5 #include "threadctx.hpp"
11 #include "fft_cufft.h"
14 #include "fft_opencv.h"
22 static bool kcf_debug = false;
24 #define DEBUG_PRINT(obj) \
26 std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl; \
28 #define DEBUG_PRINTM(obj) \
30 std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl \
31 /*<< (obj)*/ << std::endl; \
35 T clamp(const T& n, const T& lower, const T& upper)
37 return std::max(lower, std::min(n, upper));
41 void clamp2(T& n, const T& lower, const T& upper)
43 n = std::max(lower, std::min(n, upper));
46 class Kcf_Tracker_Private {
48 std::vector<ThreadCtx> threadctxs;
51 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor,
52 double output_sigma_factor, int cell_size)
53 : fft(*new FFT()), p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
54 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size), d(*new Kcf_Tracker_Private)
58 KCF_Tracker::KCF_Tracker() : fft(*new FFT()), d(*new Kcf_Tracker_Private) {}
60 KCF_Tracker::~KCF_Tracker()
66 void KCF_Tracker::train(cv::Mat input_gray, cv::Mat input_rgb, double interp_factor)
68 // obtain a sub-window for training
69 // TODO: Move Mats outside from here
70 MatScaleFeats patch_feats(1, p_num_of_feats, p_roi);
71 MatScaleFeats temp(1, p_num_of_feats, p_roi);
72 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy,
73 p_windows_size.width, p_windows_size.height,
74 p_current_scale).copyTo(patch_feats.features(0));
76 fft.forward_window(patch_feats, p_xf, temp);
77 p_model_xf = p_model_xf * (1. - interp_factor) + p_xf * interp_factor;
78 DEBUG_PRINTM(p_model_xf);
80 ComplexMat alphaf_num, alphaf_den;
82 if (m_use_linearkernel) {
83 ComplexMat xfconj = p_xf.conj();
84 alphaf_num = xfconj.mul(p_yf);
85 alphaf_den = (p_xf * xfconj);
87 // Kernel Ridge Regression, calculate alphas (in Fourier domain)
88 const uint num_scales = BIG_BATCH_MODE ? p_num_scales : 1;
89 cv::Size sz(Fft::freq_size(p_roi));
90 ComplexMat kf(sz.height, sz.width, num_scales);
91 (*gaussian_correlation)(*this, kf, p_model_xf, p_model_xf, p_kernel_sigma, true);
93 p_model_alphaf_num = p_yf * kf;
94 DEBUG_PRINTM(p_model_alphaf_num);
95 p_model_alphaf_den = kf * (kf + p_lambda);
96 DEBUG_PRINTM(p_model_alphaf_den);
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 void KCF_Tracker::init(cv::Mat &img, const cv::Rect &bbox, int fit_size_x, int fit_size_y)
106 // check boundary, enforce min size
107 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
109 if (x2 > img.cols - 1) x2 = img.cols - 1;
111 if (y2 > img.rows - 1) y2 = img.rows - 1;
113 if (x2 - x1 < 2 * p_cell_size) {
114 double diff = (2 * p_cell_size - x2 + x1) / 2.;
115 if (x1 - diff >= 0 && x2 + diff < img.cols) {
118 } else if (x1 - 2 * diff >= 0) {
124 if (y2 - y1 < 2 * p_cell_size) {
125 double diff = (2 * p_cell_size - y2 + y1) / 2.;
126 if (y1 - diff >= 0 && y2 + diff < img.rows) {
129 } else if (y1 - 2 * diff >= 0) {
138 p_pose.cx = x1 + p_pose.w / 2.;
139 p_pose.cy = y1 + p_pose.h / 2.;
141 cv::Mat input_gray, input_rgb = img.clone();
142 if (img.channels() == 3) {
143 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
144 input_gray.convertTo(input_gray, CV_32FC1);
146 img.convertTo(input_gray, CV_32FC1);
148 // don't need too large image
149 if (p_pose.w * p_pose.h > 100. * 100. && (fit_size_x == -1 || fit_size_y == -1)) {
150 std::cout << "resizing image by factor of " << 1 / p_downscale_factor << std::endl;
151 p_resize_image = true;
152 p_pose.scale(p_downscale_factor);
153 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
154 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
155 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
156 if (fit_size_x % p_cell_size != 0 || fit_size_y % p_cell_size != 0) {
157 std::cerr << "Error: Fit size is not multiple of HOG cell size (" << p_cell_size << ")" << std::endl;
158 std::exit(EXIT_FAILURE);
160 p_scale_factor_x = (double)fit_size_x / round(p_pose.w * (1. + p_padding));
161 p_scale_factor_y = (double)fit_size_y / round(p_pose.h * (1. + p_padding));
162 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x << " and verticaly by factor of "
163 << p_scale_factor_y << std::endl;
165 p_pose.scale_x(p_scale_factor_x);
166 p_pose.scale_y(p_scale_factor_y);
