5 #include "threadctx.hpp"
13 DbgTracer __dbgTracer;
16 T clamp(const T& n, const T& lower, const T& upper)
18 return std::max(lower, std::min(n, upper));
22 void clamp2(T& n, const T& lower, const T& upper)
24 n = std::max(lower, std::min(n, upper));
27 #if CV_MAJOR_VERSION < 3
28 template<typename _Tp> static inline
29 cv::Size_<_Tp> operator / (const cv::Size_<_Tp>& a, _Tp b)
31 return cv::Size_<_Tp>(a.width / b, a.height / b);
34 template<typename _Tp> static inline
35 cv::Point_<_Tp> operator / (const cv::Point_<_Tp>& a, double b)
37 return cv::Point_<_Tp>(a.x / b, a.y / b);
42 class Kcf_Tracker_Private {
45 Kcf_Tracker_Private(const KCF_Tracker &kcf) : kcf(kcf) {}
47 const KCF_Tracker &kcf;
49 std::vector<ThreadCtx> threadctxs;
51 ScaleRotVector<ThreadCtx> threadctxs{kcf.p_scales, kcf.p_angles};
55 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
56 double output_sigma_factor, int cell_size)
57 : p_cell_size(cell_size), fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
58 p_lambda(lambda), p_interp_factor(interp_factor)
62 KCF_Tracker::KCF_Tracker() : fft(*new FFT()) {}
64 KCF_Tracker::~KCF_Tracker()
69 void KCF_Tracker::train(cv::Mat input_rgb, cv::Mat input_gray, double interp_factor)
73 // obtain a sub-window for training
74 get_features(input_rgb, input_gray, nullptr, p_current_center.x, p_current_center.y,
75 p_windows_size.width, p_windows_size.height,
76 p_current_scale, p_current_angle).copyTo(model->patch_feats.scale(0));
77 DEBUG_PRINT(model->patch_feats);
78 fft.forward_window(model->patch_feats, model->xf, model->temp);
79 DEBUG_PRINTM(model->xf);
80 model->model_xf = model->model_xf * (1. - interp_factor) + model->xf * interp_factor;
81 DEBUG_PRINTM(model->model_xf);
83 if (m_use_linearkernel) {
84 ComplexMat xfconj = model->xf.conj();
85 model->model_alphaf_num = xfconj.mul(model->yf);
86 model->model_alphaf_den = (model->xf * xfconj);
88 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
89 cv::Size sz(Fft::freq_size(feature_size));
90 ComplexMat kf(sz.height, sz.width, 1);
91 (*gaussian_correlation)(kf, model->model_xf, model->model_xf, p_kernel_sigma, true, *this);
93 model->model_alphaf_num = model->yf * kf;
94 model->model_alphaf_den = kf * (kf + p_lambda);
96 model->model_alphaf = model->model_alphaf_num / model->model_alphaf_den;
97 DEBUG_PRINTM(model->model_alphaf);
98 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
101 static int round_pw2_down(int x)
103 for (int i = 1; i < 32; i <<= 1)
110 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
112 __dbgTracer.debug = m_debug;
115 // check boundary, enforce min size
116 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
118 if (x2 > img.cols - 1) x2 = img.cols - 1;
120 if (y2 > img.rows - 1) y2 = img.rows - 1;
122 if (x2 - x1 < 2 * p_cell_size) {
123 double diff = (2 * p_cell_size - x2 + x1) / 2.;
124 if (x1 - diff >= 0 && x2 + diff < img.cols) {
127 } else if (x1 - 2 * diff >= 0) {
133 if (y2 - y1 < 2 * p_cell_size) {
134 double diff = (2 * p_cell_size - y2 + y1) / 2.;
135 if (y1 - diff >= 0 && y2 + diff < img.rows) {
138 } else if (y1 - 2 * diff >= 0) {
145 p_init_pose.w = x2 - x1;
146 p_init_pose.h = y2 - y1;
147 p_init_pose.cx = x1 + p_init_pose.w / 2.;
148 p_init_pose.cy = y1 + p_init_pose.h / 2.;
150 cv::Mat input_gray, input_rgb = img.clone();
151 if (img.channels() == 3) {
152 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
153 input_gray.convertTo(input_gray, CV_32FC1);
155 img.convertTo(input_gray, CV_32FC1);
157 // don't need too large image
158 if (p_init_pose.w * p_init_pose.h > 100. * 100.) {
159 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
160 p_resize_image = true;
161 p_init_pose.scale(p_downscale_factor);
162 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
163 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
166 // compute win size + fit to fhog cell size
167 p_windows_size.width = round(p_init_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
168 p_windows_size.height = round(p_init_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
170 if (fit_size_x == 0 || fit_size_y == 0) {
171 // Round down to the next highest power of 2
