5 #include "threadctx.hpp"
13 #include "fft_cufft.h"
16 #include "fft_opencv.h"
24 DbgTracer __dbgTracer;
27 T clamp(const T& n, const T& lower, const T& upper)
29 return std::max(lower, std::min(n, upper));
33 void clamp2(T& n, const T& lower, const T& upper)
35 n = std::max(lower, std::min(n, upper));
38 #if CV_MAJOR_VERSION < 3
39 template<typename _Tp> static inline
40 cv::Size_<_Tp> operator / (const cv::Size_<_Tp>& a, _Tp b)
42 return cv::Size_<_Tp>(a.width / b, a.height / b);
45 template<typename _Tp> static inline
46 cv::Point_<_Tp> operator / (const cv::Point_<_Tp>& a, double b)
48 return cv::Point_<_Tp>(a.x / b, a.y / b);
53 class Kcf_Tracker_Private {
56 Kcf_Tracker_Private(const KCF_Tracker &kcf) : kcf(kcf) {}
58 const KCF_Tracker &kcf;
60 std::vector<ThreadCtx> threadctxs;
62 ScaleRotVector<ThreadCtx> threadctxs{kcf.p_scales, kcf.p_angles};
66 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
67 double output_sigma_factor, int cell_size)
68 : p_cell_size(cell_size), fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
69 p_lambda(lambda), p_interp_factor(interp_factor)
73 KCF_Tracker::KCF_Tracker() : fft(*new FFT()) {}
75 KCF_Tracker::~KCF_Tracker()
80 void KCF_Tracker::train(cv::Mat input_rgb, cv::Mat input_gray, double interp_factor)
84 // obtain a sub-window for training
85 get_features(input_rgb, input_gray, nullptr, p_current_center.x, p_current_center.y,
86 p_windows_size.width, p_windows_size.height,
87 p_current_scale, p_current_angle).copyTo(model->patch_feats.scale(0));
88 DEBUG_PRINT(model->patch_feats);
89 fft.forward_window(model->patch_feats, model->xf, model->temp);
90 DEBUG_PRINTM(model->xf);
91 model->model_xf = model->model_xf * (1. - interp_factor) + model->xf * interp_factor;
92 DEBUG_PRINTM(model->model_xf);
94 if (m_use_linearkernel) {
95 ComplexMat xfconj = model->xf.conj();
96 model->model_alphaf_num = xfconj.mul(model->yf);
97 model->model_alphaf_den = (model->xf * xfconj);
99 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
100 cv::Size sz(Fft::freq_size(feature_size));
101 ComplexMat kf(sz.height, sz.width, 1);
102 (*gaussian_correlation)(kf, model->model_xf, model->model_xf, p_kernel_sigma, true, *this);
104 model->model_alphaf_num = model->yf * kf;
105 model->model_alphaf_den = kf * (kf + p_lambda);
107 model->model_alphaf = model->model_alphaf_num / model->model_alphaf_den;
108 DEBUG_PRINTM(model->model_alphaf);
109 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
112 static int round_pw2_down(int x)
114 for (int i = 1; i < 32; i <<= 1)
121 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
123 __dbgTracer.debug = m_debug;
126 // check boundary, enforce min size
127 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
129 if (x2 > img.cols - 1) x2 = img.cols - 1;
131 if (y2 > img.rows - 1) y2 = img.rows - 1;
133 if (x2 - x1 < 2 * p_cell_size) {
134 double diff = (2 * p_cell_size - x2 + x1) / 2.;
135 if (x1 - diff >= 0 && x2 + diff < img.cols) {
138 } else if (x1 - 2 * diff >= 0) {
144 if (y2 - y1 < 2 * p_cell_size) {
145 double diff = (2 * p_cell_size - y2 + y1) / 2.;
146 if (y1 - diff >= 0 && y2 + diff < img.rows) {
149 } else if (y1 - 2 * diff >= 0) {
156 p_init_pose.w = x2 - x1;
157 p_init_pose.h = y2 - y1;
158 p_init_pose.cx = x1 + p_init_pose.w / 2.;
159 p_init_pose.cy = y1 + p_init_pose.h / 2.;
161 cv::Mat input_gray, input_rgb = img.clone();
162 if (img.channels() == 3) {
163 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
164 input_gray.convertTo(input_gray, CV_32FC1);
166 img.convertTo(input_gray, CV_32FC1);
168 // don't need too large image
169 if (p_init_pose.w * p_init_pose.h > 100. * 100.) {
170 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
171 p_resize_image = true;
172 p_init_pose.scale(p_downscale_factor);
173 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