167 if (fabs(p_scale_factor_x - 1) > p_floating_error || fabs(p_scale_factor_y - 1) > p_floating_error) {
168 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
169 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
170 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
172 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
173 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
178 // compute win size + fit to fhog cell size
179 p_windows_size.width = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
180 p_windows_size.height = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
181 p_roi.width = p_windows_size.width / p_cell_size;
182 p_roi.height = p_windows_size.height / p_cell_size;
186 for (int i = -int(p_num_scales) / 2; i <= int(p_num_scales) / 2; ++i)
187 p_scales.push_back(std::pow(p_scale_step, i));
189 p_scales.push_back(1.);
192 if (p_roi.height * (p_roi.width / 2 + 1) > 1024) {
193 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
194 "the window dimensions so its size is less or equal to "
195 << 1024 * p_cell_size * p_cell_size * 2 + 1
196 << " pixels . Currently the size of the window is: " << p_windows_size.width << "x" << p_windows_size.height
197 << " which is " << p_windows_size.width * p_windows_size.height << " pixels. " << std::endl;
198 std::exit(EXIT_FAILURE);
201 if (m_use_linearkernel) {
202 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
203 std::exit(EXIT_FAILURE);
205 CudaSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
207 p_xf.create(p_roi.height, p_roi.height / 2 + 1, p_num_of_feats);
210 #if defined(CUFFT) || defined(FFTW)
211 uint width = p_roi.width / 2 + 1;
213 uint width = p_roi.width;
215 p_model_xf.create(p_roi.height, width, p_num_of_feats);
216 p_yf.create(p_roi.height, width, 1);
217 p_xf.create(p_roi.height, width, p_num_of_feats);
220 for (auto scale: p_scales)
221 d.threadctxs.emplace_back(p_roi, p_num_of_feats, scale);
223 d.threadctxs.emplace_back(p_roi, p_num_of_feats, p_num_scales);
226 gaussian_correlation.reset(
227 new GaussianCorrelation(IF_BIG_BATCH(p_num_scales, 1), p_roi));
229 p_current_scale = 1.;
231 double min_size_ratio = std::max(5. * p_cell_size / p_windows_size.width, 5. * p_cell_size / p_windows_size.height);
232 double max_size_ratio =
233 std::min(floor((img.cols + p_windows_size.width / 3) / p_cell_size) * p_cell_size / p_windows_size.width,
234 floor((img.rows + p_windows_size.height / 3) / p_cell_size) * p_cell_size / p_windows_size.height);
235 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
236 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
238 std::cout << "init: img size " << img.cols << "x" << img.rows << std::endl;
239 std::cout << "init: win size " << p_windows_size.width << "x" << p_windows_size.height << std::endl;
240 std::cout << "init: FFT size " << p_roi.width << "x" << p_roi.height << std::endl;
241 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
243 p_output_sigma = std::sqrt(p_pose.w * p_pose.h) * p_output_sigma_factor / p_cell_size;
245 fft.init(p_roi.width, p_roi.height, p_num_of_feats, p_num_scales);
246 fft.set_window(MatDynMem(cosine_window_function(p_roi.width, p_roi.height)));
248 // window weights, i.e. labels
249 MatScales gsl(1, p_roi);
250 gaussian_shaped_labels(p_output_sigma, p_roi.width, p_roi.height).copyTo(gsl.plane(0));
251 fft.forward(gsl, p_yf);
254 // train initial model
255 train(input_gray, input_rgb, 1.0);
258 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat &img, int fit_size_x, int fit_size_y)
260 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
263 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
265 if (p_resize_image) {
267 tmp.scale(p_downscale_factor);
270 } else if (p_fit_to_pw2) {
272 tmp.scale_x(p_scale_factor_x);
273 tmp.scale_y(p_scale_factor_y);
282 BBox_c KCF_Tracker::getBBox()
285 tmp.w *= p_current_scale;
286 tmp.h *= p_current_scale;
288 if (p_resize_image) tmp.scale(1 / p_downscale_factor);
290 tmp.scale_x(1 / p_scale_factor_x);
291 tmp.scale_y(1 / p_scale_factor_y);
297 double KCF_Tracker::getFilterResponse() const
299 return this->max_response;
302 void KCF_Tracker::resizeImgs(cv::Mat &input_rgb, cv::Mat &input_gray)
304 if (p_resize_image) {
305 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
