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 for (int i = 0;i < p_num_scales;++i) {
38 CudaSafeCall(cudaFreeHost(scale_vars[i].xf_sqr_norm));
39 CudaSafeCall(cudaFreeHost(scale_vars[i].yf_sqr_norm));
40 CudaSafeCall(cudaFree(scale_vars[i].gauss_corr_res));
43 for (int i = 0;i < p_num_scales;++i) {
44 free(scale_vars[i].xf_sqr_norm);
45 free(scale_vars[i].yf_sqr_norm);
50 void KCF_Tracker::init(cv::Mat &img, const cv::Rect & bbox, int fit_size_x, int fit_size_y)
52 //check boundary, enforce min size
53 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
55 if (x2 > img.cols-1) x2 = img.cols - 1;
57 if (y2 > img.rows-1) y2 = img.rows - 1;
59 if (x2-x1 < 2*p_cell_size) {
60 double diff = (2*p_cell_size -x2+x1)/2.;
61 if (x1 - diff >= 0 && x2 + diff < img.cols){
64 } else if (x1 - 2*diff >= 0) {
70 if (y2-y1 < 2*p_cell_size) {
71 double diff = (2*p_cell_size -y2+y1)/2.;
72 if (y1 - diff >= 0 && y2 + diff < img.rows){
75 } else if (y1 - 2*diff >= 0) {
84 p_pose.cx = x1 + p_pose.w/2.;
85 p_pose.cy = y1 + p_pose.h /2.;
88 cv::Mat input_gray, input_rgb = img.clone();
89 if (img.channels() == 3){
90 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
91 input_gray.convertTo(input_gray, CV_32FC1);
93 img.convertTo(input_gray, CV_32FC1);
95 // don't need too large image
96 if (p_pose.w * p_pose.h > 100.*100. && (fit_size_x == -1 || fit_size_y == -1)) {
97 std::cout << "resizing image by factor of " << 1/p_downscale_factor << std::endl;
98 p_resize_image = true;
99 p_pose.scale(p_downscale_factor);
100 cv::resize(input_gray, input_gray, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
101 cv::resize(input_rgb, input_rgb, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
102 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
103 if (fit_size_x%p_cell_size != 0 || fit_size_y%p_cell_size != 0) {
104 std::cerr << "Fit size does not fit to hog cell size. The dimensions have to be divisible by HOG cell size, which is: " << p_cell_size << std::endl;;
105 std::exit(EXIT_FAILURE);
108 if (( tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_x)
109 p_scale_factor_x = fit_size_x/tmp;
110 if (( tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_y)
111 p_scale_factor_y = fit_size_y/tmp;
112 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x
113 << " and verticaly by factor of " << p_scale_factor_y << std::endl;
115 p_pose.scale_x(p_scale_factor_x);
116 p_pose.scale_y(p_scale_factor_y);
117 if (p_scale_factor_x != 1 && p_scale_factor_y != 1) {
118 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
119 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
120 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
122 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
123 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
128 //compute win size + fit to fhog cell size
129 p_windows_size[0] = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
130 p_windows_size[1] = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
134 for (int i = -p_num_scales/2; i <= p_num_scales/2; ++i)
135 p_scales.push_back(std::pow(p_scale_step, i));
137 p_scales.push_back(1.);
139 for (int i = 0;i<p_num_scales;++i) {
140 scale_vars.push_back(Scale_vars());
144 if(m_use_color) p_num_of_feats += 3;
145 if(m_use_cnfeat) p_num_of_feats += 10;
146 p_roi_width = p_windows_size[0]/p_cell_size;
147 p_roi_height = p_windows_size[1]/p_cell_size;
150 int alloc_size = p_num_scales;
156 if (p_windows_size[1]/p_cell_size*(p_windows_size[0]/p_cell_size/2+1) > 1024) {
157 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
158 "the window dimensions so its size is less or equal to " << 1024*p_cell_size*p_cell_size*2+1 <<
159 " pixels . Currently the size of the window is: " << p_windows_size[0] << "x" << p_windows_size[1] <<
160 " which is " << p_windows_size[0]*p_windows_size[1] << " pixels. " << std::endl;
161 std::exit(EXIT_FAILURE);
164 if (m_use_linearkernel){
165 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
166 std::exit(EXIT_FAILURE);
168 cudaSetDeviceFlags(cudaDeviceMapHost);
170 for (int i = 0;i<p_num_scales;++i) {
