diff options
Diffstat (limited to 'tracker-pt/module/point_extractor.cpp')
| -rw-r--r-- | tracker-pt/module/point_extractor.cpp | 178 |
1 files changed, 125 insertions, 53 deletions
diff --git a/tracker-pt/module/point_extractor.cpp b/tracker-pt/module/point_extractor.cpp index 7ad9925a..3329fafc 100644 --- a/tracker-pt/module/point_extractor.cpp +++ b/tracker-pt/module/point_extractor.cpp @@ -9,9 +9,11 @@ #include "point_extractor.h" #include "point_tracker.h" #include "frame.hpp" - #include "cv/numeric.hpp" #include "compat/math.hpp" +#include "compat/math-imports.hpp" + +#include <opencv2/imgproc.hpp> #undef PREVIEW //#define PREVIEW @@ -33,7 +35,7 @@ using namespace numeric_types; /* http://en.wikipedia.org/wiki/Mean-shift -In this application the idea, is to eliminate any bias of the point estimate +In this application the idea, is to eliminate any bias of the point estimate which is introduced by the rather arbitrary thresholded area. One must recognize that the thresholded area can only move in one pixel increments since it is binary. Thus, its center of mass might make "jumps" as pixels are added/removed @@ -42,9 +44,9 @@ With mean-shift, a moving "window" or kernel is multiplied with the gray-scale image, and the COM is calculated of the result. This is iterated where the kernel center is set the previously computed COM. Thus, peaks in the image intensity distribution "pull" the kernel towards themselves. Eventually it stops moving, i.e. -then the computed COM coincides with the kernel center. We hope that the +then the computed COM coincides with the kernel center. We hope that the corresponding location is a good candidate for the extracted point. -The idea similar to the window scaling suggested in Berglund et al. "Fast, bias-free +The idea similar to the window scaling suggested in Berglund et al. "Fast, bias-free algorithm for tracking single particles with variable size and shape." (2008). */ static vec2 MeanShiftIteration(const cv::Mat1b &frame_gray, const vec2 ¤t_center, f filter_width) @@ -87,21 +89,16 @@ PointExtractor::PointExtractor(const QString& module_name) : s(module_name) void PointExtractor::ensure_channel_buffers(const cv::Mat& orig_frame) { - if (ch[0].rows != orig_frame.rows || ch[0].cols != orig_frame.cols) - for (cv::Mat1b& x : ch) - x = cv::Mat1b(orig_frame.rows, orig_frame.cols); + for (cv::Mat1b& x : ch) + x.create(orig_frame.rows, orig_frame.cols); } void PointExtractor::ensure_buffers(const cv::Mat& frame) { const int W = frame.cols, H = frame.rows; - if (frame_gray.rows != W || frame_gray.cols != H) - { - frame_gray = cv::Mat1b(H, W); - frame_bin = cv::Mat1b(H, W); - frame_gray_unmasked = cv::Mat1b(H, W); - } + frame_gray.create(H, W); + frame_bin.create(H, W); } void PointExtractor::extract_single_channel(const cv::Mat& orig_frame, int idx, cv::Mat1b& dest) @@ -115,8 +112,43 @@ void PointExtractor::extract_single_channel(const cv::Mat& orig_frame, int idx, cv::mixChannels(&orig_frame, 1, &dest, 1, from_to, 1); } +void PointExtractor::filter_single_channel(const cv::Mat& orig_frame, float r, float g, float b, bool overexp, cv::Mat1b& dest) +{ + ensure_channel_buffers(orig_frame); + + // just filter for colour or also include overexposed regions? + if (!overexp) + cv::transform(orig_frame, dest, cv::Mat(cv::Matx13f(b, g, r))); + else + { + for (int i = 0; i < orig_frame.rows; i++) + { + cv::Vec3b const* const __restrict orig_ptr = orig_frame.ptr<cv::Vec3b>(i); + uint8_t* const __restrict dest_ptr = dest.ptr(i); + for (int j = 0; j < orig_frame.cols; j++) + { + // get the intensity of the key color (i.e. +ve coefficients) + uchar blue = orig_ptr[j][0], green = orig_ptr[j][1], red = orig_ptr[j][2]; + float key = std::max(b, 0.0f) * blue + std::max(g, 0.0f) * green + std::max(r, 0.0f) * red; + // get the intensity of the non-key color (i.e. -ve coefficients) + float nonkey = std::max(-b, 0.0f) * blue + std::max(-g, 0.0f) * green + std::max(-r, 0.0f) * red; + // the result is key color minus non-key color inversely weighted by key colour intensity + dest_ptr[j] = std::max(0.0f, std::min(255.0f, key - (255.0f - key) / 255.0f * nonkey)); + } + } + } +} + void PointExtractor::color_to_grayscale(const cv::Mat& frame, cv::Mat1b& output) { + if (frame.channels() == 1) + { + output.create(frame.rows, frame.cols); + frame.copyTo(output); + return; + } + + const float half_chr_key_str = *s.chroma_key_strength * 0.5; switch (s.blob_color) { case pt_color_green_only: @@ -134,18 +166,44 @@ void PointExtractor::color_to_grayscale(const cv::Mat& frame, cv::Mat1b& output) extract_single_channel(frame, 2, output); break; } - case pt_color_average: + case pt_color_red_chromakey: + { + filter_single_channel(frame, 1, -half_chr_key_str, -half_chr_key_str, s.chroma_key_overexposed, output); + break; + } + case pt_color_green_chromakey: + { + filter_single_channel(frame, -half_chr_key_str, 1, -half_chr_key_str, s.chroma_key_overexposed, output); + break; + } + case pt_color_blue_chromakey: + { + filter_single_channel(frame, -half_chr_key_str, -half_chr_key_str, 1, s.chroma_key_overexposed, output); + break; + } + case pt_color_cyan_chromakey: + { + filter_single_channel(frame, -*s.chroma_key_strength, 0.5, 0.5, s.chroma_key_overexposed, output); + break; + } + case pt_color_yellow_chromakey: + { + filter_single_channel(frame, 0.5, 0.5, -*s.chroma_key_strength, s.chroma_key_overexposed, output); + break; + } + case pt_color_magenta_chromakey: { - const int W = frame.cols, H = frame.rows, sz = W*H; - cv::reduce(frame.reshape(1, sz), - output.reshape(1, sz), - 1, cv::REDUCE_AVG); + filter_single_channel(frame, 0.5, -*s.chroma_key_strength, 0.5, s.chroma_key_overexposed, output); break; } + case pt_color_hardware: + eval_once(qDebug() << "camera driver doesn't support grayscale"); + goto do_grayscale; default: eval_once(qDebug() << "wrong pt_color_type enum value" << int(s.blob_color)); [[fallthrough]]; - case pt_color_natural: + case pt_color_bt709: +do_grayscale: cv::cvtColor(frame, output, cv::COLOR_BGR2GRAY); break; } @@ -177,11 +235,10 @@ void PointExtractor::threshold_image(const cv::Mat& frame_gray, cv::Mat1b& outpu const f radius = threshold_radius_value(frame_gray.cols, frame_gray.rows, threshold_slider_value); float const* const __restrict ptr = hist.ptr<float>(0); - const unsigned area = uround(3 * pi * radius*radius); + const unsigned area = unsigned(iround(3 * pi * radius*radius)); const unsigned sz = unsigned(hist.cols * hist.rows); - constexpr unsigned min_thres = 64; - unsigned thres = min_thres; - for (unsigned i = sz-1, cnt = 0; i > 32; i--) + unsigned thres = 1; + for (unsigned i = sz-1, cnt = 0; i > 1; i--) { cnt += (unsigned)ptr[i]; if (cnt >= area) @@ -209,20 +266,15 @@ static void draw_blobs(cv::Mat& preview_frame, const blob* blobs, unsigned nblob cy = preview_frame.rows / f(size.height), c = std::fmax(f(1), cx+cy)/2; - constexpr unsigned fract_bits = 8; - constexpr int c_fract(1 << fract_bits); - - cv::Point p(iround(b.pos[0] * cx * c_fract), iround(b.pos[1] * cy * c_fract)); + cv::Point p(iround(b.pos[0] * cx), iround(b.pos[1] * cy)); - auto circle_color = k >= PointModel::N_POINTS - ? cv::Scalar(192, 192, 192) - : cv::Scalar(255, 