306 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
307 } else if (p_fit_to_pw2 && fabs(p_scale_factor_x - 1) > p_floating_error &&
308 fabs(p_scale_factor_y - 1) > p_floating_error) {
309 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
310 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
311 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
313 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
314 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
319 void KCF_Tracker::findMaxReponse(uint &max_idx, cv::Point2f &new_location) const
323 for (uint j = 0; j < d.threadctxs.size(); ++j) {
324 if (d.threadctxs[j].max.response > max) {
325 max = d.threadctxs[j].max.response;
330 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
331 for (uint j = 0; j < p_scales.size(); ++j) {
332 if (d.threadctxs[0].max[j].response > max) {
333 max = d.threadctxs[0].max[j].response;
338 cv::Point2i &max_response_pt = IF_BIG_BATCH(d.threadctxs[0].max[max_idx].loc, d.threadctxs[max_idx].max.loc);
339 cv::Mat max_response_map = IF_BIG_BATCH(d.threadctxs[0].response.plane(max_idx), d.threadctxs[max_idx].response);
341 DEBUG_PRINTM(max_response_map);
342 DEBUG_PRINT(max_response_pt);
344 // sub pixel quadratic interpolation from neighbours
345 if (max_response_pt.y > max_response_map.rows / 2) // wrap around to negative half-space of vertical axis
346 max_response_pt.y = max_response_pt.y - max_response_map.rows;
347 if (max_response_pt.x > max_response_map.cols / 2) // same for horizontal axis
348 max_response_pt.x = max_response_pt.x - max_response_map.cols;
351 if (m_use_subpixel_localization) {
352 new_location = sub_pixel_peak(max_response_pt, max_response_map);
354 new_location = max_response_pt;
356 DEBUG_PRINT(new_location);
359 void KCF_Tracker::track(cv::Mat &img)
362 if (m_debug) std::cout << "NEW FRAME" << '\n';
364 cv::Mat input_gray, input_rgb = img.clone();
365 if (img.channels() == 3) {
366 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
367 input_gray.convertTo(input_gray, CV_32FC1);
369 img.convertTo(input_gray, CV_32FC1);
371 // don't need too large image
372 resizeImgs(input_rgb, input_gray);
375 for (auto &it : d.threadctxs)
376 it.async_res = std::async(std::launch::async, [this, &input_gray, &input_rgb, &it]() -> void {
377 it.track(*this, input_rgb, input_gray);
379 for (auto const &it : d.threadctxs)
383 // FIXME: Iterate correctly in big batch mode - perhaps have only one element in the list
384 NORMAL_OMP_PARALLEL_FOR
385 for (uint i = 0; i < d.threadctxs.size(); ++i)
386 d.threadctxs[i].track(*this, input_rgb, input_gray);
389 cv::Point2f new_location;
391 findMaxReponse(max_idx, new_location);
393 p_pose.cx += p_current_scale * p_cell_size * double(new_location.x);
394 p_pose.cy += p_current_scale * p_cell_size * double(new_location.y);
396 clamp2(p_pose.cx, 0.0, (img.cols * p_scale_factor_x) - 1);
397 clamp2(p_pose.cy, 0.0, (img.rows * p_scale_factor_y) - 1);
399 clamp2(p_pose.cx, 0.0, img.cols - 1.0);
400 clamp2(p_pose.cy, 0.0, img.rows - 1.0);
403 // sub grid scale interpolation
404 if (m_use_subgrid_scale) {
405 p_current_scale *= sub_grid_scale(max_idx);
407 p_current_scale *= p_scales[max_idx];
410 clamp2(p_current_scale, p_min_max_scale[0], p_min_max_scale[1]);
412 // train at newly estimated target position
413 train(input_rgb, input_gray, p_interp_factor);
416 void ThreadCtx::track(const KCF_Tracker &kcf, cv::Mat &input_rgb, cv::Mat &input_gray)
418 // TODO: Move matrices to thread ctx
419 MatScaleFeats patch_feats(IF_BIG_BATCH(kcf.p_num_scales, 1), kcf.p_num_of_feats, kcf.p_roi);
420 MatScaleFeats temp(IF_BIG_BATCH(kcf.p_num_scales, 1), kcf.p_num_of_feats, kcf.p_roi);
422 BIG_BATCH_OMP_PARALLEL_FOR
423 for (uint i = 0; i < IF_BIG_BATCH(kcf.p_num_scales, 1); ++i)
425 kcf.get_features(input_rgb, input_gray, kcf.p_pose.cx, kcf.p_pose.cy,
426 kcf.p_windows_size.width, kcf.p_windows_size.height,
427 kcf.p_current_scale * IF_BIG_BATCH(kcf.p_scales[i], scale))
428 .copyTo(patch_feats.features(i));
431 kcf.fft.forward_window(patch_feats, zf, temp);
434 if (kcf.m_use_linearkernel) {
435 kzf = zf.mul(kcf.p_model_alphaf).sum_over_channels();
437 gaussian_correlation(kcf, kzf, zf, kcf.p_model_xf, kcf.p_kernel_sigma);
438 DEBUG_PRINTM(kcf.p_model_alphaf);
