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()
38 CudaSafeCall(cudaFreeHost(xf_sqr_norm));
39 CudaSafeCall(cudaFreeHost(yf_sqr_norm));
40 CudaSafeCall(cudaFree(gauss_corr_res));
48 void KCF_Tracker::init(cv::Mat &img, const cv::Rect & bbox, int fit_size_x, int fit_size_y)
50 //check boundary, enforce min size
51 double x1 = bbox.x, x2 = bbox.x + bbox.width, y1 = bbox.y, y2 = bbox.y + bbox.height;
53 if (x2 > img.cols-1) x2 = img.cols - 1;
55 if (y2 > img.rows-1) y2 = img.rows - 1;
57 if (x2-x1 < 2*p_cell_size) {
58 double diff = (2*p_cell_size -x2+x1)/2.;
59 if (x1 - diff >= 0 && x2 + diff < img.cols){
62 } else if (x1 - 2*diff >= 0) {
68 if (y2-y1 < 2*p_cell_size) {
69 double diff = (2*p_cell_size -y2+y1)/2.;
70 if (y1 - diff >= 0 && y2 + diff < img.rows){
73 } else if (y1 - 2*diff >= 0) {
82 p_pose.cx = x1 + p_pose.w/2.;
83 p_pose.cy = y1 + p_pose.h /2.;
86 cv::Mat input_gray, input_rgb = img.clone();
87 if (img.channels() == 3){
88 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
89 input_gray.convertTo(input_gray, CV_32FC1);
91 img.convertTo(input_gray, CV_32FC1);
93 // don't need too large image
94 if (p_pose.w * p_pose.h > 100.*100. && (fit_size_x == -1 || fit_size_y == -1)) {
95 std::cout << "resizing image by factor of " << 1/p_downscale_factor << std::endl;
96 p_resize_image = true;
97 p_pose.scale(p_downscale_factor);
98 cv::resize(input_gray, input_gray, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
99 cv::resize(input_rgb, input_rgb, cv::Size(0,0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
100 } else if (!(fit_size_x == -1 && fit_size_y == -1)) {
101 if (fit_size_x%p_cell_size != 0 || fit_size_y%p_cell_size != 0) {
102 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;;
103 std::exit(EXIT_FAILURE);
106 if (( tmp = (p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_x)
107 p_scale_factor_x = fit_size_x/tmp;
108 if (( tmp = (p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size ) != fit_size_y)
109 p_scale_factor_y = fit_size_y/tmp;
110 std::cout << "resizing image horizontaly by factor of " << p_scale_factor_x
111 << " and verticaly by factor of " << p_scale_factor_y << std::endl;
113 p_pose.scale_x(p_scale_factor_x);
114 p_pose.scale_y(p_scale_factor_y);
115 if (p_scale_factor_x != 1 && p_scale_factor_y != 1) {
116 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
117 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
118 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
120 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
121 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
126 //compute win size + fit to fhog cell size
127 p_windows_size[0] = round(p_pose.w * (1. + p_padding) / p_cell_size) * p_cell_size;
128 p_windows_size[1] = round(p_pose.h * (1. + p_padding) / p_cell_size) * p_cell_size;
132 for (int i = -p_num_scales/2; i <= p_num_scales/2; ++i)
133 p_scales.push_back(std::pow(p_scale_step, i));
135 p_scales.push_back(1.);
139 if (p_windows_size[1]/p_cell_size*(p_windows_size[0]/p_cell_size/2+1) > 1024) {
140 std::cerr << "Window after forward FFT is too big for CUDA kernels. Plese use -f to set "
141 "the window dimensions so its size is less or equal to " << 1024*p_cell_size*p_cell_size*2+1 <<
142 " pixels . Currently the size of the window is: " << p_windows_size[0] << "x" << p_windows_size[1] <<
143 " which is " << p_windows_size[0]*p_windows_size[1] << " pixels. " << std::endl;
144 std::exit(EXIT_FAILURE);
147 if (m_use_linearkernel){
148 std::cerr << "cuFFT supports only Gaussian kernel." << std::endl;
149 std::exit(EXIT_FAILURE);
151 cudaSetDeviceFlags(cudaDeviceMapHost);
152 CudaSafeCall(cudaHostAlloc((void**)&xf_sqr_norm, p_num_scales*sizeof(float), cudaHostAllocMapped));