172 fit_size = cv::Size(round_pw2_down(p_windows_size.width),
173 round_pw2_down(p_windows_size.height));
174 } else if (fit_size_x == -1 || fit_size_y == -1) {
175 fit_size = p_windows_size;
177 fit_size = cv::Size(fit_size_x, fit_size_y);
180 feature_size = fit_size / p_cell_size;
183 for (int i = -int(p_num_scales - 1) / 2; i <= int(p_num_scales) / 2; ++i)
184 p_scales.push_back(std::pow(p_scale_step, i));
187 for (int i = -int(p_num_angles - 1) / 2; i <= int(p_num_angles) / 2; ++i)
188 p_angles.push_back(i * p_angle_step);
191 if (m_use_linearkernel) {
192 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
193 std::exit(EXIT_FAILURE);
197 model.reset(new Model(feature_size, p_num_of_feats));
198 d.reset(new Kcf_Tracker_Private(*this));
201 for (auto scale: p_scales)
202 for (auto angle : p_angles)
203 d->threadctxs.emplace_back(feature_size, (int)p_num_of_feats, scale, angle);
205 d->threadctxs.emplace_back(feature_size, (int)p_num_of_feats, p_scales, p_angles);
208 gaussian_correlation.reset(new GaussianCorrelation(1, p_num_of_feats, feature_size));
210 p_current_center = p_init_pose.center();
211 p_current_scale = 1.;
213 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
214 double max_size_ratio =
215 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
216 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
217 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
218 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
220 std::cout << "init: img size " << img.size() << std::endl;
221 std::cout << "init: win size " << p_windows_size;
222 if (p_windows_size != fit_size)
223 std::cout << " resized to " << fit_size;
224 std::cout << std::endl;
225 std::cout << "init: FFT size " << feature_size << std::endl;
226 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
228 p_output_sigma = std::sqrt(p_init_pose.w * p_init_pose.h * double(fit_size.area()) / p_windows_size.area())
229 * p_output_sigma_factor / p_cell_size;
231 fft.init(feature_size.width, feature_size.height, p_num_of_feats, p_num_scales * p_num_angles);
232 fft.set_window(MatDynMem(cosine_window_function(feature_size.width, feature_size.height)));
234 // window weights, i.e. labels
235 MatScales gsl(1, feature_size);
236 gaussian_shaped_labels(p_output_sigma, feature_size.width, feature_size.height).copyTo(gsl.plane(0));
237 fft.forward(gsl, model->yf);
238 DEBUG_PRINTM(model->yf);
240 // train initial model
241 train(input_rgb, input_gray, 1.0);
244 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
246 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
249 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
252 if (p_resize_image) {
253 tmp.scale(p_downscale_factor);
255 p_current_center = tmp.center();
258 BBox_c KCF_Tracker::getBBox()
261 tmp.cx = p_current_center.x;
262 tmp.cy = p_current_center.y;
263 tmp.w = p_init_pose.w * p_current_scale;
264 tmp.h = p_init_pose.h * p_current_scale;
265 tmp.a = p_current_angle;
268 tmp.scale(1 / p_downscale_factor);
273 double KCF_Tracker::getFilterResponse() const
275 return this->max_response;
278 void KCF_Tracker::resizeImgs(cv::Mat &input_rgb, cv::Mat &input_gray)
280 if (p_resize_image) {
281 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
282 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
286 static void drawCross(cv::Mat &img, cv::Point center, bool green)
288 cv::Scalar col = green ? cv::Scalar(0, 1, 0) : cv::Scalar(0, 0, 1);
289 cv::line(img, cv::Point(center.x, 0), cv::Point(center.x, img.size().height), col);
290 cv::line(img, cv::Point(0, center.y), cv::Point(img.size().height, center.y), col);
293 static cv::Point2d wrapAroundFreq(cv::Point2d pt, cv::Mat &resp_map)
295 if (pt.y > resp_map.rows / 2) // wrap around to negative half-space of vertical axis
296 pt.y = pt.y - resp_map.rows;
297 if (pt.x > resp_map.cols / 2) // same for horizontal axis
298 pt.x = pt.x - resp_map.cols;
302 double KCF_Tracker::findMaxReponse(uint &max_idx, cv::Point2d &new_location) const
305 const auto &vec = IF_BIG_BATCH(d->threadctxs[0].max, d->threadctxs);
308 auto max_it = std::max_element(vec.begin(), vec.end(),