174 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
177 // compute win size + fit to fhog cell size
178 p_windows_size.width = round(p_init_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
179 p_windows_size.height = round(p_init_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
181 if (fit_size_x == 0 || fit_size_y == 0) {
182 // Round down to the next highest power of 2
183 fit_size = cv::Size(round_pw2_down(p_windows_size.width),
184 round_pw2_down(p_windows_size.height));
185 } else if (fit_size_x == -1 || fit_size_y == -1) {
186 fit_size = p_windows_size;
188 fit_size = cv::Size(fit_size_x, fit_size_y);
191 feature_size = fit_size / p_cell_size;
194 for (int i = -int(p_num_scales - 1) / 2; i <= int(p_num_scales) / 2; ++i)
195 p_scales.push_back(std::pow(p_scale_step, i));
198 for (int i = -int(p_num_angles - 1) / 2; i <= int(p_num_angles) / 2; ++i)
199 p_angles.push_back(i * p_angle_step);
202 if (m_use_linearkernel) {
203 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
204 std::exit(EXIT_FAILURE);
208 model.reset(new Model(feature_size, p_num_of_feats));
209 d.reset(new Kcf_Tracker_Private(*this));
212 for (auto scale: p_scales)
213 for (auto angle : p_angles)
214 d->threadctxs.emplace_back(feature_size, p_num_of_feats, scale, angle);
216 d->threadctxs.emplace_back(feature_size, p_num_of_feats, p_scales, p_angles);
219 gaussian_correlation.reset(new GaussianCorrelation(1, p_num_of_feats, feature_size));
221 p_current_center = p_init_pose.center();
222 p_current_scale = 1.;
224 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
225 double max_size_ratio =
226 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
227 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
228 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
229 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
231 std::cout << "init: img size " << img.size() << std::endl;
232 std::cout << "init: win size " << p_windows_size;
233 if (p_windows_size != fit_size)
234 std::cout << " resized to " << fit_size;
235 std::cout << std::endl;
236 std::cout << "init: FFT size " << feature_size << std::endl;
237 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
239 p_output_sigma = std::sqrt(p_init_pose.w * p_init_pose.h * double(fit_size.area()) / p_windows_size.area())
240 * p_output_sigma_factor / p_cell_size;
242 fft.init(feature_size.width, feature_size.height, p_num_of_feats, p_num_scales * p_num_angles);
243 fft.set_window(MatDynMem(cosine_window_function(feature_size.width, feature_size.height)));
245 // window weights, i.e. labels
246 MatScales gsl(1, feature_size);
247 gaussian_shaped_labels(p_output_sigma, feature_size.width, feature_size.height).copyTo(gsl.plane(0));
248 fft.forward(gsl, model->yf);
249 DEBUG_PRINTM(model->yf);
251 // train initial model
252 train(input_rgb, input_gray, 1.0);
255 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
257 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
260 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
263 if (p_resize_image) {
264 tmp.scale(p_downscale_factor);
266 p_current_center = tmp.center();
269 BBox_c KCF_Tracker::getBBox()
272 tmp.cx = p_current_center.x;
273 tmp.cy = p_current_center.y;
274 tmp.w = p_init_pose.w * p_current_scale;
275 tmp.h = p_init_pose.h * p_current_scale;
279 tmp.scale(1 / p_downscale_factor);
284 double KCF_Tracker::getFilterResponse() const
286 return this->max_response;
289 void KCF_Tracker::resizeImgs(cv::Mat &input_rgb, cv::Mat &input_gray)
291 if (p_resize_image) {
292 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
293 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
297 double KCF_Tracker::findMaxReponse(uint &max_idx, cv::Point2d &new_location) const
300 const auto &vec = IF_BIG_BATCH(d->threadctxs[0].max, d->threadctxs);
303 auto max_it = std::max_element(vec.begin(), vec.end(),
