<blob> merge(std::vector<simple_blob>& blobs)
+        {
+            std::vector<blob> ret;
+            for (unsigned i = 0; i < blobs.size(); i++)
+            {
+                auto& b = blobs[i];
+                if (b.taken)
+                    continue;
+                b.taken = true;
+                blob b_;
+                b_.pos.push_back(b.pos);
+                b_.confids.push_back(b.confid);
+                b_.max_radius = b.radius;
+                
+                for (unsigned j = i+1; j < blobs.size(); j++)
+                {
+                    auto& b2 = blobs[j];
+                    if (b2.taken)
+                        continue;
+                    if (b.inside(b2) || b2.inside(b))
+                    {
+                        b2.taken = true;
+                        b_.pos.push_back(b2.pos);
+                        b_.confids.push_back(b2.confid);
+                        b_.max_radius = std::max(b_.max_radius, b2.radius);
+                    }
+                }
+                if (b_.pos.size() >= successes)
+                    ret.push_back(b_);
+            }
+            return ret;
+        }
+    };
+    
+    // mask for everything that passes the threshold (or: the upper threshold of the hysteresis)
+	Mat frame_bin = cv::Mat::zeros(H, W, CV_8U);
+    
+    const int min = std::max(0, thres - diff/2);
+    const int max = std::min(255, thres + diff/2);
+    const int step = std::max(1, diff / steps);
+    
+    std::vector<simple_blob> blobs;
+    
+    // this code is based on OpenCV SimpleBlobDetector
+    for (int i = min; i < max; i += step)
+    {
+        Mat frame_bin_;
+        threshold(frame_gray, frame_bin_, i, 255, THRESH_BINARY);
+        frame_bin.setTo(170, frame_bin_);
+        
+        std::vector<std::vector<cv::Point>> contours;
+        cv::findContours(frame_bin_, contours, CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE);
+        
+        for (auto& c : contours)
+        {
+            auto m = cv::moments(cv::Mat(c));
+            const double area = m.m00;
+            if (area == 0.)
+                continue;
+            cv::Vec2d pos(m.m10 / m.m00, m.m01 / m.m00);
+            if (area < region_size_min || area > region_size_max)
+                continue;
+            
+            double radius = 0;
+            
+            for (auto& k : c)
+            {
+                cv::Vec2d pos_(k.x, k.y);
+                cv::Vec2d tmp = pos_ - pos;
+                radius = std::max(radius, sqrt(1e-2 + tmp.dot(tmp)));
+            }
+            double confid = 1;
+            {
+                double denominator = std::sqrt(std::pow(2 * m.mu11, 2) + std::pow(m.mu20 - m.mu02, 2));
+                const double eps = 1e-2;
+                if (denominator > eps)
+                {
+                    double cosmin = (m.mu20 - m.mu02) / denominator;
+                    double sinmin = 2 * m.mu11 / denominator;
+                    double cosmax = -cosmin;
+                    double sinmax = -sinmin;
+    
+                    double imin = 0.5 * (m.mu20 + m.mu02) - 0.5 * (m.mu20 - m.mu02) * cosmin - m.mu11 * sinmin;
+                    double imax = 0.5 * (m.mu20 + m.mu02) - 0.5 * (m.mu20 - m.mu02) * cosmax - m.mu11 * sinmax;
+                    confid = imin / imax;
+                }
+            }
+            blobs.push_back(simple_blob(radius, pos, confid));
+        }
+    }
     
     // clear old points
 	points.clear();
     
-    for (auto& c : contours)
+    for (auto& b : simple_blob::merge(blobs))
     {
-        auto m = cv::moments(cv::Mat(c));
-        const double area = m.m00;
-        if (area == 0.)
-            continue;
-        cv::Vec2f pos(m.m10 / m.m00, m.m01 / m.m00);
-        if (area < region_size_min || area > region_size_max)
-            continue;
-        pos[0] =  (pos[0] - W/2)/W;
-        pos[1] = -(pos[1] - H/2)/W;
-        
-        points.push_back(pos);
+        auto pos = b.effective_pos();
+        Vec2f p((pos[0] - W/2)/W, -(pos[1] - H/2)/W);
+        points.push_back(p);
     }
-	
-	// draw output image
+    
+    // draw output image
     vector<Mat> channels;
-    frame_bin.setTo(170, frame_bin);
     channels.push_back(frame_gray + frame_bin);
     channels.push_back(frame_gray - frame_bin);
     channels.push_back(frame_gray - frame_bin);