153 CudaSafeCall(cudaHostGetDevicePointer((void**)&xf_sqr_norm_d, (void*)xf_sqr_norm, 0));
155 CudaSafeCall(cudaHostAlloc((void**)&yf_sqr_norm, sizeof(float), cudaHostAllocMapped));
156 CudaSafeCall(cudaHostGetDevicePointer((void**)&yf_sqr_norm_d, (void*)yf_sqr_norm, 0));
158 xf_sqr_norm = (float*) malloc(p_num_scales*sizeof(float));
159 yf_sqr_norm = (float*) malloc(sizeof(float));
162 p_current_scale = 1.;
164 double min_size_ratio = std::max(5.*p_cell_size/p_windows_size[0], 5.*p_cell_size/p_windows_size[1]);
165 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]);
166 p_min_max_scale[0] = std::pow(p_scale_step, std::ceil(std::log(min_size_ratio) / log(p_scale_step)));
167 p_min_max_scale[1] = std::pow(p_scale_step, std::floor(std::log(max_size_ratio) / log(p_scale_step)));
169 std::cout << "init: img size " << img.cols << " " << img.rows << std::endl;
170 std::cout << "init: win size. " << p_windows_size[0] << " " << p_windows_size[1] << std::endl;
171 std::cout << "init: min max scales factors: " << p_min_max_scale[0] << " " << p_min_max_scale[1] << std::endl;
173 p_output_sigma = std::sqrt(p_pose.w*p_pose.h) * p_output_sigma_factor / static_cast<double>(p_cell_size);
175 //window weights, i.e. labels
177 if(m_use_color) p_num_of_feats += 3;
178 if(m_use_cnfeat) p_num_of_feats += 10;
179 p_roi_width = p_windows_size[0]/p_cell_size;
180 p_roi_height = p_windows_size[1]/p_cell_size;
182 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);
183 p_yf = fft.forward(gaussian_shaped_labels(p_output_sigma, p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
184 fft.set_window(cosine_window_function(p_windows_size[0]/p_cell_size, p_windows_size[1]/p_cell_size));
187 CudaSafeCall(cudaMalloc((void**)&gauss_corr_res, (p_windows_size[0]/p_cell_size)*(p_windows_size[1]/p_cell_size)*p_num_scales*sizeof(float)));
189 //obtain a sub-window for training initial model
190 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]);
191 p_model_xf = fft.forward_window(path_feat);
192 DEBUG_PRINTM(p_model_xf);
194 if (m_use_linearkernel) {
195 ComplexMat xfconj = p_model_xf.conj();
196 p_model_alphaf_num = xfconj.mul(p_yf);
197 p_model_alphaf_den = (p_model_xf * xfconj);
199 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
200 ComplexMat kf = gaussian_correlation(p_model_xf, p_model_xf, p_kernel_sigma, true);
202 p_model_alphaf_num = p_yf * kf;
203 DEBUG_PRINTM(p_model_alphaf_num);
204 p_model_alphaf_den = kf * (kf + p_lambda);
205 DEBUG_PRINTM(p_model_alphaf_den);
207 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
208 DEBUG_PRINTM(p_model_alphaf);
209 // p_model_alphaf = p_yf / (kf + p_lambda); //equation for fast training
212 void KCF_Tracker::setTrackerPose(BBox_c &bbox, cv::Mat & img, int fit_size_x, int fit_size_y)
214 init(img, bbox.get_rect(), fit_size_x, fit_size_y);
217 void KCF_Tracker::updateTrackerPosition(BBox_c &bbox)
219 if (p_resize_image) {
221 tmp.scale(p_downscale_factor);
224 } else if (p_fit_to_pw2) {
226 tmp.scale_x(p_scale_factor_x);
227 tmp.scale_y(p_scale_factor_y);
236 BBox_c KCF_Tracker::getBBox()
239 tmp.w *= p_current_scale;
240 tmp.h *= p_current_scale;
243 tmp.scale(1/p_downscale_factor);
245 tmp.scale_x(1/p_scale_factor_x);
246 tmp.scale_y(1/p_scale_factor_y);
252 void KCF_Tracker::track(cv::Mat &img)
255 std::cout << "NEW FRAME" << '\n';
256 cv::Mat input_gray, input_rgb = img.clone();
257 if (img.channels() == 3){
258 cv::cvtColor(img, input_gray, CV_BGR2GRAY);
259 input_gray.convertTo(input_gray, CV_32FC1);
261 img.convertTo(input_gray, CV_32FC1);
263 // don't need too large image
264 if (p_resize_image) {
265 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
266 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_downscale_factor, p_downscale_factor, cv::INTER_AREA);
