5 #include "threadctx.hpp"
12 #include "fft_cufft.h"
15 #include "fft_opencv.h"
23 DbgTracer __dbgTracer;
26 T clamp(const T& n, const T& lower, const T& upper)
28 return std::max(lower, std::min(n, upper));
32 void clamp2(T& n, const T& lower, const T& upper)
34 n = std::max(lower, std::min(n, upper));
37 #if CV_VERSION_EPOCH < 3
38 template<typename _Tp> static inline
39 cv::Size_<_Tp> operator / (const cv::Size_<_Tp>& a, _Tp b)
41 return cv::Size_<_Tp>(a.width / b, a.height / b);
45 class Kcf_Tracker_Private {
47 std::vector<ThreadCtx> threadctxs;
50 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
51 double output_sigma_factor, int cell_size)
52 : p_cell_size(cell_size), fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
53 p_lambda(lambda), p_interp_factor(interp_factor), d(*new Kcf_Tracker_Private)
57 KCF_Tracker::KCF_Tracker() : fft(*new FFT()), d(*new Kcf_Tracker_Private) {}
59 KCF_Tracker::~KCF_Tracker()
65 void KCF_Tracker::train(cv::Mat input_rgb, cv::Mat input_gray, double interp_factor)
69 // obtain a sub-window for training
70 // TODO: Move Mats outside from here
71 MatScaleFeats patch_feats(1, p_num_of_feats, feature_size);
72 DEBUG_PRINT(patch_feats);
73 MatScaleFeats temp(1, p_num_of_feats, feature_size);
74 get_features(input_rgb, input_gray, p_current_center.x, p_current_center.y,
75 p_windows_size.width, p_windows_size.height,
76 p_current_scale).copyTo(patch_feats.scale(0));
77 DEBUG_PRINT(patch_feats);
78 fft.forward_window(patch_feats, p_xf, temp);
80 p_model_xf = p_model_xf * (1. - interp_factor) + p_xf * interp_factor;
81 DEBUG_PRINTM(p_model_xf);
83 ComplexMat alphaf_num, alphaf_den;
85 if (m_use_linearkernel) {
86 ComplexMat xfconj = p_xf.conj();
87 alphaf_num = xfconj.mul(p_yf);
88 alphaf_den = (p_xf * xfconj);
90 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
91 cv::Size sz(Fft::freq_size(feature_size));
92 ComplexMat kf(sz.height, sz.width, 1);
93 (*gaussian_correlation)(kf, p_model_xf, p_model_xf, p_kernel_sigma, true, *this);
95 p_model_alphaf_num = p_yf * kf;
96 p_model_alphaf_den = kf * (kf + p_lambda);
98 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
99 DEBUG_PRINTM(p_model_alphaf);
100 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
103 static int round_pw2_down(int x)
105 for (int i = 1; i < 32; i <<= 1)
112 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
114 __dbgTracer.debug = m_debug;
117 // check boundary, enforce min size
118 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
120 if (x2 > img.cols - 1) x2 = img.cols - 1;
122 if (y2 > img.rows - 1) y2 = img.rows - 1;
124 if (x2 - x1 < 2 * p_cell_size) {
125 double diff = (2 * p_cell_size - x2 + x1) / 2.;
126 if (x1 - diff >= 0 && x2 + diff < img.cols) {
129 } else if (x1 - 2 * diff >= 0) {
135 if (y2 - y1 < 2 * p_cell_size) {
136 double diff = (2 * p_cell_size - y2 + y1) / 2.;
137 if (y1 - diff >= 0 && y2 + diff < img.rows) {
140 } else if (y1 - 2 * diff >= 0) {
147 p_init_pose.w = x2 - x1;
148 p_init_pose.h = y2 - y1;
149 p_init_pose.cx = x1 + p_init_pose.w / 2.;
150 p_init_pose.cy = y1 + p_init_pose.h / 2.;
152 cv::Mat input_gray, input_rgb = img.clone();
153 if (img.channels() == 3) {
154 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
155 input_gray.convertTo(input_gray, CV_32FC1);
157 img.convertTo(input_gray, CV_32FC1);
159 // don't need too large image
160 if (p_init_pose.w * p_init_pose.h > 100. * 100.) {
161 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
162 p_resize_image = true;
163 p_init_pose.scale(p_downscale_factor);
164 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
165 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
168 // compute win size + fit to fhog cell size
169 p_windows_size.width = round(p_init_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
170 p_windows_size.height = round(p_init_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
172 if (fit_size_x == 0 || fit_size_y == 0) {
173 // Round down to the next highest power of 2
174 fit_size = cv::Size(round_pw2_down(p_windows_size.width),
175 round_pw2_down(p_windows_size.height));
176 } else if (fit_size_x == -1 || fit_size_y == -1) {
177 fit_size = p_windows_size;
179 fit_size = cv::Size(fit_size_x, fit_size_y);
182 feature_size = fit_size / p_cell_size;
185 for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
