11 #include "fft_cufft.h"
14 #include "fft_opencv.h"
22 #define DEBUG_PRINT(obj) if (m_debug) {std::cout << #obj << " @" << __LINE__ << std::endl << (obj) << std::endl;}
23 #define DEBUG_PRINTM(obj) if (m_debug) {std::cout << #obj << " @" << __LINE__ << " " << (obj).size() << " CH: " << (obj).channels() << std::endl << (obj) << std::endl;}
25 KCF_Tracker::KCF_Tracker(double padding, double kernel_sigma, double lambda, double interp_factor, double output_sigma_factor, int cell_size) :
27 p_padding(padding), p_output_sigma_factor(output_sigma_factor), p_kernel_sigma(kernel_sigma),
28 p_lambda(lambda), p_interp_factor(interp_factor), p_cell_size(cell_size) {}
30 KCF_Tracker::KCF_Tracker()
33 KCF_Tracker::~KCF_Tracker()
37 CudaSafeCall(cudaFreeHost(xf_sqr_norm));
38 CudaSafeCall(cudaFreeHost(yf_sqr_norm));
45 void KCF_Tracker::init(cv::Mat &img, const cv::Rect & bbox)
47 //check boundary, enforce min size
48 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
50 if (x2 > img.cols-1) x2 = img.cols - 1;
52 if (y2 > img.rows-1) y2 = img.rows - 1;
54 if (x2-x1 < 2*p_cell_size) {
55 double diff = (2*p_cell_size -x2+x1)/2.;
56 if (x1 - diff >= 0 && x2 + diff < img.cols){
59 } else if (x1 - 2*diff >= 0) {
65 if (y2-y1 < 2*p_cell_size) {
66 double diff = (2*p_cell_size -y2+y1)/2.;
67 if (y1 - diff >= 0 && y2 + diff < img.rows){
70 } else if (y1 - 2*diff >= 0) {
79 p_pose.cx = x1 + p_pose.w/2.;
80 p_pose.cy = y1 + p_pose.h /2.;
83 cv::Mat input_gray, input_rgb = img.clone();
84 if (img.channels() == 3){
85 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
86 input_gray.convertTo(input_gray, CV_32FC1);
88 img.convertTo(input_gray, CV_32FC1);
90 // don't need too large image
91 if (p_pose.w * p_pose.h > 100.*100.) {
92 std::cout << "resizing image by factor of " << 1/p_downscale_factor << std::endl;
93 p_resize_image = true;
94 p_pose.scale(p_downscale_factor);
95 cv::resize(input_gray, input_gray, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
96 cv::resize(input_rgb, input_rgb, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
99 //compute win size + fit to fhog cell size
100 p_windows_size[0] = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
101 p_windows_size[1] = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
105 for (int i = -p_num_scales/2; i <= p_num_scales/2; ++i)
106 p_scales.push_back(std::pow(p_scale_step, i));
108 p_scales.push_back(1.);
111 cudaSetDeviceFlags(cudaDeviceMapHost);
112 CudaSafeCall(cudaHostAlloc((void**)&xf_sqr_norm, p_num_scales*sizeof(float), cudaHostAllocMapped));
113 CudaSafeCall(cudaHostGetDevicePointer((void**)&xf_sqr_norm_d, (void*)xf_sqr_norm, 0));
114 std::cout << &xf_sqr_norm << std::endl;
115 CudaSafeCall(cudaHostAlloc((void**)&yf_sqr_norm, sizeof(float), cudaHostAllocMapped));
116 CudaSafeCall(cudaHostGetDevicePointer((void**)&yf_sqr_norm_d, (void*)yf_sqr_norm, 0));
118 xf_sqr_norm = (float*) malloc(p_num_scales*sizeof(float));
119 xf_sqr_norm = (float*) malloc(sizeof(float));
122 p_current_scale = 1.;
124 double min_size_ratio = std::max(5.*p_cell_size/p_windows_size[0], 5.*p_cell_size/p_windows_size[1]);
125 double max_size_ratio = std::min(floor((img.cols + p_windows_size[0]/3)/p_cell_size)*p_cell_size/p_windows_size[0], floor((img.rows + p_windows_size[1]/3)/p_cell_size)*p_cell_size/p_windows_size[1]);
126 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
127 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
129 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
130 std::cout << "init: win size. " << p_windows_size[0] << " " << p_windows_size[1] << std::endl;
131 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
133 p_output_sigma = std::sqrt(p_pose.w*p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
135 //window weights, i.e. labels
137 if(m_use_color) p_num_of_feats += 3;
138 if(m_use_cnfeat) p_num_of_feats += 10;
139 p_poi_width = p_windows_size[0]/p_cell_size;
140 p_poi_height = p_windows_size[1]/p_cell_size;