267 } else if (p_fit_to_pw2 && p_scale_factor_x != 1 && p_scale_factor_y != 1) {
268 if (p_scale_factor_x < 1 && p_scale_factor_y < 1) {
269 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
270 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_AREA);
272 cv::resize(input_gray, input_gray, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
273 cv::resize(input_rgb, input_rgb, cv::Size(0, 0), p_scale_factor_x, p_scale_factor_y, cv::INTER_LINEAR);
278 std::vector<cv::Mat> patch_feat;
279 double max_response = -1.;
280 cv::Mat max_response_map;
281 cv::Point2i max_response_pt;
283 std::vector<double> scale_responses;
285 if (m_use_multithreading){
286 std::vector<std::future<cv::Mat>> async_res(p_scales.size());
287 for (size_t i = 0; i < p_scales.size(); ++i) {
288 async_res[i] = std::async(std::launch::async,
289 [this, &input_gray, &input_rgb, i]() -> cv::Mat
291 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],
292 this->p_windows_size[1], this->p_current_scale * this->p_scales[i]);
293 ComplexMat zf = fft.forward_window(patch_feat_async);
294 if (m_use_linearkernel)
295 return fft.inverse((p_model_alphaf * zf).sum_over_channels());
297 ComplexMat kzf = gaussian_correlation(zf, this->p_model_xf, this->p_kernel_sigma);
298 return fft.inverse(this->p_model_alphaf * kzf);
303 for (size_t i = 0; i < p_scales.size(); ++i) {
306 cv::Mat response = async_res[i].get();
308 double min_val, max_val;
309 cv::Point2i min_loc, max_loc;
310 cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
312 double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
313 if (max_val*weight > max_response) {
314 max_response = max_val*weight;
315 max_response_map = response;
316 max_response_pt = max_loc;
319 scale_responses.push_back(max_val*weight);
321 } else if (m_use_big_batch){
322 #pragma omp parallel for ordered
323 for (size_t i = 0; i < p_scales.size(); ++i) {
324 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]);
326 patch_feat.insert(std::end(patch_feat), std::begin(tmp), std::end(tmp));
328 ComplexMat zf = fft.forward_window(patch_feat);
332 if (m_use_linearkernel)
333 response = fft.inverse((zf.mul2(p_model_alphaf)).sum_over_channels());
335 ComplexMat kzf = gaussian_correlation(zf, p_model_xf, p_kernel_sigma);
336 DEBUG_PRINTM(p_model_alphaf);
338 response = fft.inverse(kzf.mul(p_model_alphaf));
340 DEBUG_PRINTM(response);
341 std::vector<cv::Mat> scales;
342 cv::split(response,scales);
344 /* target location is at the maximum response. we must take into
345 account the fact that, if the target doesn't move, the peak
346 will appear at the top-left corner, not at the center (this is
347 discussed in the paper). the responses wrap around cyclically. */
348 for (size_t i = 0; i < p_scales.size(); ++i) {
349 double min_val, max_val;
350 cv::Point2i min_loc, max_loc;
351 cv::minMaxLoc(scales[i], &min_val, &max_val, &min_loc, &max_loc);
352 DEBUG_PRINT(max_loc);
354 double weight = p_scales[i] < 1. ? p_scales[i] : 1./p_scales[i];
356 if (max_val*weight > max_response) {
357 max_response = max_val*weight;
358 max_response_map = scales[i];
359 max_response_pt = max_loc;
362 scale_responses.push_back(max_val*weight);
365 #pragma omp parallel for ordered private(patch_feat) schedule(dynamic)
366 for (size_t i = 0; i < p_scales.size(); ++i) {
367 patch_feat = get_features(input_rgb, input_gray, this->p_pose.cx, this->p_pose.cy, this->p_windows_size[0], this->p_windows_size[1], this->p_current_scale * this->p_scales[i]);
368 ComplexMat zf = fft.forward_window(patch_feat);
371 if (m_use_linearkernel)
372 response = fft.inverse((p_model_alphaf * zf).sum_over_channels());
374 ComplexMat kzf = gaussian_correlation(zf, this->p_model_xf, this->p_kernel_sigma);