304 [](const ThreadCtx &a, const ThreadCtx &b)
305 { return a.max.response < b.max.response; });
307 auto max_it = std::max_element(vec.begin(), vec.end(),
308 [](const ThreadCtx::Max &a, const ThreadCtx::Max &b)
309 { return a.response < b.response; });
311 assert(max_it != vec.end());
312 max = max_it->IF_BIG_BATCH(response, max.response);
314 if (m_visual_debug) {
315 const bool rgb = true;
316 int type = rgb ? d->threadctxs[0].IF_BIG_BATCH(dbg_patch[0], dbg_patch).type()
317 : d->threadctxs[0].response.type();
318 int w = true ? 100 : (rgb ? fit_size.width : feature_size.width);
319 int h = true ? 100 : (rgb ? fit_size.height : feature_size.height);
320 cv::Mat all_responses((h + 1) * p_num_scales - 1,
321 (w + 1) * p_num_angles - 1, type, cv::Scalar::all(0));
322 for (size_t i = 0; i < p_num_scales; ++i) {
323 for (size_t j = 0; j < p_num_angles; ++j) {
326 tmp = d->IF_BIG_BATCH(threadctxs[0], threadctxs(i, j)).dbg_patch IF_BIG_BATCH((i, j),);
328 tmp = d->IF_BIG_BATCH(threadctxs[0], threadctxs(i, j)).response.plane(IF_BIG_BATCH(d->threadctxs[0].max.getIdx(i, j), 0));
329 tmp = circshift(tmp, -tmp.cols/2, -tmp.rows/2);
331 cv::resize(tmp, tmp, cv::Size(w, h));
332 cv::Mat resp_roi(all_responses, cv::Rect(j * (w+1), i * (h+1), w, h));
333 tmp.copyTo(resp_roi);
336 cv::namedWindow("KCF visual debug", CV_WINDOW_AUTOSIZE);
337 cv::imshow("KCF visual debug", all_responses);
340 max_idx = std::distance(vec.begin(), max_it);
342 cv::Point2i max_response_pt = IF_BIG_BATCH(max_it->loc, max_it->max.loc);
343 cv::Mat max_response_map = IF_BIG_BATCH(d->threadctxs[0].response.plane(max_idx),
344 max_it->response.plane(0));
346 DEBUG_PRINTM(max_response_map);
347 DEBUG_PRINT(max_response_pt);
349 // sub pixel quadratic interpolation from neighbours
350 if (max_response_pt.y > max_response_map.rows / 2) // wrap around to negative half-space of vertical axis
351 max_response_pt.y = max_response_pt.y - max_response_map.rows;
352 if (max_response_pt.x > max_response_map.cols / 2) // same for horizontal axis
353 max_response_pt.x = max_response_pt.x - max_response_map.cols;
356 if (m_use_subpixel_localization) {
357 new_location = sub_pixel_peak(max_response_pt, max_response_map);
359 new_location = max_response_pt;
361 DEBUG_PRINT(new_location);
365 void KCF_Tracker::track(cv::Mat &img)
367 __dbgTracer.debug = m_debug;
370 cv::Mat input_gray, input_rgb = img.clone();
371 if (img.channels() == 3) {
372 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
373 input_gray.convertTo(input_gray, CV_32FC1);
375 img.convertTo(input_gray, CV_32FC1);
377 // don't need too large image
378 resizeImgs(input_rgb, input_gray);
381 for (auto &it : d->threadctxs)
382 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
383 it.track(*this, input_rgb, input_gray);
385 for (auto const &it : d->threadctxs)
389 NORMAL_OMP_PARALLEL_FOR
390 for (uint i = 0; i < d->threadctxs.size(); ++i)
391 d->threadctxs[i].track(*this, input_rgb, input_gray);
394 cv::Point2d new_location;
396 max_response = findMaxReponse(max_idx, new_location);
398 new_location.x *= double(p_windows_size.width) / fit_size.width;
399 new_location.y *= double(p_windows_size.height) / fit_size.height;
401 p_current_center += p_current_scale * p_cell_size * new_location;
403 clamp2(p_current_center.x, 0.0, img.cols - 1.0);
404 clamp2(p_current_center.y, 0.0, img.rows - 1.0);
406 // sub grid scale interpolation
407 if (m_use_subgrid_scale) {
408 p_current_scale *= sub_grid_scale(max_idx);
410 p_current_scale *= d->IF_BIG_BATCH(threadctxs[0].max, threadctxs).scale(max_idx);
413 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
415 p_current_angle += d->IF_BIG_BATCH(threadctxs[0].max, threadctxs).angle(max_idx);
417 // train at newly estimated target position
418 train(input_rgb, input_gray, p_interp_factor);