309 [](const ThreadCtx &a, const ThreadCtx &b)
310 { return a.max.response < b.max.response; });
312 auto max_it = std::max_element(vec.begin(), vec.end(),
313 [](const ThreadCtx::Max &a, const ThreadCtx::Max &b)
314 { return a.response < b.response; });
316 assert(max_it != vec.end());
317 max = max_it->IF_BIG_BATCH(response, max.response);
319 max_idx = std::distance(vec.begin(), max_it);
321 cv::Point2i max_response_pt = IF_BIG_BATCH(max_it->loc, max_it->max.loc);
322 cv::Mat max_response_map = IF_BIG_BATCH(d->threadctxs[0].response.plane(max_idx),
323 max_it->response.plane(0));
325 DEBUG_PRINTM(max_response_map);
326 DEBUG_PRINT(max_response_pt);
328 max_response_pt = wrapAroundFreq(max_response_pt, max_response_map);
330 // sub pixel quadratic interpolation from neighbours
331 if (m_use_subpixel_localization) {
332 new_location = sub_pixel_peak(max_response_pt, max_response_map);
334 new_location = max_response_pt;
336 DEBUG_PRINT(new_location);
338 if (m_visual_debug != vd::NONE) {
340 int w = fit ? 100 : (m_visual_debug == vd::PATCH ? fit_size.width : feature_size.width);
341 int h = fit ? 100 : (m_visual_debug == vd::PATCH ? fit_size.height : feature_size.height);
342 cv::Mat all_responses((h + 1) * p_num_scales - 1,
343 (w + 1) * p_num_angles - 1, CV_32FC3, cv::Scalar::all(0));
344 for (size_t i = 0; i < p_num_scales; ++i) {
345 for (size_t j = 0; j < p_num_angles; ++j) {
346 auto &threadctx = d->IF_BIG_BATCH(threadctxs[0], threadctxs(i, j));
348 cv::Point2d cross = threadctx.IF_BIG_BATCH(max(i, j), max).loc;
349 cross = wrapAroundFreq(cross, max_response_map);
350 if (m_visual_debug == vd::PATCH ) {
351 threadctx.dbg_patch IF_BIG_BATCH((i, j),)
352 .convertTo(tmp, all_responses.type(), 1.0 / 255);
353 cross.x = cross.x / fit_size.width * tmp.cols + tmp.cols / 2;
354 cross.y = cross.y / fit_size.height * tmp.rows + tmp.rows / 2;
356 cv::cvtColor(threadctx.response.plane(IF_BIG_BATCH(threadctx.max.getIdx(i, j), 0)),
357 tmp, cv::COLOR_GRAY2BGR);
358 tmp /= max; // Normalize to 1
359 cross += cv::Point2d(tmp.size())/2;
360 tmp = circshift(tmp, -tmp.cols/2, -tmp.rows/2);
361 //drawCross(tmp, cross, false);
364 if (&*max_it == &IF_BIG_BATCH(threadctx.max(i, j), threadctx)) {
365 // Show the green cross at position of sub-pixel interpolation (if enabled)
366 cross = new_location + cv::Point2d(tmp.size())/2;
369 // Move to the center of pixes (if scaling up) and scale
370 cross.x = (cross.x + 0.5) * double(w)/tmp.cols;
371 cross.y = (cross.y + 0.5) * double(h)/tmp.rows;
372 cv::resize(tmp, tmp, cv::Size(w, h)); //, 0, 0, cv::INTER_NEAREST);
373 drawCross(tmp, cross, green);
374 cv::Mat resp_roi(all_responses, cv::Rect(j * (w+1), i * (h+1), w, h));
375 tmp.copyTo(resp_roi);
378 cv::namedWindow("KCF visual debug", CV_WINDOW_AUTOSIZE);
379 cv::imshow("KCF visual debug", all_responses);
385 void KCF_Tracker::track(cv::Mat &img)
387 __dbgTracer.debug = m_debug;
390 cv::Mat input_gray, input_rgb = img.clone();
391 if (img.channels() == 3) {
392 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
393 input_gray.convertTo(input_gray, CV_32FC1);
395 img.convertTo(input_gray, CV_32FC1);
397 // don't need too large image
398 resizeImgs(input_rgb, input_gray);
401 for (auto &it : d->threadctxs)
402 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
403 it.track(*this, input_rgb, input_gray);
405 for (auto const &it : d->threadctxs)
409 NORMAL_OMP_PARALLEL_FOR
410 for (uint i = 0; i < d->threadctxs.size(); ++i)
411 d->threadctxs[i].track(*this, input_rgb, input_gray);
414 cv::Point2d new_location;
416 max_response = findMaxReponse(max_idx, new_location);
418 double angle_change = m_use_subgrid_angle ? sub_grid_angle(max_idx)
419 : d->IF_BIG_BATCH(threadctxs[0].max, threadctxs).angle(max_idx);
420 p_current_angle += angle_change;
422 new_location.x = new_location.x * cos(-p_current_angle/180*M_PI) + new_location.y * sin(-p_current_angle/180*M_PI);
423 new_location.y = new_location.y * cos(-p_current_angle/180*M_PI) - new_location.x * sin(-p_current_angle/180*M_PI);
425 new_location.x *= double(p_windows_size.width) / fit_size.width;