142 fft.init(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size, p_num_of_feats, p_num_scales, m_use_big_batch);
143 p_yf = fft.forward(gaussian_shaped_labels(p_output_sigma, p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
144 fft.set_window(cosine_window_function(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
146 //obtain a sub-window for training initial model
147 std::vector<cv::Mat> path_feat = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1]);
148 p_model_xf = fft.forward_window(path_feat);
149 DEBUG_PRINTM(p_model_xf);
151 if (m_use_linearkernel) {
152 ComplexMat xfconj = p_model_xf.conj();
153 p_model_alphaf_num = xfconj.mul(p_yf);
154 p_model_alphaf_den = (p_model_xf * xfconj);
156 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
157 ComplexMat kf = gaussian_correlation(p_model_xf, p_model_xf, p_kernel_sigma, true);
159 p_model_alphaf_num = p_yf * kf;
160 DEBUG_PRINTM(p_model_alphaf_num);
161 p_model_alphaf_den = kf * (kf + p_lambda);
162 DEBUG_PRINTM(p_model_alphaf_den);
164 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
165 DEBUG_PRINTM(p_model_alphaf);
166 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
169 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat & img)
171 init(img, bbox.get_rect());
174 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
176 if (p_resize_image) {
178 tmp.scale(p_downscale_factor);
187 BBox_c KCF_Tracker::getBBox()
190 tmp.w *= p_current_scale;
191 tmp.h *= p_current_scale;
194 tmp.scale(1/p_downscale_factor);
199 void KCF_Tracker::track(cv::Mat &img)
202 std::cout << "NEW FRAME" << '\n';
203 cv::Mat input_gray, input_rgb = img.clone();
204 if (img.channels() == 3){
205 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
206 input_gray.convertTo(input_gray, CV_32FC1);
208 img.convertTo(input_gray, CV_32FC1);
210 // don't need too large image
211 if (p_resize_image) {
212 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
213 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
217 std::vector<cv::Mat> patch_feat;
218 double max_response = -1.;
219 cv::Mat max_response_map;
220 cv::Point2i max_response_pt;
222 std::vector<double> scale_responses;
224 if (m_use_multithreading){
225 std::vector<std::future<cv::Mat>> async_res(p_scales.size());
226 for (size_t i = 0; i < p_scales.size(); ++i) {
227 async_res[i] = std::async(std::launch::async,
228 [this, &input_gray, &input_rgb, i]() -> cv::Mat
230 std::vector<cv::Mat> patch_feat_async = get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size[0],
231 this->p_windows_size[1], this->p_current_scale * this->p_scales[i]);
232 ComplexMat zf = fft.forward_window(patch_feat_async);
233 if (m_use_linearkernel)
234 return fft.inverse((p_model_alphaf * zf).sum_over_channels());
236 ComplexMat kzf = gaussian_correlation(zf, this->p_model_xf, this->p_kernel_sigma);
237 return fft.inverse(this->p_model_alphaf * kzf);
242 for (size_t i = 0; i < p_scales.size(); ++i) {
245 cv::Mat response = async_res[i].get();
247 double min_val, max_val;
248 cv::Point2i min_loc, max_loc;
249 cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
251 double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
252 if (max_val*weight > max_response) {
253 max_response = max_val*weight;
254 max_response_map = response;
255 max_response_pt = max_loc;
258 scale_responses.push_back(max_val*weight);
260 } else if(m_use_big_batch){
261 #pragma omp parallel for ordered
262 for (size_t i = 0; i < p_scales.size(); ++i) {
263 std::vector<cv::Mat> tmp = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], p_current_scale * p_scales[i]);
265 patch_feat.insert(std::end(patch_feat), std::begin(tmp), std::end(tmp));
267 ComplexMat zf = fft.forward_window(patch_feat);
271 if (m_use_linearkernel)
272 response = fft.inverse((zf.mul2(p_model_alphaf)).sum_over_channels());
274 ComplexMat kzf = gaussian_correlation(zf, p_model_xf, p_kernel_sigma);