421 void ThreadCtx::track(const KCF_Tracker &kcf, cv::Mat &input_rgb, cv::Mat &input_gray)
425 BIG_BATCH_OMP_PARALLEL_FOR
426 for (uint i = 0; i < IF_BIG_BATCH(max.size(), 1); ++i)
428 kcf.get_features(input_rgb, input_gray, &dbg_patch IF_BIG_BATCH([i],),
429 kcf.p_current_center.x, kcf.p_current_center.y,
430 kcf.p_windows_size.width, kcf.p_windows_size.height,
431 kcf.p_current_scale * IF_BIG_BATCH(max.scale(i), scale),
432 kcf.p_current_angle + IF_BIG_BATCH(max.angle(i), angle))
433 .copyTo(patch_feats.scale(i));
434 DEBUG_PRINT(patch_feats.scale(i));
437 kcf.fft.forward_window(patch_feats, zf, temp);
440 if (kcf.m_use_linearkernel) {
441 kzf = zf.mul(kcf.model->model_alphaf).sum_over_channels();
443 gaussian_correlation(kzf, zf, kcf.model->model_xf, kcf.p_kernel_sigma, false, kcf);
445 kzf = kzf.mul(kcf.model->model_alphaf);
447 kcf.fft.inverse(kzf, response);
449 DEBUG_PRINTM(response);
451 /* target location is at the maximum response. we must take into
452 account the fact that, if the target doesn't move, the peak
453 will appear at the top-left corner, not at the center (this is
454 discussed in the paper). the responses wrap around cyclically. */
455 double min_val, max_val;
456 cv::Point2i min_loc, max_loc;
458 for (size_t i = 0; i < max.size(); ++i) {
459 cv::minMaxLoc(response.plane(i), &min_val, &max_val, &min_loc, &max_loc);
460 DEBUG_PRINT(max_loc);
461 double weight = kcf.p_scales[i] < 1. ? kcf.p_scales[i] : 1. / kcf.p_scales[i];
462 max[i].response = max_val * weight;
463 max[i].loc = max_loc;
466 cv::minMaxLoc(response.plane(0), &min_val, &max_val, &min_loc, &max_loc);
468 DEBUG_PRINT(max_loc);
469 DEBUG_PRINT(max_val);
471 double weight = scale < 1. ? scale : 1. / scale;
472 max.response = max_val * weight;
477 // ****************************************************************************
479 cv::Mat KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, cv::Mat *dbg_patch,
480 int cx, int cy, int size_x, int size_y, double scale, double angle) const
482 cv::Size scaled = cv::Size(floor(size_x * scale), floor(size_y * scale));
484 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, scaled.width, scaled.height, angle);
485 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, scaled.width, scaled.height, angle);
488 patch_rgb.copyTo(*dbg_patch);
490 // resize to default size
491 if (scaled.area() > fit_size.area()) {
492 // if we downsample use INTER_AREA interpolation
493 // note: this is just a guess - we may downsample in X and upsample in Y (or vice versa)
494 cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_AREA);
496 cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_LINEAR);
499 // get hog(Histogram of Oriented Gradients) features
500 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
502 // get color rgb features (simple r,g,b channels)
503 std::vector<cv::Mat> color_feat;
504 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
505 // resize to default size
506 if (scaled.area() > (fit_size / p_cell_size).area()) {
507 // if we downsample use INTER_AREA interpolation
508 cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_AREA);
510 cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_LINEAR);
514 if (m_use_color && input_rgb.channels() == 3) {
515 // use rgb color space
516 cv::Mat patch_rgb_norm;
517 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
518 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
519 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
520 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
521 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
522 cv::split(patch_rgb_norm, rgb);
523 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
526 if (m_use_cnfeat && input_rgb.channels() == 3) {
527 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
528 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