426 new_location.y *= double(p_windows_size.height) / fit_size.height;
428 p_current_center += p_current_scale * p_cell_size * new_location;
430 clamp2(p_current_center.x, 0.0, img.cols - 1.0);
431 clamp2(p_current_center.y, 0.0, img.rows - 1.0);
433 // sub grid scale interpolation
434 if (m_use_subgrid_scale) {
435 p_current_scale *= sub_grid_scale(max_idx);
437 p_current_scale *= d->IF_BIG_BATCH(threadctxs[0].max, threadctxs).scale(max_idx);
440 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
443 // train at newly estimated target position
444 train(input_rgb, input_gray, p_interp_factor);
447 void ThreadCtx::track(const KCF_Tracker &kcf, cv::Mat &input_rgb, cv::Mat &input_gray)
451 BIG_BATCH_OMP_PARALLEL_FOR
452 for (uint i = 0; i < IF_BIG_BATCH(max.size(), 1); ++i)
454 kcf.get_features(input_rgb, input_gray, &dbg_patch IF_BIG_BATCH([i],),
455 kcf.p_current_center.x, kcf.p_current_center.y,
456 kcf.p_windows_size.width, kcf.p_windows_size.height,
457 kcf.p_current_scale * IF_BIG_BATCH(max.scale(i), scale),
458 kcf.p_current_angle + IF_BIG_BATCH(max.angle(i), angle))
459 .copyTo(patch_feats.scale(i));
460 DEBUG_PRINT(patch_feats.scale(i));
463 kcf.fft.forward_window(patch_feats, zf, temp);
466 if (kcf.m_use_linearkernel) {
467 kzf = zf.mul(kcf.model->model_alphaf).sum_over_channels();
469 gaussian_correlation(kzf, zf, kcf.model->model_xf, kcf.p_kernel_sigma, false, kcf);
471 kzf = kzf.mul(kcf.model->model_alphaf);
473 kcf.fft.inverse(kzf, response);
475 DEBUG_PRINTM(response);
477 /* target location is at the maximum response. we must take into
478 account the fact that, if the target doesn't move, the peak
479 will appear at the top-left corner, not at the center (this is
480 discussed in the paper). the responses wrap around cyclically. */
481 double min_val, max_val;
482 cv::Point2i min_loc, max_loc;
484 for (size_t i = 0; i < max.size(); ++i) {
485 cv::minMaxLoc(response.plane(i), &min_val, &max_val, &min_loc, &max_loc);
486 DEBUG_PRINT(max_loc);
487 double weight = max.scale(i) < 1. ? max.scale(i) : 1. / max.scale(i);
488 max[i].response = max_val * weight;
489 max[i].loc = max_loc;
492 cv::minMaxLoc(response.plane(0), &min_val, &max_val, &min_loc, &max_loc);
494 DEBUG_PRINT(max_loc);
495 DEBUG_PRINT(max_val);
497 double weight = scale < 1. ? scale : 1. / scale;
498 max.response = max_val * weight;
503 // ****************************************************************************
505 cv::Mat KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, cv::Mat *dbg_patch,
506 int cx, int cy, int size_x, int size_y, double scale, double angle) const
508 cv::Size scaled = cv::Size(floor(size_x * scale), floor(size_y * scale));
510 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, scaled.width, scaled.height, angle);
511 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, scaled.width, scaled.height, angle);
513 // resize to default size
514 if (scaled.area() > fit_size.area()) {
515 // if we downsample use INTER_AREA interpolation
516 // note: this is just a guess - we may downsample in X and upsample in Y (or vice versa)
517 cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_AREA);
519 cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_LINEAR);
522 // get hog(Histogram of Oriented Gradients) features
523 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
525 // get color rgb features (simple r,g,b channels)
526 std::vector<cv::Mat> color_feat;
527 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
528 // resize to default size
529 if (scaled.area() > (fit_size / p_cell_size).area()) {
530 // if we downsample use INTER_AREA interpolation
531 cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_AREA);
533 cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_LINEAR);
538 patch_rgb.copyTo(*dbg_patch);
540 if (m_use_color && input_rgb.channels() == 3) {
541 // use rgb color space
542 cv::Mat patch_rgb_norm;
543 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
544 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
545 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
546 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
547 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