440 kzf = kzf.mul(kcf.p_model_alphaf);
442 kcf.fft.inverse(kzf, response);
444 DEBUG_PRINTM(response);
446 /* target location is at the maximum response. we must take into
447 account the fact that, if the target doesn't move, the peak
448 will appear at the top-left corner, not at the center (this is
449 discussed in the paper). the responses wrap around cyclically. */
450 double min_val, max_val;
451 cv::Point2i min_loc, max_loc;
453 for (size_t i = 0; i < kcf.p_scales.size(); ++i) {
454 cv::minMaxLoc(response.plane(i), &min_val, &max_val, &min_loc, &max_loc);
455 DEBUG_PRINT(max_loc);
456 double weight = kcf.p_scales[i] < 1. ? kcf.p_scales[i] : 1. / kcf.p_scales[i];
457 max[i].response = max_val * weight;
458 max[i].loc = max_loc;
461 cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
463 DEBUG_PRINT(max_loc);
465 double weight = scale < 1. ? scale : 1. / scale;
466 max.response = max_val * weight;
471 // ****************************************************************************
473 cv::Mat KCF_Tracker::get_features(cv::Mat &input_rgb, cv::Mat &input_gray, int cx, int cy,
474 int size_x, int size_y, double scale) const
476 int size_x_scaled = floor(size_x * scale);
477 int size_y_scaled = floor(size_y * scale);
479 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
480 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
482 // resize to default size
484 // if we downsample use INTER_AREA interpolation
485 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
487 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
490 // get hog(Histogram of Oriented Gradients) features
491 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
493 // get color rgb features (simple r,g,b channels)
494 std::vector<cv::Mat> color_feat;
495 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
496 // resize to default size
498 // if we downsample use INTER_AREA interpolation
499 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_AREA);
501 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x / p_cell_size, size_y / p_cell_size), 0., 0., cv::INTER_LINEAR);
505 if (m_use_color && input_rgb.channels() == 3) {
506 // use rgb color space
507 cv::Mat patch_rgb_norm;
508 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
509 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
510 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
511 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
512 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
513 cv::split(patch_rgb_norm, rgb);
514 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
517 if (m_use_cnfeat && input_rgb.channels() == 3) {
518 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
519 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
522 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
524 int size[] = {p_num_of_feats, p_roi.height, p_roi.width};
525 cv::Mat result(3, size, CV_32F);
526 for (uint i = 0; i < hog_feat.size(); ++i)
527 hog_feat[i].copyTo(cv::Mat(size[1], size[2], CV_32FC1, result.ptr(i)));
532 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
534 cv::Mat labels(dim2, dim1, CV_32FC1);
535 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
536 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
538 double sigma_s = sigma * sigma;
540 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j) {
541 float *row_ptr = labels.ptr<float>(j);
543 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i) {
544 row_ptr[i] = std::exp(-0.5 * (y_s + x * x) / sigma_s); //-1/2*e^((y^2+x^2)/sigma^2)
548 // rotate so that 1 is at top-left corner (see KCF paper for explanation)
549 MatDynMem rot_labels = circshift(labels, range_x[0], range_y[0]);
550 // sanity check, 1 at top left corner
551 assert(rot_labels.at<float>(0, 0) >= 1.f - 1e-10f);
556 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
558 cv::Mat rot_patch(patch.size(), CV_32FC1);
559 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
561 // circular rotate x-axis
563 // move part that does not rotate over the edge
564 cv::Range orig_range(-x_rot, patch.cols);
565 cv::Range rot_range(0, patch.cols - (-x_rot));
566 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
569 orig_range = cv::Range(0, -x_rot);