275 DEBUG_PRINTM(p_model_alphaf);
277 response = fft.inverse(kzf.mul(p_model_alphaf));
279 DEBUG_PRINTM(response);
280 std::vector<cv::Mat> scales;
281 cv::split(response,scales);
283 /* target location is at the maximum response. we must take into
284 account the fact that, if the target doesn't move, the peak
285 will appear at the top-left corner, not at the center (this is
286 discussed in the paper). the responses wrap around cyclically. */
287 for (size_t i = 0; i < p_scales.size(); ++i) {
288 double min_val, max_val;
289 cv::Point2i min_loc, max_loc;
290 cv::minMaxLoc(scales[i], &min_val, &max_val, &min_loc, &max_loc);
291 DEBUG_PRINT(max_loc);
293 double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
295 if (max_val*weight > max_response) {
296 max_response = max_val*weight;
297 max_response_map = scales[i];
298 max_response_pt = max_loc;
301 scale_responses.push_back(max_val*weight);
304 #pragma omp parallel for ordered private(patch_feat) schedule(dynamic)
305 for (size_t i = 0; i < p_scales.size(); ++i) {
306 patch_feat = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], p_current_scale * p_scales[i]);
307 ComplexMat zf = fft.forward_window(patch_feat);
310 if (m_use_linearkernel)
311 response = fft.inverse((p_model_alphaf * zf).sum_over_channels());
313 ComplexMat kzf = gaussian_correlation(zf, p_model_xf, p_kernel_sigma);
314 DEBUG_PRINTM(p_model_alphaf);
316 DEBUG_PRINTM(p_model_alphaf * kzf);
317 response = fft.inverse(p_model_alphaf * kzf);
319 DEBUG_PRINTM(response);
321 /* target location is at the maximum response. we must take into
322 account the fact that, if the target doesn't move, the peak
323 will appear at the top-left corner, not at the center (this is
324 discussed in the paper). the responses wrap around cyclically. */
325 double min_val, max_val;
326 cv::Point2i min_loc, max_loc;
327 cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
328 DEBUG_PRINT(max_loc);
330 double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
333 if (max_val*weight > max_response) {
334 max_response = max_val*weight;
335 max_response_map = response;
336 max_response_pt = max_loc;
341 scale_responses.push_back(max_val*weight);
344 DEBUG_PRINTM(max_response_map);
345 DEBUG_PRINT(max_response_pt);
346 //sub pixel quadratic interpolation from neighbours
347 if (max_response_pt.y > max_response_map.rows / 2) //wrap around to negative half-space of vertical axis
348 max_response_pt.y = max_response_pt.y - max_response_map.rows;
349 if (max_response_pt.x > max_response_map.cols / 2) //same for horizontal axis
350 max_response_pt.x = max_response_pt.x - max_response_map.cols;
352 cv::Point2f new_location(max_response_pt.x, max_response_pt.y);
353 DEBUG_PRINT(new_location);
355 if (m_use_subpixel_localization)
356 new_location = sub_pixel_peak(max_response_pt, max_response_map);
357 DEBUG_PRINT(new_location);
359 p_pose.cx += p_current_scale*p_cell_size*new_location.x;
360 p_pose.cy += p_current_scale*p_cell_size*new_location.y;
361 if (p_pose.cx < 0) p_pose.cx = 0;
362 if (p_pose.cx > img.cols-1) p_pose.cx = img.cols-1;
363 if (p_pose.cy < 0) p_pose.cy = 0;
364 if (p_pose.cy > img.rows-1) p_pose.cy = img.rows-1;
366 //sub grid scale interpolation
367 double new_scale = p_scales[scale_index];
368 if (m_use_subgrid_scale)
369 new_scale = sub_grid_scale(scale_responses, scale_index);
371 p_current_scale *= new_scale;
373 if (p_current_scale < p_min_max_scale[0])
374 p_current_scale = p_min_max_scale[0];
375 if (p_current_scale > p_min_max_scale[1])
376 p_current_scale = p_min_max_scale[1];
377 //obtain a subwindow for training at newly estimated target position
378 patch_feat = get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], p_current_scale);
379 ComplexMat xf = fft.forward_window(patch_feat);
381 //subsequent frames, interpolate model
382 p_model_xf = p_model_xf * (1. - p_interp_factor) + xf * p_interp_factor;
384 ComplexMat alphaf_num, alphaf_den;