186 p_scales.push_back(std::pow(p_scale_step, i));
189 if (Fft::freq_size(feature_size).area() > 1024) {
190 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
191 "the window dimensions so its size is less or equal to "
192 << 1024 * p_cell_size * p_cell_size * 2 + 1
193 << " pixels. Currently the size of the window is: " << fit_size
194 << " which is " << fit_size.area() << " pixels. " << std::endl;
195 std::exit(EXIT_FAILURE);
198 if (m_use_linearkernel) {
199 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
200 std::exit(EXIT_FAILURE);
204 #if defined(CUFFT) || defined(FFTW)
205 uint width = feature_size.width / 2 + 1;
207 uint width = feature_size.width;
209 p_model_xf.create(feature_size.height, width, p_num_of_feats);
210 p_yf.create(feature_size.height, width, 1);
211 p_xf.create(feature_size.height, width, p_num_of_feats);
214 for (auto scale: p_scales)
215 d.threadctxs.emplace_back(feature_size, p_num_of_feats, scale);
217 d.threadctxs.emplace_back(feature_size, p_num_of_feats, p_num_scales);
220 gaussian_correlation.reset(new GaussianCorrelation(1, feature_size));
222 p_current_center = p_init_pose.center();
223 p_current_scale = 1.;
225 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
226 double max_size_ratio =
227 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
228 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
229 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
230 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
232 std::cout << "init: img size " << img.size() << std::endl;
233 std::cout << "init: win size " << p_windows_size;
234 if (p_windows_size != fit_size)
235 std::cout << " resized to " << fit_size;
236 std::cout << std::endl;
237 std::cout << "init: FFT size " << feature_size << std::endl;
238 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
240 p_output_sigma = std::sqrt(p_init_pose.w * p_init_pose.h * double(fit_size.area()) / p_windows_size.area())
241 * p_output_sigma_factor / p_cell_size;
243 fft.init(feature_size.width, feature_size.height, p_num_of_feats, p_num_scales);
244 fft.set_window(MatDynMem(cosine_window_function(feature_size.width, feature_size.height)));
246 // window weights, i.e. labels
247 MatScales gsl(1, feature_size);
248 gaussian_shaped_labels(p_output_sigma, feature_size.width, feature_size.height).copyTo(gsl.plane(0));
249 fft.forward(gsl, p_yf);
252 // train initial model
253 train(input_rgb, input_gray, 1.0);
256 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
258 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
261 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
264 if (p_resize_image) {
265 tmp.scale(p_downscale_factor);
267 p_current_center = tmp.center();
270 BBox_c KCF_Tracker::getBBox()
273 tmp.cx = p_current_center.x;
274 tmp.cy = p_current_center.y;
275 tmp.w = p_init_pose.w * p_current_scale;
276 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
301 for (uint j = 0; j < d.threadctxs.size(); ++j) {
302 if (d.threadctxs[j].max.response > max) {
303 max = d.threadctxs[j].max.response;
308 for (uint j = 0; j < p_scales.size(); ++j) {
309 if (d.threadctxs[0].max[j].response > max) {
310 max = d.threadctxs[0].max[j].response;
315 cv::Point2i &max_response_pt = IF_BIG_BATCH(d.threadctxs[0].max[max_idx].loc, d.threadctxs[max_idx].max.loc);
316 cv::Mat max_response_map = IF_BIG_BATCH(d.threadctxs[0].response.plane(max_idx), d.threadctxs[max_idx].response.plane(0));
318 DEBUG_PRINTM(max_response_map);
319 DEBUG_PRINT(max_response_pt);
321 // sub pixel quadratic interpolation from neighbours
322 if (max_response_pt.y > max_response_map.rows / 2) // wrap around to negative half-space of vertical axis
323 max_response_pt.y = max_response_pt.y - max_response_map.rows;
324 if (max_response_pt.x > max_response_map.cols / 2) // same for horizontal axis
325 max_response_pt.x = max_response_pt.x - max_response_map.cols;
328 if (m_use_subpixel_localization) {
329 new_location = sub_pixel_peak(max_response_pt, max_response_map);
331 new_location = max_response_pt;
333 DEBUG_PRINT(new_location);
337 void KCF_Tracker::track(cv::Mat &img)
339 __dbgTracer.debug = m_debug;
342 cv::Mat input_gray, input_rgb = img.clone();
343 if (img.channels() == 3) {
344 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
345 input_gray.convertTo(input_gray, CV_32FC1);