255, 0); + auto outline_color = k >= PointModel::N_POINTS + ? cv::Scalar(192, 192, 192) + : cv::Scalar(255, 255, 0); - const int overlay_size = iround(dpi); - - cv::circle(preview_frame, p, iround((b.radius + f(3.3) * c) * c_fract), - circle_color, overlay_size, - cv::LINE_AA, fract_bits); + cv::ellipse(preview_frame, p, + {iround(b.rect.width/(f)2+2*c), iround(b.rect.height/(f)2+2*c)}, + 0, 0, 360, outline_color, iround(dpi), cv::LINE_AA); char buf[16]; std::snprintf(buf, sizeof(buf), "%.2fpx", (double)b.radius); @@ -233,25 +285,51 @@ static void draw_blobs(cv::Mat& preview_frame, const blob* blobs, unsigned nblob cv::Point pos(iround(b.pos[0]*cx+offx), iround(b.pos[1]*cy+offy)); cv::putText(preview_frame, buf, pos, - cv::FONT_HERSHEY_PLAIN, overlay_size, text_color, + cv::FONT_HERSHEY_PLAIN, iround(dpi), text_color, 1); } } -void PointExtractor::extract_points(const pt_frame& frame_, pt_preview& preview_frame_, std::vector<vec2>& points) +static vec2 meanshift_initial_guess(const cv::Rect rect, cv::Mat& frame_roi) +{ + vec2 ret = {rect.width/(f)2, rect.height/(f)2}; + + // compute center initial guess + double ynorm = 0, xnorm = 0, y = 0, x = 0; + for (int j = 0; j < rect.height; j++) + { + const unsigned char* __restrict ptr = frame_roi.ptr<unsigned char>(j); + for (int i = 0; i < rect.width; i++) + { + double val = ptr[i] * 1./255; + x += i * val; + y += j * val; + xnorm += val; + ynorm += val; + } + } + constexpr double eps = 1e-4; + if (xnorm > eps && ynorm > eps) + ret = { (f)(x / xnorm), (f)(y / ynorm) }; + return ret; +} + +void PointExtractor::extract_points(const pt_frame& frame_, + pt_preview& preview_frame_, + bool preview_visible, + std::vector<vec2>& points) { const cv::Mat& frame = frame_.as_const<Frame>()->mat; ensure_buffers(frame); - color_to_grayscale(frame, frame_gray_unmasked); + color_to_grayscale(frame, frame_gray); #if defined PREVIEW cv::imshow("capture", frame_gray); cv::waitKey(1); #endif - threshold_image(frame_gray_unmasked, frame_bin); - frame_gray_unmasked.copyTo(frame_gray, frame_bin); + threshold_image(frame_gray, frame_bin); const f region_size_min = (f)s.min_point_size; const f region_size_max = (f)s.max_point_size; @@ -299,7 +377,7 @@ void PointExtractor::extract_points(const pt_frame& frame_, pt_preview& preview_ } } - const f radius = std::sqrt(cnt / pi); + const f radius = std::sqrt((f)cnt) / std::sqrt(pi); if (radius > region_size_max || radius < region_size_min) continue; @@ -329,21 +407,14 @@ end: for (idx = 0; idx < sz; ++idx) { blob& b = blobs[idx]; - cv::Rect rect = b.rect; - - rect.x -= rect.width / 2; - rect.y -= rect.height / 2; - rect.width *= 2; - rect.height *= 2; - rect &= cv::Rect(0, 0, W, H); // crop at frame boundaries - + cv::Rect rect = b.rect & cv::Rect(0, 0, W, H); // crop at frame boundaries cv::Mat frame_roi = frame_gray(rect); // smaller values mean more changes. 1 makes too many changes while 1.5 makes about .1 static constexpr f radius_c = f(1.75); const f kernel_radius = b.radius * radius_c; - vec2 pos(rect.width/f(2), rect.height/f(2)); // position relative to ROI. + vec2 pos = meanshift_initial_guess(rect, frame_roi); // position relative to ROI. for (int iter = 0; iter < 10; ++iter) { @@ -358,9 +429,10 @@ end: b.pos[1] = pos[1] + rect.y; } - draw_blobs(preview_frame_.as<Frame>()->mat, - blobs.data(), blobs.size(), - frame_gray.size()); + if (preview_visible) + draw_blobs(preview_frame_.as<Frame>()->mat, + blobs.data(), blobs.size(), + frame_gray.size()); // End of mean shift code. At this point, blob positions are updated with hopefully less noisy less biased values. |