171 CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].xf_sqr_norm, alloc_size*sizeof(float), cudaHostAllocMapped));
172 CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].xf_sqr_norm_d, (void*)scale_vars[i].xf_sqr_norm, 0));
174 CudaSafeCall(cudaHostAlloc((void**)&scale_vars[i].yf_sqr_norm, sizeof(float), cudaHostAllocMapped));
175 CudaSafeCall(cudaHostGetDevicePointer((void**)&scale_vars[i].yf_sqr_norm_d, (void*)scale_vars[i].yf_sqr_norm, 0));
177 CudaSafeCall(cudaMalloc((void**)&scale_vars[i].gauss_corr_res, (p_windows_size[0]/p_cell_size)*(p_windows_size[1]/p_cell_size)*alloc_size*sizeof(float)));
180 for (int i = 0;i<p_num_scales;++i) {
181 scale_vars[i].xf_sqr_norm = (float*) malloc(alloc_size*sizeof(float));
182 scale_vars[i].yf_sqr_norm = (float*) malloc(sizeof(float));
184 scale_vars[i].patch_feats.reserve(p_num_of_feats);
186 scale_vars[i].zf = ComplexMat(p_windows_size[1]/p_cell_size, p_windows_size[0]/p_cell_size, p_num_of_feats);
190 p_current_scale = 1.;
192 double min_size_ratio = std::max(5.*p_cell_size/p_windows_size[0], 5.*p_cell_size/p_windows_size[1]);
193 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]);
194 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
195 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
197 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
198 std::cout << "init: win size. " << p_windows_size[0] << " " << p_windows_size[1] << std::endl;
199 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
201 p_output_sigma = std::sqrt(p_pose.w*p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
203 //window weights, i.e. labels
204 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);
205 p_yf = fft.forward(gaussian_shaped_labels(p_output_sigma, p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
206 fft.set_window(cosine_window_function(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
208 //obtain a sub-window for training initial model
209 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], scale_vars[0]);
210 p_model_xf = fft.forward_window(scale_vars[0].patch_feats);
211 DEBUG_PRINTM(p_model_xf);
212 scale_vars[0].flag = Track_flags::AUTO_CORRELATION;
214 if (m_use_linearkernel) {
215 ComplexMat xfconj = p_model_xf.conj();
216 p_model_alphaf_num = xfconj.mul(p_yf);
217 p_model_alphaf_den = (p_model_xf * xfconj);
219 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
220 gaussian_correlation(scale_vars[0], p_model_xf, p_model_xf, p_kernel_sigma, true);
221 DEBUG_PRINTM(scale_vars[0].kf);
222 p_model_alphaf_num = p_yf * scale_vars[0].kf;
223 DEBUG_PRINTM(p_model_alphaf_num);
224 p_model_alphaf_den = scale_vars[0].kf * (scale_vars[0].kf + p_lambda);
225 DEBUG_PRINTM(p_model_alphaf_den);
227 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
228 DEBUG_PRINTM(p_model_alphaf);
229 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
232 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat & img, int fit_size_x, int fit_size_y)
234 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
237 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
239 if (p_resize_image) {
241 tmp.scale(p_downscale_factor);
244 } else if (p_fit_to_pw2) {
246 tmp.scale_x(p_scale_factor_x);
247 tmp.scale_y(p_scale_factor_y);
256 BBox_c KCF_Tracker::getBBox()
259 tmp.w *= p_current_scale;
260 tmp.h *= p_current_scale;
263 tmp.scale(1/p_downscale_factor);
265 tmp.scale_x(1/p_scale_factor_x);
266 tmp.scale_y(1/p_scale_factor_y);
272 void KCF_Tracker::track(cv::Mat &img)
275 std::cout << "NEW FRAME" << '\n';
276 cv::Mat input_gray, input_rgb = img.clone();
277 if (img.channels() == 3){
278 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
279 input_gray.convertTo(input_gray, CV_32FC1);
281 img.convertTo(input_gray, CV_32FC1);
283 // don't need too large image
284 if (p_resize_image) {
285 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
286 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
287 } else if (p_fit_to_pw2 && p_scale_factor_x != 1 && p_scale_factor_y != 1) {
288 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
289 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
290 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