570 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
571 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
572 } else if (x_rot > 0) {
573 // move part that does not rotate over the edge
574 cv::Range orig_range(0, patch.cols - x_rot);
575 cv::Range rot_range(x_rot, patch.cols);
576 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
579 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
580 rot_range = cv::Range(0, x_rot);
581 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
582 } else { // zero rotation
583 // move part that does not rotate over the edge
584 cv::Range orig_range(0, patch.cols);
585 cv::Range rot_range(0, patch.cols);
586 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
589 // circular rotate y-axis
591 // move part that does not rotate over the edge
592 cv::Range orig_range(-y_rot, patch.rows);
593 cv::Range rot_range(0, patch.rows - (-y_rot));
594 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
597 orig_range = cv::Range(0, -y_rot);
598 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
599 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
600 } else if (y_rot > 0) {
601 // move part that does not rotate over the edge
602 cv::Range orig_range(0, patch.rows - y_rot);
603 cv::Range rot_range(y_rot, patch.rows);
604 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
607 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
608 rot_range = cv::Range(0, y_rot);
609 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
610 } else { // zero rotation
611 // move part that does not rotate over the edge
612 cv::Range orig_range(0, patch.rows);
613 cv::Range rot_range(0, patch.rows);
614 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
620 // hann window actually (Power-of-cosine windows)
621 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
623 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
624 double N_inv = 1. / (static_cast<double>(dim1) - 1.);
625 for (int i = 0; i < dim1; ++i)
626 m1.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
627 N_inv = 1. / (static_cast<double>(dim2) - 1.);
628 for (int i = 0; i < dim2; ++i)
629 m2.at<float>(i) = float(0.5 * (1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv)));
630 cv::Mat ret = m2 * m1;
634 // Returns sub-window of image input centered at [cx, cy] coordinates),
635 // with size [width, height]. If any pixels are outside of the image,
636 // they will replicate the values at the borders.
637 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height) const
641 int x1 = cx - width / 2;
642 int y1 = cy - height / 2;
643 int x2 = cx + width / 2;
644 int y2 = cy + height / 2;
647 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
648 patch.create(height, width, input.type());
649 patch.setTo(double(0.f));
653 int top = 0, bottom = 0, left = 0, right = 0;
655 // fit to image coordinates, set border extensions;
664 if (x2 >= input.cols) {
665 right = x2 - input.cols + width % 2;
670 if (y2 >= input.rows) {
671 bottom = y2 - input.rows + height % 2;
676 if (x2 - x1 == 0 || y2 - y1 == 0)
677 patch = cv::Mat::zeros(height, width, CV_32FC1);
679 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right,
680 cv::BORDER_REPLICATE);
681 // imshow( "copyMakeBorder", patch);
686 assert(patch.cols == width && patch.rows == height);
691 void KCF_Tracker::GaussianCorrelation::operator()(const KCF_Tracker &kcf, ComplexMat &result, const ComplexMat &xf,
692 const ComplexMat &yf, double sigma, bool auto_correlation)
694 xf.sqr_norm(xf_sqr_norm);
695 if (auto_correlation) {
696 yf_sqr_norm = xf_sqr_norm;
698 yf.sqr_norm(yf_sqr_norm);
700 xyf = auto_correlation ? xf.sqr_mag() : xf * yf.conj(); // xf.muln(yf.conj());
702 kcf.fft.inverse(xyf, ifft_res);
704 cuda_gaussian_correlation(ifft_res.deviceMem(), k.deviceMem(), xf_sqr_norm.deviceMem(),
705 auto_correlation ? xf_sqr_norm.deviceMem() : yf_sqr_norm.deviceMem(), sigma,
706 xf.n_channels, xf.n_scales, kcf.p_roi.height, kcf.p_roi.width);
708 // ifft2 and sum over 3rd dimension, we dont care about individual channels
709 //DEBUG_PRINTM(ifft_res);
711 if (xf.channels() != kcf.p_num_scales * kcf.p_num_of_feats)
712 xy_sum.create(ifft_res.size(), CV_32FC1);
714 xy_sum.create(ifft_res.size(), CV_32FC(kcf.p_scales.size()));