375 DEBUG_PRINTM(p_model_alphaf);
377 DEBUG_PRINTM(p_model_alphaf * kzf);
378 response = fft.inverse(this->p_model_alphaf * kzf);
380 DEBUG_PRINTM(response);
382 /* target location is at the maximum response. we must take into
383 account the fact that, if the target doesn't move, the peak
384 will appear at the top-left corner, not at the center (this is
385 discussed in the paper). the responses wrap around cyclically. */
386 double min_val, max_val;
387 cv::Point2i min_loc, max_loc;
388 cv::minMaxLoc(response, &min_val, &max_val, &min_loc, &max_loc);
389 DEBUG_PRINT(max_loc);
391 double weight = this->p_scales[i] < 1. ? this->p_scales[i] : 1./this->p_scales[i];
394 if (max_val*weight > max_response) {
395 max_response = max_val*weight;
396 max_response_map = response;
397 max_response_pt = max_loc;
402 scale_responses.push_back(max_val*weight);
405 DEBUG_PRINTM(max_response_map);
406 DEBUG_PRINT(max_response_pt);
407 //sub pixel quadratic interpolation from neighbours
408 if (max_response_pt.y > max_response_map.rows / 2) //wrap around to negative half-space of vertical axis
409 max_response_pt.y = max_response_pt.y - max_response_map.rows;
410 if (max_response_pt.x > max_response_map.cols / 2) //same for horizontal axis
411 max_response_pt.x = max_response_pt.x - max_response_map.cols;
413 cv::Point2f new_location(max_response_pt.x, max_response_pt.y);
414 DEBUG_PRINT(new_location);
416 if (m_use_subpixel_localization)
417 new_location = sub_pixel_peak(max_response_pt, max_response_map);
418 DEBUG_PRINT(new_location);
420 p_pose.cx += p_current_scale*p_cell_size*new_location.x;
421 p_pose.cy += p_current_scale*p_cell_size*new_location.y;
423 if (p_pose.cx < 0) p_pose.cx = 0;
424 if (p_pose.cx > (img.cols*p_scale_factor_x)-1) p_pose.cx = (img.cols*p_scale_factor_x)-1;
425 if (p_pose.cy < 0) p_pose.cy = 0;
426 if (p_pose.cy > (img.rows*p_scale_factor_y)-1) p_pose.cy = (img.rows*p_scale_factor_y)-1;
428 if (p_pose.cx < 0) p_pose.cx = 0;
429 if (p_pose.cx > img.cols-1) p_pose.cx = img.cols-1;
430 if (p_pose.cy < 0) p_pose.cy = 0;
431 if (p_pose.cy > img.rows-1) p_pose.cy = img.rows-1;
434 //sub grid scale interpolation
435 double new_scale = p_scales[scale_index];
436 if (m_use_subgrid_scale)
437 new_scale = sub_grid_scale(scale_responses, scale_index);
439 p_current_scale *= new_scale;
441 if (p_current_scale < p_min_max_scale[0])
442 p_current_scale = p_min_max_scale[0];
443 if (p_current_scale > p_min_max_scale[1])
444 p_current_scale = p_min_max_scale[1];
445 //obtain a subwindow for training at newly estimated target position
446 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);
447 ComplexMat xf = fft.forward_window(patch_feat);
449 //subsequent frames, interpolate model
450 p_model_xf = p_model_xf * (1. - p_interp_factor) + xf * p_interp_factor;
452 ComplexMat alphaf_num, alphaf_den;
454 if (m_use_linearkernel) {
455 ComplexMat xfconj = xf.conj();
456 alphaf_num = xfconj.mul(p_yf);
457 alphaf_den = (xf * xfconj);
459 //Kernel Ridge Regression, calculate alphas (in Fourier domain)
460 ComplexMat kf = gaussian_correlation(xf, xf, p_kernel_sigma, true);
461 // ComplexMat alphaf = p_yf / (kf + p_lambda); //equation for fast training
462 // p_model_alphaf = p_model_alphaf * (1. - p_interp_factor) + alphaf * p_interp_factor;
463 alphaf_num = p_yf * kf;
464 alphaf_den = kf * (kf + p_lambda);
467 p_model_alphaf_num = p_model_alphaf_num * (1. - p_interp_factor) + alphaf_num * p_interp_factor;
468 p_model_alphaf_den = p_model_alphaf_den * (1. - p_interp_factor) + alphaf_den * p_interp_factor;
469 p_model_alphaf = p_model_alphaf_num / p_model_alphaf_den;
472 // ****************************************************************************
474 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)