386 if (m_use_linearkernel) {
387 ComplexMat xfconj = xf.conj();
388 alphaf_num = xfconj.mul(p_yf);
389 alphaf_den = (xf * xfconj);
391 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
392 ComplexMat kf = gaussian_correlation(xf, xf, p_kernel_sigma, true);
393 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
394 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
395 alphaf_num = p_yf * kf;
396 alphaf_den = kf * (kf + p_lambda);
399 p_model_alphaf_num = p_model_alphaf_num * (1. - p_interp_factor) + alphaf_num * p_interp_factor;
400 p_model_alphaf_den = p_model_alphaf_den * (1. - p_interp_factor) + alphaf_den * p_interp_factor;
401 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
404 // ****************************************************************************
406 std::vector<cv::Mat> KCF_Tracker::get_features(cv::Mat & input_rgb, cv::Mat & input_gray, int cx, int cy, int size_x, int size_y, double scale)
408 int size_x_scaled = floor(size_x*scale);
409 int size_y_scaled = floor(size_y*scale);
411 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
412 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
414 //resize to default size
416 //if we downsample use INTER_AREA interpolation
417 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
419 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
422 // get hog(Histogram of Oriented Gradients) features
423 std::vector<cv::Mat> hog_feat = FHoG::extract(patch_gray, 2, p_cell_size, 9);
425 //get color rgb features (simple r,g,b channels)
426 std::vector<cv::Mat> color_feat;
427 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
428 //resize to default size
430 //if we downsample use INTER_AREA interpolation
431 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_AREA);
433 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_LINEAR);
437 if (m_use_color && input_rgb.channels() == 3) {
438 //use rgb color space
439 cv::Mat patch_rgb_norm;
440 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
441 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
442 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
443 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
444 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
445 cv::split(patch_rgb_norm, rgb);
446 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
449 if (m_use_cnfeat && input_rgb.channels() == 3) {
450 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
451 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
454 hog_feat.insert(hog_feat.end(), color_feat.begin(), color_feat.end());
458 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
460 cv::Mat labels(dim2, dim1, CV_32FC1);
461 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
462 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
464 double sigma_s = sigma*sigma;
466 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j){
467 float * row_ptr = labels.ptr<float>(j);
469 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i){
470 row_ptr[i] = std::exp(-0.5 * (y_s + x*x) / sigma_s);//-1/2*e^((y^2+x^2)/sigma^2)
474 //rotate so that 1 is at top-left corner (see KCF paper for explanation)
475 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
476 //sanity check, 1 at top left corner
477 assert(rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
482 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
484 cv::Mat rot_patch(patch.size(), CV_32FC1);
485 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
487 //circular rotate x-axis
489 //move part that does not rotate over the edge
490 cv::Range orig_range(-x_rot, patch.cols);
491 cv::Range rot_range(0, patch.cols - (-x_rot));
492 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
495 orig_range = cv::Range(0, -x_rot);
496 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
497 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
498 }else if (x_rot > 0){
499 //move part that does not rotate over the edge