548 cv::split(patch_rgb_norm, rgb);
549 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
552 if (m_use_cnfeat && input_rgb.channels() == 3) {
553 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
554 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
557 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
559 int size[] = {p_num_of_feats, feature_size.height, feature_size.width};
560 cv::Mat result(3, size, CV_32F);
561 for (uint i = 0; i < hog_feat.size(); ++i)
562 hog_feat[i].copyTo(cv::Mat(size[1], size[2], CV_32FC1, result.ptr(i)));
567 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
569 cv::Mat labels(dim2, dim1, CV_32FC1);
570 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
571 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
573 double sigma_s = sigma * sigma;
575 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
576 float *row_ptr = labels.ptr<float>(j);
578 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
579 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
583 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
584 MatDynMem rot_labels = circshift(labels, range_x[0], range_y[0]);
585 // sanity check, 1 at top left corner
586 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
591 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot) const
593 cv::Mat rot_patch(patch.size(), patch.type());
594 cv::Mat tmp_x_rot(patch.size(), patch.type());
596 // circular rotate x-axis
598 // move part that does not rotate over the edge
599 cv::Range orig_range(-x_rot, patch.cols);
600 cv::Range rot_range(0, patch.cols - (-x_rot));
601 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
604 orig_range = cv::Range(0, -x_rot);
605 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
606 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
607 } else if (x_rot > 0) {
608 // move part that does not rotate over the edge
609 cv::Range orig_range(0, patch.cols - x_rot);
610 cv::Range rot_range(x_rot, patch.cols);
611 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
614 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
615 rot_range = cv::Range(0, x_rot);
616 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
617 } else { // zero rotation
618 // move part that does not rotate over the edge
619 cv::Range orig_range(0, patch.cols);
620 cv::Range rot_range(0, patch.cols);
621 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
624 // circular rotate y-axis
626 // move part that does not rotate over the edge
627 cv::Range orig_range(-y_rot, patch.rows);
628 cv::Range rot_range(0, patch.rows - (-y_rot));
629 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
632 orig_range = cv::Range(0, -y_rot);
633 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
634 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
635 } else if (y_rot > 0) {
636 // move part that does not rotate over the edge
637 cv::Range orig_range(0, patch.rows - y_rot);
638 cv::Range rot_range(y_rot, patch.rows);
639 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
642 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
643 rot_range = cv::Range(0, y_rot);
644 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
645 } else { // zero rotation
646 // move part that does not rotate over the edge
647 cv::Range orig_range(0, patch.rows);
648 cv::Range rot_range(0, patch.rows);
649 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
655 // hann window actually (Power-of-cosine windows)
656 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
658 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
659 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
660 for (int i = 0; i < dim1; ++i)
661 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
662 N_inv = 1. / (static_cast<double>(dim2) - 1.);
663 for (int i = 0; i < dim2; ++i)
664 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
665 cv::Mat ret = m2 * m1;
669 // Returns sub-window of image input centered at [cx, cy] coordinates),
670 // with size [width, height]. If any pixels are outside of the image,
671 // they will replicate the values at the borders.