347 img.convertTo(input_gray, CV_32FC1);
349 // don't need too large image
350 resizeImgs(input_rgb, input_gray);
353 for (auto &it : d.threadctxs)
354 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
355 it.track(*this, input_rgb, input_gray);
357 for (auto const &it : d.threadctxs)
361 NORMAL_OMP_PARALLEL_FOR
362 for (uint i = 0; i < d.threadctxs.size(); ++i)
363 d.threadctxs[i].track(*this, input_rgb, input_gray);
366 cv::Point2d new_location;
368 max_response = findMaxReponse(max_idx, new_location);
370 new_location.x *= double(p_windows_size.width) / fit_size.width;
371 new_location.y *= double(p_windows_size.height) / fit_size.height;
373 p_current_center += p_current_scale * p_cell_size * new_location;
375 clamp2(p_current_center.x, 0.0, img.cols - 1.0);
376 clamp2(p_current_center.y, 0.0, img.rows - 1.0);
378 // sub grid scale interpolation
379 if (m_use_subgrid_scale) {
380 p_current_scale *= sub_grid_scale(max_idx);
382 p_current_scale *= p_scales[max_idx];
385 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
387 // train at newly estimated target position
388 train(input_rgb, input_gray, p_interp_factor);
391 void ThreadCtx::track(const KCF_Tracker &kcf, cv::Mat &input_rgb, cv::Mat &input_gray)
395 BIG_BATCH_OMP_PARALLEL_FOR
396 for (uint i = 0; i < IF_BIG_BATCH(kcf.p_num_scales, 1); ++i)
398 kcf.get_features(input_rgb, input_gray, kcf.p_current_center.x, kcf.p_current_center.y,
399 kcf.p_windows_size.width, kcf.p_windows_size.height,
400 kcf.p_current_scale * IF_BIG_BATCH(kcf.p_scales[i], scale))
401 .copyTo(patch_feats.scale(i));
402 DEBUG_PRINT(patch_feats.scale(i));
405 kcf.fft.forward_window(patch_feats, zf, temp);
408 if (kcf.m_use_linearkernel) {
409 kzf = zf.mul(kcf.p_model_alphaf).sum_over_channels();
411 gaussian_correlation(kzf, zf, kcf.p_model_xf, kcf.p_kernel_sigma, false, kcf);
413 kzf = kzf.mul(kcf.p_model_alphaf);
415 kcf.fft.inverse(kzf, response);
417 DEBUG_PRINTM(response);
419 /* target location is at the maximum response. we must take into
420 account the fact that, if the target doesn't move, the peak
421 will appear at the top-left corner, not at the center (this is
422 discussed in the paper). the responses wrap around cyclically. */
423 double min_val, max_val;
424 cv::Point2i min_loc, max_loc;
426 for (size_t i = 0; i < kcf.p_scales.size(); ++i) {
427 cv::minMaxLoc(response.plane(i), &min_val, &max_val, &min_loc, &max_loc);
428 DEBUG_PRINT(max_loc);
429 double weight = kcf.p_scales[i] < 1. ? kcf.p_scales[i] : 1. / kcf.p_scales[i];
430 max[i].response = max_val * weight;
431 max[i].loc = max_loc;
434 cv::minMaxLoc(response.plane(0), &min_val, &max_val, &min_loc, &max_loc);
436 DEBUG_PRINT(max_loc);
437 DEBUG_PRINT(max_val);
439 double weight = scale < 1. ? scale : 1. / scale;
440 max.response = max_val * weight;
445 // ****************************************************************************
447 cv::Mat KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy,
448 int size_x, int size_y, double scale) const
450 cv::Size scaled = cv::Size(floor(size_x * scale), floor(size_y * scale));
452 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, scaled.width, scaled.height);
453 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, scaled.width, scaled.height);
455 // resize to default size
456 if (scaled.area() > fit_size.area()) {
457 // if we downsample use INTER_AREA interpolation
458 // note: this is just a guess - we may downsample in X and upsample in Y (or vice versa)
459 cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_AREA);
461 cv::resize(patch_gray, patch_gray, fit_size, 0., 0., cv::INTER_LINEAR);
464 // get hog(Histogram of Oriented Gradients) features
465 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
467 // get color rgb features (simple r,g,b channels)
468 std::vector<cv::Mat> color_feat;
469 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
470 // resize to default size
471 if (scaled.area() > (fit_size / p_cell_size).area()) {
472 // if we downsample use INTER_AREA interpolation
473 cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_AREA);
475 cv::resize(patch_rgb, patch_rgb, fit_size / p_cell_size, 0., 0., cv::INTER_LINEAR);
479 if (m_use_color && input_rgb.channels() == 3) {
480 // use rgb color space
481 cv::Mat patch_rgb_norm;
482 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