500 cv::Range orig_range(0, patch.cols - x_rot);
501 cv::Range rot_range(x_rot, patch.cols);
502 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
505 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
506 rot_range = cv::Range(0, x_rot);
507 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
508 }else { //zero rotation
509 //move part that does not rotate over the edge
510 cv::Range orig_range(0, patch.cols);
511 cv::Range rot_range(0, patch.cols);
512 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
515 //circular rotate y-axis
517 //move part that does not rotate over the edge
518 cv::Range orig_range(-y_rot, patch.rows);
519 cv::Range rot_range(0, patch.rows - (-y_rot));
520 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
523 orig_range = cv::Range(0, -y_rot);
524 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
525 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
526 }else if (y_rot > 0){
527 //move part that does not rotate over the edge
528 cv::Range orig_range(0, patch.rows - y_rot);
529 cv::Range rot_range(y_rot, patch.rows);
530 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
533 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
534 rot_range = cv::Range(0, y_rot);
535 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
536 }else { //zero rotation
537 //move part that does not rotate over the edge
538 cv::Range orig_range(0, patch.rows);
539 cv::Range rot_range(0, patch.rows);
540 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
546 //hann window actually (Power-of-cosine windows)
547 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
549 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
550 double N_inv = 1./(static_cast<double>(dim1)-1.);
551 for (int i = 0; i < dim1; ++i)
552 m1.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
553 N_inv = 1./(static_cast<double>(dim2)-1.);
554 for (int i = 0; i < dim2; ++i)
555 m2.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
560 // Returns sub-window of image input centered at [cx, cy] coordinates),
561 // with size [width, height]. If any pixels are outside of the image,
562 // they will replicate the values at the borders.
563 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
567 int x1 = cx - width/2;
568 int y1 = cy - height/2;
569 int x2 = cx + width/2;
570 int y2 = cy + height/2;
573 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
574 patch.create(height, width, input.type());
579 int top = 0, bottom = 0, left = 0, right = 0;
581 //fit to image coordinates, set border extensions;
590 if (x2 >= input.cols) {
591 right = x2 - input.cols + width % 2;
596 if (y2 >= input.rows) {
597 bottom = y2 - input.rows + height % 2;
602 if (x2 - x1 == 0 || y2 - y1 == 0)
603 patch = cv::Mat::zeros(height, width, CV_32FC1);
606 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right, cv::BORDER_REPLICATE);
607 // imshow( "copyMakeBorder", patch);
612 assert(patch.cols == width && patch.rows == height);
617 ComplexMat KCF_Tracker::gaussian_correlation(const ComplexMat &xf, const ComplexMat &yf, double sigma, bool auto_correlation)
620 xf.sqr_norm(xf_sqr_norm_d);
622 xf.sqr_norm(xf_sqr_norm);
624 if(auto_correlation){
625 yf_sqr_norm[0] = xf_sqr_norm[0];
628 yf.sqr_norm(yf_sqr_norm_d);
630 yf.sqr_norm(yf_sqr_norm);
635 xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
638 //ifft2 and sum over 3rd dimension, we dont care about individual channels
639 cv::Mat ifft2_res = fft.inverse(xyf);
641 if(xf.channels() != p_num_scales*p_num_of_feats)
642 xy_sum.create(ifft2_res.size(), CV_32FC1);
644 xy_sum.create(ifft2_res.size(), CV_32FC(p_scales.size()));
646 for (int y = 0; y < ifft2_res.rows; ++y) {
647 float * row_ptr = ifft2_res.ptr<float>(y);
648 float * row_ptr_sum = xy_sum.ptr<float>(y);
649 for (int x = 0; x < ifft2_res.cols; ++x) {
650 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
