summaryrefslogtreecommitdiffhomepage
path: root/tracker-pt/point_extractor.cpp
blob: 4bd92d44d2f2ca39cd03476ede9730a6a33f62fc (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
/* Copyright (c) 2012 Patrick Ruoff
 * Copyright (c) 2015-2016 Stanislaw Halik <sthalik@misaki.pl>
 *
 * Permission to use, copy, modify, and/or distribute this software for any
 * purpose with or without fee is hereby granted, provided that the above
 * copyright notice and this permission notice appear in all copies.
 */

#include "point_extractor.h"
#include "compat/util.hpp"
#include "point_tracker.h"
#include <QDebug>

#include <opencv2/videoio.hpp>

#include <cmath>
#include <algorithm>
#include <cinttypes>

#include <QDebug>

using namespace types;

/*
http://en.wikipedia.org/wiki/Mean-shift
In this application the idea, is to eliminate any bias of the point estimate 
which is introduced by the rather arbitrary thresholded area. One must recognize
that the thresholded area can only move in one pixel increments since it is
binary. Thus, its center of mass might make "jumps" as pixels are added/removed
from the thresholded area.
With mean-shift, a moving "window" or kernel is multiplied with the gray-scale
image, and the COM is calculated of the result. This is iterated where the
kernel center is set the previously computed COM. Thus, peaks in the image intensity
distribution "pull" the kernel towards themselves. Eventually it stops moving, i.e.
then the computed COM coincides with the kernel center. We hope that the 
corresponding location is a good candidate for the extracted point.
The idea similar to the window scaling suggested in  Berglund et al. "Fast, bias-free 
algorithm for tracking single particles with variable size and shape." (2008).
*/
static cv::Vec2d MeanShiftIteration(const cv::Mat &frame_gray, const cv::Vec2d &current_center, double filter_width)
{
    // Most amazingling this function runs faster with doubles than with floats.
    const double s = 1.0 / filter_width;

    double m = 0;
    cv::Vec2d com(0.0, 0.0);
    for (int i = 0; i < frame_gray.rows; i++)
    {
        auto frame_ptr = (uint8_t *)frame_gray.ptr(i);
        for (int j = 0; j < frame_gray.cols; j++)
        {
            double val = frame_ptr[j];
            val = val * val; // taking the square wights brighter parts of the image stronger.
            {
                double dx = (j - current_center[0])*s;
                double dy = (i - current_center[1])*s;
                double f = std::max(0.0, 1.0 - dx*dx - dy*dy);
                val *= f;
            }
            m += val;
            com[0] += j * val;
            com[1] += i * val;
        }
    }
    if (m > 0.1)
    {
        com *= 1.0 / m;
        return com;
    }
    else
        return current_center;
}

PointExtractor::PointExtractor()
{
    blobs.reserve(max_blobs);
}

void PointExtractor::extract_points(const cv::Mat& frame, cv::Mat& preview_frame, std::vector<vec2>& points)
{
    using std::sqrt;
    using std::max;
    using std::round;
    using std::sort;

    if (frame_gray.rows != frame.rows || frame_gray.cols != frame.cols)
    {
        frame_gray = cv::Mat(frame.rows, frame.cols, CV_8U);
        frame_bin = cv::Mat(frame.rows, frame.cols, CV_8U);
        frame_blobs = cv::Mat(frame.rows, frame.cols, CV_8U);
    }

    // convert to grayscale
    cv::cvtColor(frame, frame_gray, cv::COLOR_RGB2GRAY);

    const double region_size_min = s.min_point_size;
    const double region_size_max = s.max_point_size;

    if (!s.auto_threshold)
    {
        const int thres = s.threshold;
        cv::threshold(frame_gray, frame_bin, thres, 255, cv::THRESH_BINARY);
    }
    else
    {
        cv::calcHist(std::vector<cv::Mat> { frame_gray },
                     std::vector<int> { 0 },
                     cv::Mat(),
                     hist,
                     std::vector<int> { 256 },
                     std::vector<float> { 0, 256 },
                     false);

        static constexpr double min_radius = 2.5;
        static constexpr double max_radius = 15;

        const double radius = max(0., (max_radius-min_radius) * s.threshold / 255 + min_radius);
        const float* ptr = reinterpret_cast<const float*>(hist.ptr(0));
        const unsigned area = unsigned(round(3 * M_PI * radius*radius));
        const unsigned sz = unsigned(hist.cols * hist.rows);
        unsigned thres = 1;
        for (unsigned i = sz-1, cnt = 0; i > 1; i--)
        {
            cnt += ptr[i];
            if (cnt >= area)
            {
                thres = i;
                break;
            }
        }
        //val *= 240./256.;
        //qDebug() << "thres" << thres;

        cv::threshold(frame_gray, frame_bin, thres, 255, CV_THRESH_BINARY);
    }

    blobs.clear();
    frame_bin.copyTo(frame_blobs);

    unsigned idx = 0;
    for (int y=0; y < frame_blobs.rows; y++)
    {
        const unsigned char* ptr_bin = frame_blobs.ptr(y);
        for (int x=0; x < frame_blobs.cols; x++)
        {
            if (ptr_bin[x] != 255)
                continue;
            idx = blobs.size() + 1;

            cv::Rect rect;
            cv::floodFill(frame_blobs,
                          cv::Point(x,y),
                          cv::Scalar(idx),
                          &rect,
                          cv::Scalar(0),
                          cv::Scalar(0),
                          8);