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());
526 cv::Mat KCF_Tracker::gaussian_shaped_labels(double sigma, int dim1, int dim2)
528 cv::Mat labels(dim2, dim1, CV_32FC1);
529 int range_y[2] = {-dim2 / 2, dim2 - dim2 / 2};
530 int range_x[2] = {-dim1 / 2, dim1 - dim1 / 2};
532 double sigma_s = sigma*sigma;
534 for (int y = range_y[0], j = 0; y < range_y[1]; ++y, ++j){
535 float * row_ptr = labels.ptr<float>(j);
537 for (int x = range_x[0], i = 0; x < range_x[1]; ++x, ++i){
538 row_ptr[i] = std::exp(-0.5 * (y_s + x*x) / sigma_s);//-1/2*e^((y^2+x^2)/sigma^2)
542 //rotate so that 1 is at top-left corner (see KCF paper for explanation)
543 cv::Mat rot_labels = circshift(labels, range_x[0], range_y[0]);
544 //sanity check, 1 at top left corner
545 assert(rot_labels.at<float>(0,0) >= 1.f - 1e-10f);
550 cv::Mat KCF_Tracker::circshift(const cv::Mat &patch, int x_rot, int y_rot)
552 cv::Mat rot_patch(patch.size(), CV_32FC1);
553 cv::Mat tmp_x_rot(patch.size(), CV_32FC1);
555 //circular rotate x-axis
557 //move part that does not rotate over the edge
558 cv::Range orig_range(-x_rot, patch.cols);
559 cv::Range rot_range(0, patch.cols - (-x_rot));
560 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
563 orig_range = cv::Range(0, -x_rot);
564 rot_range = cv::Range(patch.cols - (-x_rot), patch.cols);
565 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
566 }else if (x_rot > 0){
567 //move part that does not rotate over the edge
568 cv::Range orig_range(0, patch.cols - x_rot);
569 cv::Range rot_range(x_rot, patch.cols);
570 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
573 orig_range = cv::Range(patch.cols - x_rot, patch.cols);
574 rot_range = cv::Range(0, x_rot);
575 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
576 }else { //zero rotation
577 //move part that does not rotate over the edge
578 cv::Range orig_range(0, patch.cols);
579 cv::Range rot_range(0, patch.cols);
580 patch(cv::Range::all(), orig_range).copyTo(tmp_x_rot(cv::Range::all(), rot_range));
583 //circular rotate y-axis
585 //move part that does not rotate over the edge
586 cv::Range orig_range(-y_rot, patch.rows);
587 cv::Range rot_range(0, patch.rows - (-y_rot));
588 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
591 orig_range = cv::Range(0, -y_rot);
592 rot_range = cv::Range(patch.rows - (-y_rot), patch.rows);
593 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
594 }else if (y_rot > 0){
595 //move part that does not rotate over the edge
596 cv::Range orig_range(0, patch.rows - y_rot);
597 cv::Range rot_range(y_rot, patch.rows);
598 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
601 orig_range = cv::Range(patch.rows - y_rot, patch.rows);
602 rot_range = cv::Range(0, y_rot);
603 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
604 }else { //zero rotation
605 //move part that does not rotate over the edge
606 cv::Range orig_range(0, patch.rows);
607 cv::Range rot_range(0, patch.rows);
608 tmp_x_rot(orig_range, cv::Range::all()).copyTo(rot_patch(rot_range, cv::Range::all()));
614 //hann window actually (Power-of-cosine windows)
615 cv::Mat KCF_Tracker::cosine_window_function(int dim1, int dim2)
617 cv::Mat m1(1, dim1, CV_32FC1), m2(dim2, 1, CV_32FC1);
618 double N_inv = 1./(static_cast<double>(dim1)-1.);
619 for (int i = 0; i < dim1; ++i)
620 m1.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
621 N_inv = 1./(static_cast<double>(dim2)-1.);
622 for (int i = 0; i < dim2; ++i)
623 m2.at<float>(i) = 0.5*(1. - std::cos(2. * CV_PI * static_cast<double>(i) * N_inv));
628 // Returns sub-window of image input centered at [cx, cy] coordinates),
629 // with size [width, height]. If any pixels are outside of the image,
630 // they will replicate the values at the borders.