483 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
484 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
485 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
486 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
487 cv::split(patch_rgb_norm, rgb);
488 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
491 if (m_use_cnfeat && input_rgb.channels() == 3) {
492 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
493 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
496 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
498 int size[] = {p_num_of_feats, feature_size.height, feature_size.width};
499 cv::Mat result(3, size, CV_32F);
500 for (uint i = 0; i < hog_feat.size(); ++i)
501 hog_feat[i].copyTo(cv::Mat(size[1], size[2], CV_32FC1, result.ptr(i)));
506 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
508 cv::Mat labels(dim2, dim1, CV_32FC1);
509 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
510 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
512 double sigma_s = sigma * sigma;
514 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
515 float *row_ptr = labels.ptr<float>(j);
517 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
518 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
522 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
523 MatDynMem rot_labels = circshift(labels, range_x[0], range_y[0]);
524 // sanity check, 1 at top left corner
525 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
530 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
532 cv::Mat rot_patch(patch.size(), CV_32FC1);
533 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
535 // circular rotate x-axis
537 // move part that does not rotate over the edge
538 cv::Range orig_range(-x_rot, patch.cols);
539 cv::Range rot_range(0, patch.cols - (-x_rot));
540 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
543 orig_range = cv::Range(0, -x_rot);
544 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
545 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
546 } else if (x_rot > 0) {
547 // move part that does not rotate over the edge
548 cv::Range orig_range(0, patch.cols - x_rot);
549 cv::Range rot_range(x_rot, patch.cols);
550 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
553 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
554 rot_range = cv::Range(0, x_rot);
555 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
556 } else { // zero rotation
557 // move part that does not rotate over the edge
558 cv::Range orig_range(0, patch.cols);
559 cv::Range rot_range(0, patch.cols);
560 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
563 // circular rotate y-axis
565 // move part that does not rotate over the edge
566 cv::Range orig_range(-y_rot, patch.rows);
567 cv::Range rot_range(0, patch.rows - (-y_rot));
568 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
571 orig_range = cv::Range(0, -y_rot);
572 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
573 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
574 } else if (y_rot > 0) {
575 // move part that does not rotate over the edge
576 cv::Range orig_range(0, patch.rows - y_rot);
577 cv::Range rot_range(y_rot, patch.rows);
578 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
581 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
582 rot_range = cv::Range(0, y_rot);
583 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
584 } else { // zero rotation
585 // move part that does not rotate over the edge
586 cv::Range orig_range(0, patch.rows);
587 cv::Range rot_range(0, patch.rows);
588 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
594 // hann window actually (Power-of-cosine windows)
595 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
597 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
598 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
599 for (int i = 0; i < dim1; ++i)
600 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
601 N_inv = 1. / (static_cast<double>(dim2) - 1.);
602 for (int i = 0; i < dim2; ++i)
603 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
604 cv::Mat ret = m2 * m1;
608 // Returns sub-window of image input centered at [cx, cy] coordinates),
609 // with size [width, height]. If any pixels are outside of the image,
610 // they will replicate the values at the borders.