651 row_ptr_sum[(x*xy_sum.channels())+sum_ch] += std::accumulate(row_ptr + x*ifft2_res.channels() + sum_ch*(ifft2_res.channels()/xy_sum.channels()),
652 (row_ptr + x*ifft2_res.channels() + (sum_ch+1)*(ifft2_res.channels()/xy_sum.channels())), 0.f);
656 DEBUG_PRINTM(ifft2_res);
657 DEBUG_PRINTM(xy_sum);
659 std::vector<cv::Mat> scales;
660 cv::split(xy_sum,scales);
661 cv::Mat in_all(scales[0].rows * xf.n_scales, scales[0].cols, CV_32F);
663 float numel_xf_inv = 1.f/(xf.cols * xf.rows * (xf.channels()/xf.n_scales));
664 for(int i = 0; i < xf.n_scales; ++i){
665 cv::Mat in_roi(in_all, cv::Rect(0, i*scales[0].rows, scales[0].cols, scales[0].rows));
666 cv::exp(- 1.f / (sigma * sigma) * cv::max((xf_sqr_norm[i] + yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0), in_roi);
667 DEBUG_PRINTM(in_roi);
670 DEBUG_PRINTM(in_all);
671 return fft.forward(in_all);
674 float get_response_circular(cv::Point2i & pt, cv::Mat & response)
679 x = response.cols + x;
681 y = response.rows + y;
682 if (x >= response.cols)
683 x = x - response.cols;
684 if (y >= response.rows)
685 y = y - response.rows;
687 return response.at<float>(y,x);
690 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point & max_loc, cv::Mat & response)
692 //find neighbourhood of max_loc (response is circular)
696 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);
697 cv::Point2i p4(max_loc.x-1, max_loc.y), p5(max_loc.x+1, max_loc.y);
698 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);
700 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
701 cv::Mat A = (cv::Mat_<float>(9, 6) <<
702 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
703 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
704 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
705 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
706 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
707 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
708 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
709 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
710 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);
711 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
712 get_response_circular(p1, response),
713 get_response_circular(p2, response),
714 get_response_circular(p3, response),
715 get_response_circular(p4, response),
716 get_response_circular(p5, response),
717 get_response_circular(p6, response),
718 get_response_circular(p7, response),
719 get_response_circular(p8, response),
720 get_response_circular(max_loc, response));
722 cv::solve(A, fval, x, cv::DECOMP_SVD);
724 double a = x.at<float>(0), b = x.at<float>(1), c = x.at<float>(2),
725 d = x.at<float>(3), e = x.at<float>(4);
727 cv::Point2f sub_peak(max_loc.x, max_loc.y);
728 if (b > 0 || b < 0) {
729 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
730 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
736 double KCF_Tracker::sub_grid_scale(std::vector<double> & responses, int index)
739 if (index < 0 || index > (int)p_scales.size()-1) {
740 // interpolate from all values
741 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
742 A.create(p_scales.size(), 3, CV_32FC1);
743 fval.create(p_scales.size(), 1, CV_32FC1);
744 for (size_t i = 0; i < p_scales.size(); ++i) {
745 A.at<float>(i, 0) = p_scales[i] * p_scales[i];
746 A.at<float>(i, 1) = p_scales[i];
747 A.at<float>(i, 2) = 1;
748 fval.at<float>(i) = responses[i];
751 //only from neighbours
752 if (index == 0 || index == (int)p_scales.size()-1)
753 return p_scales[index];
755 A = (cv::Mat_<float>(3, 3) <<
756 p_scales[index-1] * p_scales[index-1], p_scales[index-1], 1,
757 p_scales[index] * p_scales[index], p_scales[index], 1,
758 p_scales[index+1] * p_scales[index+1], p_scales[index+1], 1);
759 fval = (cv::Mat_<float>(3, 1) << responses[index-1], responses[index], responses[index+1]);
763 cv::solve(A, fval, x, cv::DECOMP_SVD);
764 double a = x.at<float>(0), b = x.at<float>(1);
765 double scale = p_scales[index];
767 scale = -b / (2 * a);