292 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
293 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
297 double max_response = -1.;
299 cv::Point2i *max_response_pt = nullptr;
300 cv::Mat *max_response_map = nullptr;
302 for (size_t i = 0; i < p_scales.size(); ++i) {
303 scale_track(this->scale_vars[i], input_rgb, input_gray, this->p_scales[i]);
305 if (this->scale_vars[i].max_response > max_response) {
306 max_response = this->scale_vars[i].max_response;
307 max_response_pt = & this->scale_vars[i].max_loc;
308 max_response_map = & this->scale_vars[i].response;
313 DEBUG_PRINTM(*max_response_map);
314 DEBUG_PRINT(*max_response_pt);
316 //sub pixel quadratic interpolation from neighbours
317 if (max_response_pt->y > max_response_map->rows / 2) //wrap around to negative half-space of vertical axis
318 max_response_pt->y = max_response_pt->y - max_response_map->rows;
319 if (max_response_pt->x > max_response_map->cols / 2) //same for horizontal axis
320 max_response_pt->x = max_response_pt->x - max_response_map->cols;
322 cv::Point2f new_location(max_response_pt->x, max_response_pt->y);
323 DEBUG_PRINT(new_location);
325 if (m_use_subpixel_localization)
326 new_location = sub_pixel_peak(*max_response_pt, *max_response_map);
327 DEBUG_PRINT(new_location);
329 p_pose.cx += p_current_scale*p_cell_size*new_location.x;
330 p_pose.cy += p_current_scale*p_cell_size*new_location.y;
332 if (p_pose.cx < 0) p_pose.cx = 0;
333 if (p_pose.cx > (img.cols*p_scale_factor_x)-1) p_pose.cx = (img.cols*p_scale_factor_x)-1;
334 if (p_pose.cy < 0) p_pose.cy = 0;
335 if (p_pose.cy > (img.rows*p_scale_factor_y)-1) p_pose.cy = (img.rows*p_scale_factor_y)-1;
337 if (p_pose.cx < 0) p_pose.cx = 0;
338 if (p_pose.cx > img.cols-1) p_pose.cx = img.cols-1;
339 if (p_pose.cy < 0) p_pose.cy = 0;
340 if (p_pose.cy > img.rows-1) p_pose.cy = img.rows-1;
343 //sub grid scale interpolation
344 double new_scale = p_scales[scale_index];
345 if (m_use_subgrid_scale)
346 new_scale = sub_grid_scale(scale_index);
348 p_current_scale *= new_scale;
350 if (p_current_scale < p_min_max_scale[0])
351 p_current_scale = p_min_max_scale[0];
352 if (p_current_scale > p_min_max_scale[1])
353 p_current_scale = p_min_max_scale[1];
354 //obtain a subwindow for training at newly estimated target position
355 get_features(input_rgb, input_gray, p_pose.cx, p_pose.cy, p_windows_size[0], p_windows_size[1], scale_vars[0], p_current_scale);
356 ComplexMat xf = fft.forward_window(scale_vars[0].patch_feats);
358 //subsequent frames, interpolate model
359 p_model_xf = p_model_xf * (1. - p_interp_factor) + xf * p_interp_factor;
361 ComplexMat alphaf_num, alphaf_den;
362 scale_vars[0].flag = Track_flags::AUTO_CORRELATION;
364 if (m_use_linearkernel) {
365 ComplexMat xfconj = xf.conj();
366 alphaf_num = xfconj.mul(p_yf);
367 alphaf_den = (xf * xfconj);
369 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
370 gaussian_correlation(scale_vars[0], xf, xf, p_kernel_sigma, true);
371 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
372 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
373 alphaf_num = p_yf * scale_vars[0].kf;
374 alphaf_den = scale_vars[0].kf * (scale_vars[0].kf + p_lambda);
377 p_model_alphaf_num = p_model_alphaf_num * (1. - p_interp_factor) + alphaf_num * p_interp_factor;
378 p_model_alphaf_den = p_model_alphaf_den * (1. - p_interp_factor) + alphaf_den * p_interp_factor;
379 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
382 // ****************************************************************************
384 void KCF_Tracker::scale_track(Scale_vars & vars, cv::Mat & input_rgb, cv::Mat & input_gray, double scale)
386 get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size[0], this->p_windows_size[1],
387 vars, this->p_current_scale * scale);
389 fft.forward_window(vars);
390 DEBUG_PRINTM(vars.zf);
392 if (m_use_linearkernel) {
393 vars.kzf = (vars.zf.mul2(p_model_alphaf)).sum_over_channels();
394 vars.flag = Track_flags::RESPONSE;
397 vars.flag = Track_flags::CROSS_CORRELATION;
398 gaussian_correlation(vars, vars.zf, this->p_model_xf, this->p_kernel_sigma);
399 DEBUG_PRINTM(p_model_alphaf);