672 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height, double angle) const
676 cv::Size sz(width, height);
677 cv::RotatedRect rr(cv::Point2f(cx, cy), sz, angle);
678 cv::Rect bb = rr.boundingRect();
686 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
687 patch.create(height, width, input.type());
688 patch.setTo(double(0.f));
692 int top = 0, bottom = 0, left = 0, right = 0;
694 // fit to image coordinates, set border extensions;
703 if (x2 >= input.cols) {
704 right = x2 - input.cols + width % 2;
709 if (y2 >= input.rows) {
710 bottom = y2 - input.rows + height % 2;
715 if (x2 - x1 == 0 || y2 - y1 == 0)
716 patch = cv::Mat::zeros(height, width, CV_32FC1);
718 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
719 cv::BORDER_REPLICATE);
720 // imshow( "copyMakeBorder", patch);
724 cv::Point2f src_pts[4];
725 cv::RotatedRect(cv::Point2f(patch.size()) / 2.0, sz, angle).points(src_pts);
726 cv::Point2f dst_pts[3] = { cv::Point2f(0, height), cv::Point2f(0, 0), cv::Point2f(width, 0)};
727 auto rot = cv::getAffineTransform(src_pts, dst_pts);
728 cv::warpAffine(patch, patch, rot, sz);
731 assert(patch.cols == width && patch.rows == height);
736 void KCF_Tracker::GaussianCorrelation::operator()(ComplexMat &result, const ComplexMat &xf, const ComplexMat &yf,
737 double sigma, bool auto_correlation, const KCF_Tracker &kcf)
741 DEBUG_PRINT(xf_sqr_norm.num_elem);
742 xf.sqr_norm(xf_sqr_norm);
743 for (uint s = 0; s < xf.n_scales; ++s)
744 DEBUG_PRINT(xf_sqr_norm[s]);
745 if (auto_correlation) {
746 yf_sqr_norm = xf_sqr_norm;
749 yf.sqr_norm(yf_sqr_norm);
751 for (uint s = 0; s < yf.n_scales; ++s)
752 DEBUG_PRINTM(yf_sqr_norm[s]);
753 xyf = auto_correlation ? xf.sqr_mag() : xf * yf.conj(); // xf.muln(yf.conj());
756 // ifft2 and sum over 3rd dimension, we dont care about individual channels
757 ComplexMat xyf_sum = xyf.sum_over_channels();
758 DEBUG_PRINTM(xyf_sum);
759 kcf.fft.inverse(xyf_sum, ifft_res);
760 DEBUG_PRINTM(ifft_res);
762 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
763 for (uint i = 0; i < xf.n_scales; ++i) {
764 cv::Mat plane = ifft_res.plane(i);
765 DEBUG_PRINT(ifft_res.plane(i));
766 cv::exp(-1. / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * ifft_res.plane(i))
767 * numel_xf_inv, 0), plane);
771 kcf.fft.forward(ifft_res, result);
774 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
778 assert(response.dims == 2); // ensure .cols and .rows are valid
779 if (x < 0) x = response.cols + x;
780 if (y < 0) y = response.rows + y;
781 if (x >= response.cols) x = x - response.cols;
782 if (y >= response.rows) y = y - response.rows;
784 return response.at<float>(y, x);
787 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const
789 // find neighbourhood of max_loc (response is circular)
793 cv::Point2i p1(max_loc.x - 1, max_loc.y - 1), p2(max_loc.x, max_loc.y - 1), p3(max_loc.x + 1, max_loc.y - 1);
794 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
795 cv::Point2i p6(max_loc.x - 1, max_loc.y + 1), p7(max_loc.x, max_loc.y + 1), p8(max_loc.x + 1, max_loc.y + 1);
798 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
799 cv::Mat A = (cv::Mat_<float>(9, 6) <<
800 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
801 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
802 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
803 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
804 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
805 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
806 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
807 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
808 max_loc.x*max_loc.x, max_loc.x*max_loc.y, max_loc.y*max_loc.y, max_loc.x, max_loc.y, 1.f);
809 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
810 get_response_circular(p1, response),
811 get_response_circular(p2, response),
812 get_response_circular(p3, response),
813 get_response_circular(p4, response),
