diff --git a/src/detector.c b/src/detector.c index 7d504122..5f2cec83 100644 --- a/src/detector.c +++ b/src/detector.c @@ -399,7 +399,7 @@ void validate_detector_recall(char *datacfg, char *cfgfile, char *weightfile) int m = plist->size; int i=0; - float thresh = .2;// .001; + float thresh = .001;// .001; // .2; float iou_thresh = .5; float nms = .4; diff --git a/src/yolo_console_dll.cpp b/src/yolo_console_dll.cpp index 25497ce2..ebafe110 100644 --- a/src/yolo_console_dll.cpp +++ b/src/yolo_console_dll.cpp @@ -2,6 +2,7 @@ #include #include #include +#include #include #include #include @@ -10,8 +11,10 @@ #ifdef _WIN32 #define OPENCV +#include "windows.h" #endif +#define TRACK_OPTFLOW #include "yolo_v2_class.hpp" // imported functions from DLL #ifdef OPENCV @@ -21,6 +24,11 @@ #include "opencv2/videoio/videoio.hpp" #define OPENCV_VERSION CVAUX_STR(CV_VERSION_MAJOR)""CVAUX_STR(CV_VERSION_MINOR)""CVAUX_STR(CV_VERSION_REVISION) #pragma comment(lib, "opencv_world" OPENCV_VERSION ".lib") +#pragma comment(lib, "opencv_cudaoptflow" OPENCV_VERSION ".lib") +#pragma comment(lib, "opencv_cudaimgproc" OPENCV_VERSION ".lib") +#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib") +#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib") +#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib") #else #define OPENCV_VERSION CVAUX_STR(CV_VERSION_EPOCH)""CVAUX_STR(CV_VERSION_MAJOR)""CVAUX_STR(CV_VERSION_MINOR) #pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib") @@ -85,11 +93,15 @@ int main(int argc, char *argv[]) std::string filename; if (argc > 1) filename = argv[1]; - Detector detector("cfg/yolo-voc.cfg", "yolo-voc.weights"); + //Detector detector("cfg/yolo-voc.cfg", "yolo-voc.weights"); + Detector detector("tiny-yolo-voc_air.cfg", "backup/tiny-yolo-voc_air_5000.weights"); auto obj_names = objects_names_from_file("data/voc.names"); std::string out_videofile = "result.avi"; bool const save_output_videofile = false; +#ifdef TRACK_OPTFLOW + Tracker_optflow tracker_flow; +#endif while (true) { @@ -105,6 +117,8 @@ int main(int argc, char *argv[]) protocol == "rtmp://" || protocol == "rtsp://" || protocol == "http://" || protocol == "https:/") // video network stream { cv::Mat cap_frame, cur_frame, det_frame, write_frame; + std::queue track_optflow_queue; + int passed_flow_frames = 0; std::shared_ptr det_image; std::vector result_vec, thread_result_vec; detector.nms = 0.02; // comment it - if track_id is not required @@ -126,7 +140,9 @@ int main(int argc, char *argv[]) if (save_output_videofile) output_video.open(out_videofile, CV_FOURCC('D', 'I', 'V', 'X'), std::max(35, video_fps), frame_size, true); - while (!cur_frame.empty()) { + while (!cur_frame.empty()) + { + // always sync if (t_cap.joinable()) { t_cap.join(); ++fps_cap_counter; @@ -134,22 +150,79 @@ int main(int argc, char *argv[]) } t_cap = std::thread([&]() { cap >> cap_frame; }); - // swap result and input-frame + // swap result bouned-boxes and input-frame if(consumed) { - std::unique_lock lock(mtx); - det_image = detector.mat_to_image_resize(cur_frame); - result_vec = thread_result_vec; - result_vec = detector.tracking(result_vec); // comment it - if track_id is not required - consumed = false; + { + std::unique_lock lock(mtx); + det_image = detector.mat_to_image_resize(cur_frame); + result_vec = thread_result_vec; + result_vec = detector.tracking(result_vec); // comment it - if track_id is not required + + consumed = false; + } + +#ifdef TRACK_OPTFLOW + int y = 0, x = 0; + cv::Mat show_flow = cur_frame.clone(); + auto lambda = [&x, &y](cv::Mat draw_frame, cv::Mat src_frame, std::vector result_vec) { + //if (x > 10) return; + if (result_vec.size() == 0) return; + bbox_t i = result_vec[0]; + //cv::Rect r(i.x, i.y, i.w, i.h); + cv::Rect r(i.x + (i.w-31)/2, i.y + (i.h - 31)/2, 31, 31); + cv::Rect img_rect(cv::Point2i(0, 0), src_frame.size()); + cv::Rect rect_roi = r & img_rect; + if (rect_roi.width < 1 || rect_roi.height < 1) return; + cv::Mat roi = src_frame(rect_roi); + cv::Mat dst; + cv::resize(roi, dst, cv::Size(100, 100)); + if (x > 10) x = 0, ++y; + cv::Rect dst_rect_roi(cv::Point2i(x*100, y*100), dst.size()); + cv::Mat dst_roi = draw_frame(dst_rect_roi); + dst.copyTo(dst_roi); + + ++x; + }; + + + // track optical flow + if (track_optflow_queue.size() > 0) { + std::queue new_track_optflow_queue; + std::cout << "\n !!!! all = " << track_optflow_queue.size() << ", cur = " << passed_flow_frames << std::endl; + //draw_boxes(track_optflow_queue.front().clone(), result_vec, obj_names, 3, current_det_fps, current_cap_fps); + //cv::waitKey(10); + tracker_flow.update_tracking_flow(track_optflow_queue.front()); + lambda(show_flow, track_optflow_queue.front(), result_vec); + track_optflow_queue.pop(); + while(track_optflow_queue.size() > 0) { + //draw_boxes(track_optflow_queue.front().clone(), result_vec, obj_names, 3, current_det_fps, current_cap_fps); + //cv::waitKey(10); + result_vec = tracker_flow.tracking_flow(track_optflow_queue.front(), result_vec); + if (track_optflow_queue.size() <= passed_flow_frames && new_track_optflow_queue.size() == 0) + new_track_optflow_queue = track_optflow_queue; + lambda(show_flow, track_optflow_queue.front(), result_vec); + track_optflow_queue.pop(); + } + track_optflow_queue = new_track_optflow_queue; + new_track_optflow_queue.swap(std::queue()); + passed_flow_frames = 0; + std::cout << "\n !!!! now = " << track_optflow_queue.size() << ", cur = " << passed_flow_frames << std::endl; + + cv::imshow("flow", show_flow); + cv::waitKey(3); + } +#endif + } - // launch thread once + // launch thread once - Detection if (!t_detect.joinable()) { t_detect = std::thread([&]() { auto current_image = det_image; consumed = true; while (current_image.use_count() > 0) { - auto result = detector.detect_resized(*current_image, frame_size, 0.24, true); + auto result = detector.detect_resized(*current_image, frame_size, 0.24, false); // true + Sleep(500); ++fps_det_counter; std::unique_lock lock(mtx); thread_result_vec = result; @@ -169,6 +242,13 @@ int main(int argc, char *argv[]) fps_det_counter = 0; fps_cap_counter = 0; } + +#ifdef TRACK_OPTFLOW + ++passed_flow_frames; + track_optflow_queue.push(cur_frame.clone()); + result_vec = tracker_flow.tracking_flow(cur_frame, result_vec); // track optical flow +#endif + draw_boxes(cur_frame, result_vec, obj_names, 3, current_det_fps, current_cap_fps); //show_console_result(result_vec, obj_names); @@ -183,10 +263,10 @@ int main(int argc, char *argv[]) } // wait