400 DEBUG_PRINTM(vars.kzf);
401 DEBUG_PRINTM(p_model_alphaf * vars.kzf);
402 vars.flag = Track_flags::RESPONSE;
403 vars.kzf = p_model_alphaf * vars.kzf;
407 DEBUG_PRINTM(vars.response);
409 /* target location is at the maximum response. we must take into
410 account the fact that, if the target doesn't move, the peak
411 will appear at the top-left corner, not at the center (this is
412 discussed in the paper). the responses wrap around cyclically. */
415 cv::minMaxLoc(vars.response, &min_val, &vars.max_val, &min_loc, &vars.max_loc);
417 DEBUG_PRINT(vars.max_loc);
419 double weight = scale < 1. ? scale : 1./scale;
420 vars.max_response = vars.max_val*weight;
423 void KCF_Tracker::get_features(cv::Mat & input_rgb, cv::Mat & input_gray, int cx, int cy, int size_x, int size_y, Scale_vars &vars, double scale)
425 int size_x_scaled = floor(size_x*scale);
426 int size_y_scaled = floor(size_y*scale);
428 cv::Mat patch_gray = get_subwindow(input_gray, cx, cy, size_x_scaled, size_y_scaled);
429 cv::Mat patch_rgb = get_subwindow(input_rgb, cx, cy, size_x_scaled, size_y_scaled);
431 //resize to default size
433 //if we downsample use INTER_AREA interpolation
434 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_AREA);
436 cv::resize(patch_gray, patch_gray, cv::Size(size_x, size_y), 0., 0., cv::INTER_LINEAR);
439 // get hog(Histogram of Oriented Gradients) features
440 FHoG::extract(patch_gray, vars, 2, p_cell_size, 9);
442 //get color rgb features (simple r,g,b channels)
443 std::vector<cv::Mat> color_feat;
444 if ((m_use_color || m_use_cnfeat) && input_rgb.channels() == 3) {
445 //resize to default size
447 //if we downsample use INTER_AREA interpolation
448 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_AREA);
450 cv::resize(patch_rgb, patch_rgb, cv::Size(size_x/p_cell_size, size_y/p_cell_size), 0., 0., cv::INTER_LINEAR);
454 if (m_use_color && input_rgb.channels() == 3) {
455 //use rgb color space
456 cv::Mat patch_rgb_norm;
457 patch_rgb.convertTo(patch_rgb_norm, CV_32F, 1. / 255., -0.5);
458 cv::Mat ch1(patch_rgb_norm.size(), CV_32FC1);
459 cv::Mat ch2(patch_rgb_norm.size(), CV_32FC1);
460 cv::Mat ch3(patch_rgb_norm.size(), CV_32FC1);
461 std::vector<cv::Mat> rgb = {ch1, ch2, ch3};
462 cv::split(patch_rgb_norm, rgb);
463 color_feat.insert(color_feat.end(), rgb.begin(), rgb.end());
466 if (m_use_cnfeat && input_rgb.channels() == 3) {
467 std::vector<cv::Mat> cn_feat = CNFeat::extract(patch_rgb);
468 color_feat.insert(color_feat.end(), cn_feat.begin(), cn_feat.end());
471 vars.patch_feats.insert(vars.patch_feats.end(), color_feat.begin(), color_feat.end());
475 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
477 cv::Mat labels(dim2, dim1, CV_32FC1);
478 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
479 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
481 double sigma_s = sigma*sigma;
483 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j){
484 float * row_ptr = labels.ptr<float>(j);
486 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i){
487 row_ptr[i] = std::exp(-0.5 * (y_s + x*x) / sigma_s);//-1/2*e^((y^2+x^2)/sigma^2)
491 //rotate so that 1 is at top-left corner (see KCF paper for explanation)
492 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
493 //sanity check, 1 at top left corner
494 assert(rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
499 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
501 cv::Mat rot_patch(patch.size(), CV_32FC1);
502 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
504 //circular rotate x-axis
506 //move part that does not rotate over the edge
507 cv::Range orig_range(-x_rot, patch.cols);
508 cv::Range rot_range(0, patch.cols - (-x_rot));
509 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
512 orig_range = cv::Range(0, -x_rot);
513 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
514 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
515 }else if (x_rot > 0){
516 //move part that does not rotate over the edge
517 cv::Range orig_range(0, patch.cols - x_rot);
518 cv::Range rot_range(x_rot, patch.cols);
519 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