631 cv::Mat KCF_Tracker::get_subwindow(const cv::Mat &input, int cx, int cy, int width, int height)
635 int x1 = cx - width/2;
636 int y1 = cy - height/2;
637 int x2 = cx + width/2;
638 int y2 = cy + height/2;
641 if (x1 >= input.cols || y1 >= input.rows || x2 < 0 || y2 < 0) {
642 patch.create(height, width, input.type());
647 int top = 0, bottom = 0, left = 0, right = 0;
649 //fit to image coordinates, set border extensions;
658 if (x2 >= input.cols) {
659 right = x2 - input.cols + width % 2;
664 if (y2 >= input.rows) {
665 bottom = y2 - input.rows + height % 2;
670 if (x2 - x1 == 0 || y2 - y1 == 0)
671 patch = cv::Mat::zeros(height, width, CV_32FC1);
674 cv::copyMakeBorder(input(cv::Range(y1, y2), cv::Range(x1, x2)), patch, top, bottom, left, right, cv::BORDER_REPLICATE);
675 // imshow( "copyMakeBorder", patch);
680 assert(patch.cols == width && patch.rows == height);
685 ComplexMat KCF_Tracker::gaussian_correlation(const ComplexMat &xf, const ComplexMat &yf, double sigma, bool auto_correlation)
689 xf.sqr_norm(xf_sqr_norm_d);
690 if (!auto_correlation)
691 yf.sqr_norm(yf_sqr_norm_d);
693 xf.sqr_norm(xf_sqr_norm);
694 if (auto_correlation){
695 yf_sqr_norm[0] = xf_sqr_norm[0];
697 yf.sqr_norm(yf_sqr_norm);
701 float xf_sqr_norm = xf.sqr_norm();
702 float yf_sqr_norm =auto_correlation ? xf_sqr_norm : yf.sqr_norm();
705 xyf = auto_correlation ? xf.sqr_mag() : xf.mul2(yf.conj());
709 cuda_gaussian_correlation(fft.inverse_raw(xyf), gauss_corr_res, xf_sqr_norm_d, xf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
711 cuda_gaussian_correlation(fft.inverse_raw(xyf), gauss_corr_res, xf_sqr_norm_d, yf_sqr_norm_d, sigma, xf.n_channels, xf.n_scales, p_roi_height, p_roi_width);
713 return fft.forward_raw(gauss_corr_res, xf.n_scales==p_num_scales);
715 //ifft2 and sum over 3rd dimension, we dont care about individual channels
716 cv::Mat ifft2_res = fft.inverse(xyf);
717 DEBUG_PRINTM(ifft2_res);
719 if (xf.channels() != p_num_scales*p_num_of_feats)
720 xy_sum.create(ifft2_res.size(), CV_32FC1);
722 xy_sum.create(ifft2_res.size(), CV_32FC(p_scales.size()));
724 for (int y = 0; y < ifft2_res.rows; ++y) {
725 float * row_ptr = ifft2_res.ptr<float>(y);
726 float * row_ptr_sum = xy_sum.ptr<float>(y);
727 for (int x = 0; x < ifft2_res.cols; ++x) {
728 for (int sum_ch = 0; sum_ch < xy_sum.channels(); ++sum_ch) {
729 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()), (row_ptr + x*ifft2_res.channels() + (sum_ch+1)*(ifft2_res.channels()/xy_sum.channels())), 0.f);
733 DEBUG_PRINTM(xy_sum);
735 std::vector<cv::Mat> scales;
736 cv::split(xy_sum,scales);
737 cv::Mat in_all(scales[0].rows * xf.n_scales, scales[0].cols, CV_32F);
739 float numel_xf_inv = 1.f/(xf.cols * xf.rows * (xf.channels()/xf.n_scales));
741 for (int i = 0; i < xf.n_scales; ++i){
742 cv::Mat in_roi(in_all, cv::Rect(0, i*scales[0].rows, scales[0].cols, scales[0].rows));
743 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);
744 DEBUG_PRINTM(in_roi);
747 cv::exp(- 1.f / (sigma * sigma) * cv::max((xf_sqr_norm + yf_sqr_norm - 2 * xy_sum) * numel_xf_inv, 0), in_all);
750 DEBUG_PRINTM(in_all);