531 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
533 int size[] = {p_num_of_feats, feature_size.height, feature_size.width};
534 cv::Mat result(3, size, CV_32F);
535 for (uint i = 0; i < hog_feat.size(); ++i)
536 hog_feat[i].copyTo(cv::Mat(size[1], size[2], CV_32FC1, result.ptr(i)));
541 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
543 cv::Mat labels(dim2, dim1, CV_32FC1);
544 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
545 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
547 double sigma_s = sigma * sigma;
549 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
550 float *row_ptr = labels.ptr<float>(j);
552 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
553 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
557 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
558 MatDynMem rot_labels = circshift(labels, range_x[0], range_y[0]);
559 // sanity check, 1 at top left corner
560 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
565 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot) const
567 cv::Mat rot_patch(patch.size(), CV_32FC1);
568 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
570 // circular rotate x-axis
572 // move part that does not rotate over the edge
573 cv::Range orig_range(-x_rot, patch.cols);
574 cv::Range rot_range(0, patch.cols - (-x_rot));
575 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
578 orig_range = cv::Range(0, -x_rot);
579 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
580 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
581 } else if (x_rot > 0) {
582 // move part that does not rotate over the edge
583 cv::Range orig_range(0, patch.cols - x_rot);
584 cv::Range rot_range(x_rot, patch.cols);
585 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
588 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
589 rot_range = cv::Range(0, x_rot);
590 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
591 } else { // zero rotation
592 // move part that does not rotate over the edge
593 cv::Range orig_range(0, patch.cols);
594 cv::Range rot_range(0, patch.cols);
595 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
598 // circular rotate y-axis
600 // move part that does not rotate over the edge
601 cv::Range orig_range(-y_rot, patch.rows);
602 cv::Range rot_range(0, patch.rows - (-y_rot));
603 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
606 orig_range = cv::Range(0, -y_rot);
607 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
608 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
609 } else if (y_rot > 0) {
610 // move part that does not rotate over the edge
611 cv::Range orig_range(0, patch.rows - y_rot);
612 cv::Range rot_range(y_rot, patch.rows);
613 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
616 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
617 rot_range = cv::Range(0, y_rot);
618 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
619 } else { // zero rotation
620 // move part that does not rotate over the edge
621 cv::Range orig_range(0, patch.rows);
622 cv::Range rot_range(0, patch.rows);
623 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
629 // hann window actually (Power-of-cosine windows)
630 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
632 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
633 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
634 for (int i = 0; i < dim1; ++i)
635 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
636 N_inv = 1. / (static_cast<double>(dim2) - 1.);
637 for (int i = 0; i < dim2; ++i)
638 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
639 cv::Mat ret = m2 * m1;
643 // Returns sub-window of image input centered at [cx, cy] coordinates),
644 // with size [width, height]. If any pixels are outside of the image,
645 // they will replicate the values at the borders.