716 for (int y = 0; y < ifft_res.rows; ++y) {
717 float *row_ptr = ifft_res.ptr<float>(y);
718 float *row_ptr_sum = xy_sum.ptr<float>(y);
719 for (int x = 0; x < ifft_res.cols; ++x) {
720 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
721 row_ptr_sum[(x * xy_sum.channels()) + sum_ch] += std::accumulate(
722 row_ptr + x * ifft_res.channels() + sum_ch * (ifft_res.channels() / xy_sum.channels()),
723 (row_ptr + x * ifft_res.channels() +
724 (sum_ch + 1) * (ifft_res.channels() / xy_sum.channels())),
729 DEBUG_PRINTM(xy_sum);
731 std::vector<cv::Mat> scales;
732 cv::split(xy_sum, scales);
734 float numel_xf_inv = 1.f / (xf.cols * xf.rows * (xf.channels() / xf.n_scales));
735 for (uint i = 0; i < xf.n_scales; ++i) {
736 cv::Mat k_roi = k.plane(i);
737 cv::exp(-1. / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0),
742 kcf.fft.forward(k, result);
745 float get_response_circular(cv::Point2i &pt, cv::Mat &response)
749 if (x < 0) x = response.cols + x;
750 if (y < 0) y = response.rows + y;
751 if (x >= response.cols) x = x - response.cols;
752 if (y >= response.rows) y = y - response.rows;
754 return response.at<float>(y, x);
757 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point &max_loc, cv::Mat &response) const
759 // find neighbourhood of max_loc (response is circular)
763 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);
764 cv::Point2i p4(max_loc.x - 1, max_loc.y), p5(max_loc.x + 1, max_loc.y);
765 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);
768 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
769 cv::Mat A = (cv::Mat_<float>(9, 6) <<
770 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
771 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
772 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
773 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
774 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
775 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
776 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
777 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
778 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);
779 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
780 get_response_circular(p1, response),
781 get_response_circular(p2, response),
782 get_response_circular(p3, response),
783 get_response_circular(p4, response),
784 get_response_circular(p5, response),
785 get_response_circular(p6, response),
786 get_response_circular(p7, response),
787 get_response_circular(p8, response),
788 get_response_circular(max_loc, response));
791 cv::solve(A, fval, x, cv::DECOMP_SVD);
793 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);
795 cv::Point2f sub_peak(max_loc.x, max_loc.y);
796 if (b > 0 || b < 0) {
797 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
798 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
804 double KCF_Tracker::sub_grid_scale(uint index)
807 if (index >= p_scales.size()) {
808 // interpolate from all values
809 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
810 A.create(p_scales.size(), 3, CV_32FC1);
811 fval.create(p_scales.size(), 1, CV_32FC1);
812 for (size_t i = 0; i < p_scales.size(); ++i) {
813 A.at<float>(i, 0) = float(p_scales[i] * p_scales[i]);
814 A.at<float>(i, 1) = float(p_scales[i]);
815 A.at<float>(i, 2) = 1;
816 fval.at<float>(i) = d.threadctxs.back().IF_BIG_BATCH(max[i].response, max.response);
819 // only from neighbours
820 if (index == 0 || index == p_scales.size() - 1)
821 return p_scales[index];
823 A = (cv::Mat_<float>(3, 3) <<
824 p_scales[index - 1] * p_scales[index - 1], p_scales[index - 1], 1,
825 p_scales[index + 0] * p_scales[index + 0], p_scales[index + 0], 1,
826 p_scales[index + 1] * p_scales[index + 1], p_scales[index + 1], 1);
828 fval = (cv::Mat_<float>(3, 1) <<
829 d.threadctxs.back().max[index - 1].response,
830 d.threadctxs.back().max[index + 0].response,
831 d.threadctxs.back().max[index + 1].response);
833 fval = (cv::Mat_<float>(3, 1) <<
834 d.threadctxs[index - 1].max.response,
835 d.threadctxs[index + 0].max.response,
836 d.threadctxs[index + 1].max.response);
841 cv::solve(A, fval, x, cv::DECOMP_SVD);
842 float a = x.at<float>(0), b = x.at<float>(1);
843 double scale = p_scales[index];
845 scale = -b / (2 * a);