            // these are doubles since m10 and m01 could overflow theoretically
            // log2(255^2 * 640^2 * pi) > 36
            double m10 = 0;
            double m01 = 0;
            // norm can't overflow since there's no 640^2 component
            int norm = 0;
            int cnt = 0;

            for (int i=rect.y; i < (rect.y+rect.height); i++)
            {
                unsigned char* ptr_blobs = frame_blobs.ptr(i);
                const unsigned char* ptr_gray = frame_gray.ptr(i);
                for (int j=rect.x; j < (rect.x+rect.width); j++)
                {
                    if (ptr_blobs[j] != idx)
                        continue;

                    ptr_blobs[j] = 0;

                    // square as a weight gives better results
                    const int val(int(ptr_gray[j]) * int(ptr_gray[j]));

                    norm += val;
                    m01 += i * val;
                    m10 += j * val;
                    cnt++;
                }
            }
            if (norm > 0)
            {
                const double radius = sqrt(cnt / M_PI), N = double(norm);
                if (radius > region_size_max || radius < region_size_min)
                    continue;

                blob b(radius, cv::Vec2d(m10 / N, m01 / N), N/sqrt(double(cnt)), rect);
                blobs.push_back(b);

                {
                    static const f offx = 10, offy = 7.5;
                    const f cx = preview_frame.cols / f(frame.cols),
                            cy = preview_frame.rows / f(frame.rows);
                    cv::Point p(iround(b.pos[0] * cx), iround(b.pos[1] * cy));

                    static const cv::Scalar color(0, 255, 0);

                    cv::circle(preview_frame, p, b.radius, color);

                    {
                        const int len = iround(b.radius * 1.25) + 1;
                        cv::line(preview_frame,
                                 cv::Point(p.x - len, p.y),
                                 cv::Point(p.x + len, p.y),
                                 color);
                        cv::line(preview_frame,
                                 cv::Point(p.x, p.y - len),
                                 cv::Point(p.x, p.y + len),
                                 color);
                    }

                    char buf[64];
                    sprintf(buf, "%.2fpx", radius);

                    cv::putText(preview_frame,
                                buf,
                                cv::Point(iround(b.pos[0]*cx+offx), iround(b.pos[1]*cy+offy)),
                                cv::FONT_HERSHEY_PLAIN,
                                1,
                                cv::Scalar(0, 0, 255),
                                1);
                }

                if (idx >= max_blobs) goto end;
            }
        }
    }
end:

    sort(blobs.begin(), blobs.end(), [](const blob& b1, const blob& b2) -> bool { return b2.brightness < b1.brightness; });

    const int W = frame.cols;
    const int H = frame.rows;

    for (idx = 0; idx < std::min(PointModel::N_POINTS, unsigned(blobs.size())); ++idx)
    {
        blob &b = blobs[idx];
        cv::Rect rect = b.rect;

        rect.x -= rect.width / 2;
        rect.y -= rect.height / 2;
        rect.width *= 2;
        rect.height *= 2;
        rect &= cv::Rect(0, 0, W, H);  // crop at frame boundaries

        cv::Mat frame_roi = frame_gray(rect);

        // smaller values mean more changes. 1 makes too many changes while 1.5 makes about .1
        // seems values close to 1.3 reduce noise best with about .15->.2 changes
        static constexpr double radius_c = 1.3;

        const double kernel_radius = b.radius * radius_c;
        cv::Vec2d pos(b.pos[0] - rect.x, b.pos[1] - rect.y); // position relative to ROI.

        for (int iter = 0; iter < 10; ++iter)
        {
            cv::Vec2d com_new = MeanShiftIteration(frame_roi, pos, kernel_radius);
            cv::Vec2d delta = com_new - pos;
            pos = com_new;
            if (delta.dot(delta) < 1e-3)
                break;
        }

        b.pos[0] = pos[0] + rect.x;
        b.pos[1] = pos[1] + rect.y;
    }

    // End of mean shift code. At this point, blob positions are updated which hopefully less noisy less biased values.
    points.reserve(max_blobs);
    points.clear();

    for (const auto& b : blobs)
    {
        // note: H/W is equal to fx/fy

        vec2 p((b.pos[0] - W/2)/W, -(b.pos[1] - H/2)/W);
        points.push_back(p);
    }
}

PointExtractor::blob::blob(double radius, const cv::Vec2d& pos, double brightness, cv::Rect& rect) :
    radius(radius), brightness(brightness), pos(pos), rect(rect)
{
    //qDebug() << "radius" << radius << "pos" << pos[0] << pos[1];
}