611 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height) const
615 int x1 = cx - width / 2;
616 int y1 = cy - height / 2;
617 int x2 = cx + width / 2;
618 int y2 = cy + height / 2;
621 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
622 patch.create(height, width, input.type());
623 patch.setTo(double(0.f));
627 int top = 0, bottom = 0, left = 0, right = 0;
629 // fit to image coordinates, set border extensions;
638 if (x2 >= input.cols) {
639 right = x2 - input.cols + width % 2;
644 if (y2 >= input.rows) {
645 bottom = y2 - input.rows + height % 2;
650 if (x2 - x1 == 0 || y2 - y1 == 0)
651 patch = cv::Mat::zeros(height, width, CV_32FC1);
653 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
654 cv::BORDER_REPLICATE);
655 // imshow( "copyMakeBorder", patch);
660 assert(patch.cols == width && patch.rows == height);
665 void KCF_Tracker::GaussianCorrelation::operator()(ComplexMat &result, const ComplexMat &xf, const ComplexMat &yf,
666 double sigma, bool auto_correlation, const KCF_Tracker &kcf)
669 xf.sqr_norm(xf_sqr_norm);
670 if (auto_correlation) {
671 yf_sqr_norm = xf_sqr_norm;
673 yf.sqr_norm(yf_sqr_norm);
675 xyf = auto_correlation ? xf.sqr_mag() : xf * yf.conj(); // xf.muln(yf.conj());
678 // ifft2 and sum over 3rd dimension, we dont care about individual channels
679 ComplexMat xyf_sum = xyf.sum_over_channels();
680 DEBUG_PRINTM(xyf_sum);
681 kcf.fft.inverse(xyf_sum, ifft_res);
682 DEBUG_PRINTM(ifft_res);
685 cuda_gaussian_correlation(ifft_res.deviceMem(), k.deviceMem(), xf_sqr_norm.deviceMem(),
686 auto_correlation ? xf_sqr_norm.deviceMem() : yf_sqr_norm.deviceMem(), sigma,
687 xf.n_channels, xf.n_scales, kcf.feature_size.height, kcf.feature_size.width);
690 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
691 for (uint i = 0; i < xf.n_scales; ++i) {
692 cv::Mat plane = ifft_res.plane(i);
693 DEBUG_PRINT(ifft_res.plane(i));
694 cv::exp(-1. / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * ifft_res.plane(i))
695 * numel_xf_inv, 0), plane);
699 kcf.fft.forward(ifft_res, result);
702 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
706 assert(response.dims == 2); // ensure .cols and .rows are valid
707 if (x < 0) x = response.cols + x;
708 if (y < 0) y = response.rows + y;
709 if (x >= response.cols) x = x - response.cols;
710 if (y >= response.rows) y = y - response.rows;
712 return response.at<float>(y, x);
715 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const
717 // find neighbourhood of max_loc (response is circular)
721 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);
722 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
723 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);
726 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
727 cv::Mat A = (cv::Mat_<float>(9, 6) <<
728 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
729 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
730 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
731 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
732 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
733 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
734 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
735 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
736 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);
737 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
738 get_response_circular(p1, response),
739 get_response_circular(p2, response),
740 get_response_circular(p3, response),
741 get_response_circular(p4, response),
742 get_response_circular(p5, response),
743 get_response_circular(p6, response),
744 get_response_circular(p7, response),
745 get_response_circular(p8, response),
746 get_response_circular(max_loc, response));
749 cv::solve(A, fval, x, cv::DECOMP_SVD);
751 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);
753 cv::Point2f sub_peak(max_loc.x, max_loc.y);
754 if (b > 0 || b < 0) {
755 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
756 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
762 double KCF_Tracker::sub_grid_scale(uint index)
765 if (index >= p_scales.size()) {
766 // interpolate from all values
767 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
768 A.create(p_scales.size(), 3, CV_32FC1);
769 fval.create(p_scales.size(), 1, CV_32FC1);
770 for (size_t i = 0; i < p_scales.size(); ++i) {
771 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
772 A.at<float>(i, 1) = float(p_scales[i]);
773 A.at<float>(i, 2) = 1;
774 fval.at<float>(i) = d.threadctxs.back().IF_BIG_BATCH(max[i].response, max.response);
777 // only from neighbours
778 if (index == 0 || index == p_scales.size() - 1)
779 return p_scales[index];
781 A = (cv::Mat_<float>(3, 3) <<
782 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
783 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
784 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
786 fval = (cv::Mat_<float>(3, 1) <<
787 d.threadctxs.back().max[index - 1].response,
788 d.threadctxs.back().max[index + 0].response,
789 d.threadctxs.back().max[index + 1].response);
791 fval = (cv::Mat_<float>(3, 1) <<
792 d.threadctxs[index - 1].max.response,
793 d.threadctxs[index + 0].max.response,
794 d.threadctxs[index + 1].max.response);
799 cv::solve(A, fval, x, cv::DECOMP_SVD);
800 float a = x.at<float>(0), b = x.at<float>(1);
801 double scale = p_scales[index];
803 scale = -b / (2 * a);