522 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
523 rot_range = cv::Range(0, x_rot);
524 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
525 }else { //zero rotation
526 //move part that does not rotate over the edge
527 cv::Range orig_range(0, patch.cols);
528 cv::Range rot_range(0, patch.cols);
529 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
532 //circular rotate y-axis
534 //move part that does not rotate over the edge
535 cv::Range orig_range(-y_rot, patch.rows);
536 cv::Range rot_range(0, patch.rows - (-y_rot));
537 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
540 orig_range = cv::Range(0, -y_rot);
541 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
542 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
543 }else if (y_rot > 0){
544 //move part that does not rotate over the edge
545 cv::Range orig_range(0, patch.rows - y_rot);
546 cv::Range rot_range(y_rot, patch.rows);
547 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
550 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
551 rot_range = cv::Range(0, y_rot);
552 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
553 }else { //zero rotation
554 //move part that does not rotate over the edge
555 cv::Range orig_range(0, patch.rows);
556 cv::Range rot_range(0, patch.rows);
557 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
563 //hann window actually (Power-of-cosine windows)
564 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
566 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
567 double N_inv = 1./(static_cast<double>(dim1)-1.);
568 for (int i = 0; i < dim1; ++i)
569 m1.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
570 N_inv = 1./(static_cast<double>(dim2)-1.);
571 for (int i = 0; i < dim2; ++i)
572 m2.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
577 // Returns sub-window of image input centered at [cx, cy] coordinates),
578 // with size [width, height]. If any pixels are outside of the image,
579 // they will replicate the values at the borders.
580 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat & input, int cx, int cy, int width, int height)
584 int x1 = cx - width/2;
585 int y1 = cy - height/2;
586 int x2 = cx + width/2;
587 int y2 = cy + height/2;
590 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
591 patch.create(height, width, input.type());
596 int top = 0, bottom = 0, left = 0, right = 0;
598 //fit to image coordinates, set border extensions;
607 if (x2 >= input.cols) {
608 right = x2 - input.cols + width % 2;
613 if (y2 >= input.rows) {
614 bottom = y2 - input.rows + height % 2;
619 if (x2 - x1 == 0 || y2 - y1 == 0)
620 patch = cv::Mat::zeros(height, width, CV_32FC1);
623 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right, cv::BORDER_REPLICATE);
624 // imshow( "copyMakeBorder", patch);
629 assert(patch.cols == width && patch.rows == height);
634 void KCF_Tracker::gaussian_correlation(struct Scale_vars & vars, const ComplexMat & xf, const ComplexMat & yf, double sigma, bool auto_correlation)
637 xf.sqr_norm(vars.xf_sqr_norm_d);
638 if (!auto_correlation)
639 yf.sqr_norm(vars.yf_sqr_norm_d);
641 xf.sqr_norm(vars.xf_sqr_norm);
642 if (auto_correlation){
643 vars.yf_sqr_norm[0] = vars.xf_sqr_norm[0];
645 yf.sqr_norm(vars.yf_sqr_norm);
648 vars.xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
649 DEBUG_PRINTM(vars.xyf);
652 cuda_gaussian_correlation(fft.inverse_raw(xyf), vars.gauss_corr_res, vars.xf_sqr_norm_d, vars.xf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
654 cuda_gaussian_correlation(fft.inverse_raw(xyf), vars.gauss_corr_res, vars.xf_sqr_norm_d, vars.yf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
656 return fft.forward_raw(vars.gauss_corr_res, xf.n_scales==p_num_scales);
658 //ifft2 and sum over 3rd dimension, we dont care about individual channels
660 DEBUG_PRINTM(vars.ifft2_res);
662 if (xf.channels() != p_num_scales*p_num_of_feats)
663 xy_sum.create(vars.ifft2_res.size(), CV_32FC1);
665 xy_sum.create(vars.ifft2_res.size(), CV_32FC(p_scales.size()));
667 for (int y = 0; y < vars.ifft2_res.rows; ++y) {
668 float * row_ptr = vars.ifft2_res.ptr<float>(y);