751 return fft.forward(in_all);
755 float get_response_circular(cv::Point2i & pt, cv::Mat & response)
760 x = response.cols + x;
762 y = response.rows + y;
763 if (x >= response.cols)
764 x = x - response.cols;
765 if (y >= response.rows)
766 y = y - response.rows;
768 return response.at<float>(y,x);
771 cv::Point2f KCF_Tracker::sub_pixel_peak(cv::Point & max_loc, cv::Mat & response)
773 //find neighbourhood of max_loc (response is circular)
777 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);
778 cv::Point2i p4(max_loc.x-1, max_loc.y), p5(max_loc.x+1, max_loc.y);
779 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);
782 // fit 2d quadratic function f(x, y) = a*x^2 + b*x*y + c*y^2 + d*x + e*y + f
783 cv::Mat A = (cv::Mat_<float>(9, 6) <<
784 p1.x*p1.x, p1.x*p1.y, p1.y*p1.y, p1.x, p1.y, 1.f,
785 p2.x*p2.x, p2.x*p2.y, p2.y*p2.y, p2.x, p2.y, 1.f,
786 p3.x*p3.x, p3.x*p3.y, p3.y*p3.y, p3.x, p3.y, 1.f,
787 p4.x*p4.x, p4.x*p4.y, p4.y*p4.y, p4.x, p4.y, 1.f,
788 p5.x*p5.x, p5.x*p5.y, p5.y*p5.y, p5.x, p5.y, 1.f,
789 p6.x*p6.x, p6.x*p6.y, p6.y*p6.y, p6.x, p6.y, 1.f,
790 p7.x*p7.x, p7.x*p7.y, p7.y*p7.y, p7.x, p7.y, 1.f,
791 p8.x*p8.x, p8.x*p8.y, p8.y*p8.y, p8.x, p8.y, 1.f,
792 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);
793 cv::Mat fval = (cv::Mat_<float>(9, 1) <<
794 get_response_circular(p1, response),
795 get_response_circular(p2, response),
796 get_response_circular(p3, response),
797 get_response_circular(p4, response),
798 get_response_circular(p5, response),
799 get_response_circular(p6, response),
800 get_response_circular(p7, response),
801 get_response_circular(p8, response),
802 get_response_circular(max_loc, response));
805 cv::solve(A, fval, x, cv::DECOMP_SVD);
807 double a = x.at<float>(0), b = x.at<float>(1), c = x.at<float>(2),
808 d = x.at<float>(3), e = x.at<float>(4);
810 cv::Point2f sub_peak(max_loc.x, max_loc.y);
811 if (b > 0 || b < 0) {
812 sub_peak.y = ((2.f * a * e) / b - d) / (b - (4 * a * c) / b);
813 sub_peak.x = (-2 * c * sub_peak.y - e) / b;
819 double KCF_Tracker::sub_grid_scale(std::vector<double> & responses, int index)
822 if (index < 0 || index > (int)p_scales.size()-1) {
823 // interpolate from all values
824 // fit 1d quadratic function f(x) = a*x^2 + b*x + c
825 A.create(p_scales.size(), 3, CV_32FC1);
826 fval.create(p_scales.size(), 1, CV_32FC1);
827 for (size_t i = 0; i < p_scales.size(); ++i) {
828 A.at<float>(i, 0) = p_scales[i] * p_scales[i];
829 A.at<float>(i, 1) = p_scales[i];
830 A.at<float>(i, 2) = 1;
831 fval.at<float>(i) = responses[i];
834 //only from neighbours
835 if (index == 0 || index == (int)p_scales.size()-1)
836 return p_scales[index];
838 A = (cv::Mat_<float>(3, 3) <<
839 p_scales[index-1] * p_scales[index-1], p_scales[index-1], 1,
840 p_scales[index] * p_scales[index], p_scales[index], 1,
841 p_scales[index+1] * p_scales[index+1], p_scales[index+1], 1);
842 fval = (cv::Mat_<float>(3, 1) << responses[index-1], responses[index], responses[index+1]);
846 cv::solve(A, fval, x, cv::DECOMP_SVD);
847 double a = x.at<float>(0), b = x.at<float>(1);
848 double scale = p_scales[index];
850 scale = -b / (2 * a);