814 get_response_circular(p5, response),
815 get_response_circular(p6, response),
816 get_response_circular(p7, response),
817 get_response_circular(p8, response),
818 get_response_circular(max_loc, response));
821 cv::solve(A, fval, x, cv::DECOMP_SVD);
823 float a = x.at<float>(0), b = x.at<float>(1), c = x.at<float>(2), d = x.at<float>(3), e = x.at<float>(4);
825 cv::Point2f sub_peak(max_loc.x, max_loc.y);
826 if (4 * a * c - b * b > p_floating_error) {
827 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
828 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
829 if (fabs(sub_peak.x - max_loc.x) > 1 ||
830 fabs(sub_peak.y - max_loc.y) > 1)
837 double KCF_Tracker::sub_grid_scale(uint max_index)
840 const auto &vec = d->IF_BIG_BATCH(threadctxs[0].max, threadctxs);
841 uint index = vec.getScaleIdx(max_index);
842 uint angle_idx = vec.getAngleIdx(max_index);
844 if (index >= vec.size()) {
845 // interpolate from all values
846 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
847 A.create(p_scales.size(), 3, CV_32FC1);
848 fval.create(p_scales.size(), 1, CV_32FC1);
849 for (size_t i = 0; i < p_scales.size(); ++i) {
850 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
851 A.at<float>(i, 1) = float(p_scales[i]);
852 A.at<float>(i, 2) = 1;
853 fval.at<float>(i) = d->IF_BIG_BATCH(threadctxs[0].max[i].response, threadctxs(i, angle_idx).max.response);
856 // only from neighbours
857 if (index == 0 || index == p_scales.size() - 1)
858 return p_scales[index];
860 A = (cv::Mat_<float>(3, 3) <<
861 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
862 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
863 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
865 fval = (cv::Mat_<float>(3, 1) <<
866 d->threadctxs[0].max(index - 1, angle_idx).response,
867 d->threadctxs[0].max(index + 0, angle_idx).response,
868 d->threadctxs[0].max(index + 1, angle_idx).response);
870 fval = (cv::Mat_<float>(3, 1) <<
871 d->threadctxs(index - 1, angle_idx).max.response,
872 d->threadctxs(index + 0, angle_idx).max.response,
873 d->threadctxs(index + 1, angle_idx).max.response);
878 cv::solve(A, fval, x, cv::DECOMP_SVD);
879 float a = x.at<float>(0), b = x.at<float>(1);
880 double scale = p_scales[index];
882 scale = -b / (2 * a);
886 double KCF_Tracker::sub_grid_angle(uint max_index)
889 const auto &vec = d->IF_BIG_BATCH(threadctxs[0].max, threadctxs);
890 uint scale_idx = vec.getScaleIdx(max_index);
891 uint index = vec.getAngleIdx(max_index);
893 if (index >= vec.size()) {
894 // interpolate from all values
895 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
896 A.create(p_angles.size(), 3, CV_32FC1);
897 fval.create(p_angles.size(), 1, CV_32FC1);
898 for (size_t i = 0; i < p_angles.size(); ++i) {
899 A.at<float>(i, 0) = float(p_angles[i] * p_angles[i]);
900 A.at<float>(i, 1) = float(p_angles[i]);
901 A.at<float>(i, 2) = 1;
902 fval.at<float>(i) = d->IF_BIG_BATCH(threadctxs[0].max[i].response, threadctxs(scale_idx, i).max.response);
905 // only from neighbours
906 if (index == 0 || index == p_angles.size() - 1)
907 return p_angles[index];
909 A = (cv::Mat_<float>(3, 3) <<
910 p_angles[index - 1] * p_angles[index - 1], p_angles[index - 1], 1,
911 p_angles[index + 0] * p_angles[index + 0], p_angles[index + 0], 1,
912 p_angles[index + 1] * p_angles[index + 1], p_angles[index + 1], 1);
914 fval = (cv::Mat_<float>(3, 1) <<
915 d->threadctxs[0].max(scale_idx, index - 1).response,
916 d->threadctxs[0].max(scale_idx, index + 0).response,
917 d->threadctxs[0].max(scale_idx, index + 1).response);
919 fval = (cv::Mat_<float>(3, 1) <<
920 d->threadctxs(scale_idx, index - 1).max.response,
921 d->threadctxs(scale_idx, index + 0).max.response,
922 d->threadctxs(scale_idx, index + 1).max.response);
927 cv::solve(A, fval, x, cv::DECOMP_SVD);
928 float a = x.at<float>(0), b = x.at<float>(1);
929 double angle = p_angles[index];
931 angle = -b / (2 * a);