669 float * row_ptr_sum = xy_sum.ptr<float>(y);
670 for (int x = 0; x < vars.ifft2_res.cols; ++x) {
671 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
672 row_ptr_sum[(x*xy_sum.channels())+sum_ch] += std::accumulate(row_ptr + x*vars.ifft2_res.channels() + sum_ch*(vars.ifft2_res.channels()/xy_sum.channels()),
673 (row_ptr + x*vars.ifft2_res.channels() + (sum_ch+1)*(vars.ifft2_res.channels()/xy_sum.channels())), 0.f);
677 DEBUG_PRINTM(xy_sum);
679 std::vector<cv::Mat> scales;
680 cv::split(xy_sum,scales);
681 vars.in_all = cv::Mat(scales[0].rows * xf.n_scales, scales[0].cols, CV_32F);
683 float numel_xf_inv = 1.f/(xf.cols * xf.rows * (xf.channels()/xf.n_scales));
684 for (int i = 0; i < xf.n_scales; ++i){
685 cv::Mat in_roi(vars.in_all, cv::Rect(0, i*scales[0].rows, scales[0].cols, scales[0].rows));
686 cv::exp(- 1.f / (sigma * sigma) * cv::max((vars.xf_sqr_norm[i] + vars.yf_sqr_norm[0] - 2 * scales[i]) * numel_xf_inv, 0), in_roi);
687 DEBUG_PRINTM(in_roi);
690 DEBUG_PRINTM(vars.in_all );
696 float get_response_circular(cv::Point2i & pt, cv::Mat & response)
701 x = response.cols + x;
703 y = response.rows + y;
704 if (x >= response.cols)
705 x = x - response.cols;
706 if (y >= response.rows)
707 y = y - response.rows;
709 return response.at<float>(y,x);
712 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point & max_loc, cv::Mat & response)
714 //find neighbourhood of max_loc (response is circular)
718 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);
719 cv::Point2i p4(max_loc.x-1, max_loc.y), p5(max_loc.x+1, max_loc.y);
720 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);
723 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
724 cv::Mat A = (cv::Mat_<float>(9, 6) <<
725 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
726 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
727 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
728 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
729 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
730 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
731 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
732 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
733 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);
734 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
735 get_response_circular(p1, response),
736 get_response_circular(p2, response),
737 get_response_circular(p3, response),
738 get_response_circular(p4, response),
739 get_response_circular(p5, response),
740 get_response_circular(p6, response),
741 get_response_circular(p7, response),
742 get_response_circular(p8, response),
743 get_response_circular(max_loc, response));
746 cv::solve(A, fval, x, cv::DECOMP_SVD);
748 double a = x.at<float>(0), b = x.at<float>(1), c = x.at<float>(2),
749 d = x.at<float>(3), e = x.at<float>(4);
751 cv::Point2f sub_peak(max_loc.x, max_loc.y);
752 if (b > 0 || b < 0) {
753 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
754 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
760 double KCF_Tracker::sub_grid_scale(int index)
763 if (index < 0 || index > (int)p_scales.size()-1) {
764 // interpolate from all values
765 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
766 A.create(p_scales.size(), 3, CV_32FC1);
767 fval.create(p_scales.size(), 1, CV_32FC1);
768 for (size_t i = 0; i < p_scales.size(); ++i) {
769 A.at<float>(i, 0) = p_scales[i] * p_scales[i];
770 A.at<float>(i, 1) = p_scales[i];
771 A.at<float>(i, 2) = 1;
772 fval.at<float>(i) = scale_vars[i].max_response;
775 //only from neighbours
776 if (index == 0 || index == (int)p_scales.size()-1)
777 return p_scales[index];
779 A = (cv::Mat_<float>(3, 3) <<
780 p_scales[index-1] * p_scales[index-1], p_scales[index-1], 1,
781 p_scales[index] * p_scales[index], p_scales[index], 1,
782 p_scales[index+1] * p_scales[index+1], p_scales[index+1], 1);
783 fval = (cv::Mat_<float>(3, 1) << scale_vars[index-1].max_response, scale_vars[index].max_response, scale_vars[index+1].max_response);
787 cv::solve(A, fval, x, cv::DECOMP_SVD);
788 double a = x.at<float>(0), b = x.at<float>(1);
789 double scale = p_scales[index];
791 scale = -b / (2 * a);