646 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height, double angle) const
651 int x1 = cx - width / 2;
652 int y1 = cy - height / 2;
653 int x2 = cx + width / 2;
654 int y2 = cy + height / 2;
657 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
658 patch.create(height, width, input.type());
659 patch.setTo(double(0.f));
663 int top = 0, bottom = 0, left = 0, right = 0;
665 // fit to image coordinates, set border extensions;
674 if (x2 >= input.cols) {
675 right = x2 - input.cols + width % 2;
680 if (y2 >= input.rows) {
681 bottom = y2 - input.rows + height % 2;
686 if (x2 - x1 == 0 || y2 - y1 == 0)
687 patch = cv::Mat::zeros(height, width, CV_32FC1);
689 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
690 cv::BORDER_REPLICATE);
691 // imshow( "copyMakeBorder", patch);
696 assert(patch.cols == width && patch.rows == height);
701 void KCF_Tracker::GaussianCorrelation::operator()(ComplexMat &result, const ComplexMat &xf, const ComplexMat &yf,
702 double sigma, bool auto_correlation, const KCF_Tracker &kcf)
706 DEBUG_PRINT(xf_sqr_norm.num_elem);
707 xf.sqr_norm(xf_sqr_norm);
708 for (uint s = 0; s < xf.n_scales; ++s)
709 DEBUG_PRINT(xf_sqr_norm[s]);
710 if (auto_correlation) {
711 yf_sqr_norm = xf_sqr_norm;
714 yf.sqr_norm(yf_sqr_norm);
716 for (uint s = 0; s < yf.n_scales; ++s)
717 DEBUG_PRINTM(yf_sqr_norm[s]);
718 xyf = auto_correlation ? xf.sqr_mag() : xf * yf.conj(); // xf.muln(yf.conj());
721 // ifft2 and sum over 3rd dimension, we dont care about individual channels
722 ComplexMat xyf_sum = xyf.sum_over_channels();
723 DEBUG_PRINTM(xyf_sum);
724 kcf.fft.inverse(xyf_sum, ifft_res);
725 DEBUG_PRINTM(ifft_res);
727 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
728 for (uint i = 0; i < xf.n_scales; ++i) {
729 cv::Mat plane = ifft_res.plane(i);
730 DEBUG_PRINT(ifft_res.plane(i));
731 cv::exp(-1. / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * ifft_res.plane(i))
732 * numel_xf_inv, 0), plane);
736 kcf.fft.forward(ifft_res, result);
739 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
743 assert(response.dims == 2); // ensure .cols and .rows are valid
744 if (x < 0) x = response.cols + x;
745 if (y < 0) y = response.rows + y;
746 if (x >= response.cols) x = x - response.cols;
747 if (y >= response.rows) y = y - response.rows;
749 return response.at<float>(y, x);
752 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const
754 // find neighbourhood of max_loc (response is circular)
758 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);
759 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
760 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);
763 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
764 cv::Mat A = (cv::Mat_<float>(9, 6) <<
765 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
766 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
767 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
768 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
769 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
770 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
771 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
772 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
773 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);
774 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
775 get_response_circular(p1, response),
776 get_response_circular(p2, response),
777 get_response_circular(p3, response),
778 get_response_circular(p4, response),
779 get_response_circular(p5, response),
780 get_response_circular(p6, response),
781 get_response_circular(p7, response),
782 get_response_circular(p8, response),
783 get_response_circular(max_loc, response));
786 cv::solve(A, fval, x, cv::DECOMP_SVD);
788 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);
790 cv::Point2f sub_peak(max_loc.x, max_loc.y);
791 if (b > 0 || b < 0) {
792 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
793 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
799 double KCF_Tracker::sub_grid_scale(uint max_index)
802 const auto &vec = d->IF_BIG_BATCH(threadctxs[0].max, threadctxs);
803 uint index = vec.getScaleIdx(max_index);
804 uint angle_idx = vec.getAngleIdx(index);
806 if (index >= vec.size()) {
807 // interpolate from all values
808 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
809 A.create(p_scales.size(), 3, CV_32FC1);
810 fval.create(p_scales.size(), 1, CV_32FC1);
811 for (size_t i = 0; i < p_scales.size(); ++i) {
812 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
813 A.at<float>(i, 1) = float(p_scales[i]);
814 A.at<float>(i, 2) = 1;
815 fval.at<float>(i) = d->IF_BIG_BATCH(threadctxs[0].max[i].response, threadctxs(i, angle_idx).max.response);
818 // only from neighbours
819 if (index == 0 || index == p_scales.size() - 1)
820 return p_scales[index];
822 A = (cv::Mat_<float>(3, 3) <<
823 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
824 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
825 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
827 fval = (cv::Mat_<float>(3, 1) <<
828 d->threadctxs[0].max(index - 1, angle_idx).response,
829 d->threadctxs[0].max(index + 0, angle_idx).response,
830 d->threadctxs[0].max(index + 1, angle_idx).response);
832 fval = (cv::Mat_<float>(3, 1) <<
833 d->threadctxs(index - 1, angle_idx).max.response,
834 d->threadctxs(index + 0, angle_idx).max.response,
835 d->threadctxs(index + 1, angle_idx).max.response);
840 cv::solve(A, fval, x, cv::DECOMP_SVD);
841 float a = x.at<float>(0), b = x.at<float>(1);
842 double scale = p_scales[index];
844 scale = -b / (2 * a);