detection result for video-file only (not for net-cam) - if (protocol != "rtsp://" && protocol != "http://" && protocol != "https:/") { - std::unique_lock lock(mtx); - while (!consumed) cv.wait(lock); - } + //if (protocol != "rtsp://" && protocol != "http://" && protocol != "https:/") { + // std::unique_lock lock(mtx); + // while (!consumed) cv.wait(lock); + //} } if (t_cap.joinable()) t_cap.join(); if (t_detect.joinable()) t_detect.join(); diff --git a/src/yolo_v2_class.hpp b/src/yolo_v2_class.hpp index a52d3ac4..ab446ce9 100644 --- a/src/yolo_v2_class.hpp +++ b/src/yolo_v2_class.hpp @@ -8,6 +8,11 @@ #include // C++ #include "opencv2/highgui/highgui_c.h" // C #include "opencv2/imgproc/imgproc_c.h" // C + +#include +#include +#include +#include "opencv2/core/cuda.hpp" #endif // OPENCV #ifdef YOLODLL_EXPORTS @@ -41,6 +46,7 @@ struct image_t { class Detector { std::shared_ptr detector_gpu_ptr; + std::deque> prev_bbox_vec_deque; public: float nms = .4; @@ -93,6 +99,7 @@ public: } private: + static image_t ipl_to_image(IplImage* src) { unsigned char *data = (unsigned char *)src->imageData; @@ -142,8 +149,165 @@ private: #endif // OPENCV - std::deque> prev_bbox_vec_deque; }; +#if defined(TRACK_OPTFLOW) && defined(OPENCV) + +class Tracker_optflow { +public: + + + // just to avoid extra allocations + cv::cuda::GpuMat src_mat_gpu; + cv::cuda::GpuMat dst_mat_gpu, dst_grey_gpu; + cv::cuda::GpuMat tmp_grey_gpu; + cv::cuda::GpuMat prev_pts_flow_gpu, cur_pts_flow_gpu; + cv::cuda::GpuMat status_gpu, err_gpu; + + cv::cuda::GpuMat src_grey_gpu; // used in both functions + cv::Ptr sync_PyrLKOpticalFlow_gpu; + + void update_tracking_flow(cv::Mat src_mat, int gpu_id = 0) + { + int const old_gpu_id = cv::cuda::getDevice(); + static const int gpu_count = cv::cuda::getCudaEnabledDeviceCount(); + if (gpu_count > gpu_id) + cv::cuda::setDevice(gpu_id); + + cv::cuda::Stream stream; + + if (sync_PyrLKOpticalFlow_gpu.empty()) { + sync_PyrLKOpticalFlow_gpu = cv::cuda::SparsePyrLKOpticalFlow::create(); + + //sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(31, 31)); //sync_PyrLKOpticalFlow_gpu.winSize = cv::Size(31, 31); + //sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(15, 15)); //sync_PyrLKOpticalFlow_gpu.winSize = cv::Size(15, 15); + sync_PyrLKOpticalFlow_gpu->setWinSize(cv::Size(21, 21)); + sync_PyrLKOpticalFlow_gpu->setMaxLevel(3); //sync_PyrLKOpticalFlow_gpu.maxLevel = 8; // +-32 points // def: 3 + sync_PyrLKOpticalFlow_gpu->setNumIters(6000); //sync_PyrLKOpticalFlow_gpu.iters = 8000; // def: 30 + //??? //sync_PyrLKOpticalFlow_gpu.getMinEigenVals = true; + //std::cout << "sync_PyrLKOpticalFlow_gpu.maxLevel: " << sync_PyrLKOpticalFlow_gpu.maxLevel << std::endl; + //std::cout << "sync_PyrLKOpticalFlow_gpu.iters: " << sync_PyrLKOpticalFlow_gpu.iters << std::endl; + //std::cout << "sync_PyrLKOpticalFlow_gpu.winSize: " << sync_PyrLKOpticalFlow_gpu.winSize << std::endl; + } + + if (src_mat.channels() == 3) { + if (src_mat_gpu.cols == 0) { + src_mat_gpu = cv::cuda::GpuMat(src_mat.size(), src_mat.type()); + src_grey_gpu = cv::cuda::GpuMat(src_mat.size(), CV_8UC1); + } + + src_mat_gpu.upload(src_mat, stream); + cv::cuda::cvtColor(src_mat_gpu, src_grey_gpu, CV_BGR2GRAY, 0, stream); + //std::cout << " \n\n OK !!! \n\n"; + } + cv::cuda::setDevice(old_gpu_id); + } + + + std::vector tracking_flow(cv::Mat dst_mat, std::vector cur_bbox_vec, int gpu_id = 0) + { + if (sync_PyrLKOpticalFlow_gpu.empty()) { + std::cout << "sync_PyrLKOpticalFlow_gpu isn't initialized \n"; + return cur_bbox_vec; + } + + int const old_gpu_id = cv::cuda::getDevice(); + static const int gpu_count = cv::cuda::getCudaEnabledDeviceCount(); + if (gpu_count > gpu_id) + cv::cuda::setDevice(gpu_id); + + cv::cuda::Stream stream; + + if (dst_mat_gpu.cols == 0) { + dst_mat_gpu = cv::cuda::GpuMat(dst_mat.size(), dst_mat.type()); + dst_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1); + tmp_grey_gpu = cv::cuda::GpuMat(dst_mat.size(), CV_8UC1); + } + + + dst_mat_gpu.upload(dst_mat, stream); + + cv::cuda::cvtColor(dst_mat_gpu, dst_grey_gpu, CV_BGR2GRAY, 0, stream); + + if (src_grey_gpu.rows != dst_grey_gpu.rows || src_grey_gpu.cols != dst_grey_gpu.cols) { + stream.waitForCompletion(); + src_grey_gpu = dst_grey_gpu.clone(); + cv::cuda::setDevice(old_gpu_id); + return cur_bbox_vec; + } + + cv::Mat prev_pts, prev_pts_flow_cpu, cur_pts_flow_cpu; + + for (auto &i : cur_bbox_vec) { + float x_center = (i.x + i.w / 2); + float y_center = (i.y + i.h / 2); + prev_pts.push_back(cv::Point2f(x_center, y_center)); + } + + + if (prev_pts.rows == 0) + prev_pts_flow_cpu = cv::Mat(); + else + cv::transpose(prev_pts, prev_pts_flow_cpu); + + + if (prev_pts_flow_gpu.cols < prev_pts_flow_cpu.cols) { + prev_pts_flow_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), prev_pts_flow_cpu.type()); + cur_pts_flow_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), prev_pts_flow_cpu.type()); + + status_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), CV_8UC1); + err_gpu = cv::cuda::GpuMat(prev_pts_flow_cpu.size(), CV_32FC1); + } + + prev_pts_flow_gpu.upload(cv::Mat(prev_pts_flow_cpu), stream); + + + dst_grey_gpu.copyTo(tmp_grey_gpu, stream); + + //sync_PyrLKOpticalFlow_gpu.sparse(src_grey_gpu, dst_grey_gpu, prev_pts_flow_gpu, cur_pts_flow_gpu, status_gpu, &err_gpu); // OpenCV 2.4.x + sync_PyrLKOpticalFlow_gpu->calc(src_grey_gpu, dst_grey_gpu, prev_pts_flow_gpu, cur_pts_flow_gpu, status_gpu, err_gpu, stream); // OpenCV 3.x + //std::cout << "\n 1-e \n"; + + cur_pts_flow_gpu.download(cur_pts_flow_cpu, stream); + + tmp_grey_gpu.copyTo(src_grey_gpu, stream); + + cv::Mat err_cpu, status_cpu; + err_gpu.download(err_cpu, stream); + status_gpu.download(status_cpu, stream); + + stream.waitForCompletion(); + + std::vector result_bbox_vec; + + for (size_t i = 0; i < cur_bbox_vec.size(); ++i) + { + cv::Point2f cur_key_pt = cur_pts_flow_cpu.at(0, i); + cv::Point2f prev_key_pt = prev_pts_flow_cpu.at(0, i); + + float moved_x = cur_key_pt.x - prev_key_pt.x; + float moved_y = cur_key_pt.y - prev_key_pt.y; + + if (err_cpu.cols > i && status_cpu.cols > i) + if (abs(moved_x) < 100 && abs(moved_y) < 100) + //if (err_cpu.at(0, i) < 60 && status_cpu.at(0, i) != 0) + { + cur_bbox_vec[i].x += moved_x + 0.5; + cur_bbox_vec[i].y += moved_y + 0.5; + result_bbox_vec.push_back(cur_bbox_vec[i]); + } + } + + cv::cuda::setDevice(old_gpu_id); + + return result_bbox_vec; + } + +}; +#else + +class Tracker_optflow {}; + +#endif // defined(TRACK_OPTFLOW) && defined(OPENCV)