Added CUDA-streams to Darknet-Yolo forward inference

AlexeyAB

2017-12-31 9d23aad8696268e8ce3a94fee9490fd1db000dc8

 src/yolo_console_dll.cpp

@@ -2,12 +2,19 @@
#include <iomanip> 
#include <string>
#include <vector>
#include <queue>
#include <fstream>
#include <thread>
#include <atomic>
#include <mutex>              // std::mutex, std::unique_lock
#include <condition_variable> // std::condition_variable

#ifdef _WIN32
#define OPENCV
#include "windows.h"
#endif

#define TRACK_OPTFLOW
#include "yolo_v2_class.hpp"  // imported functions from DLL

#ifdef OPENCV
@@ -15,19 +22,31 @@
#include "opencv2/core/version.hpp"
#ifndef CV_VERSION_EPOCH
#include "opencv2/videoio/videoio.hpp"
#pragma comment(lib, "opencv_world320.lib")  
#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
#pragma comment(lib, "opencv_core2413.lib")  
#pragma comment(lib, "opencv_imgproc2413.lib")  
#pragma comment(lib, "opencv_highgui2413.lib") 
#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")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
#endif


void draw_boxes(cv::Mat mat_img, std::vector<bbox_t> result_vec, std::vector<std::string> obj_names, 
   unsigned int wait_msec = 0, int current_fps = -1)
   unsigned int wait_msec = 0, int current_det_fps = -1, int current_cap_fps = -1)
{
   int const colors[6][3] = { { 1,0,1 },{ 0,0,1 },{ 0,1,1 },{ 0,1,0 },{ 1,1,0 },{ 1,0,0 } };

   for (auto &i : result_vec) {
      cv::Scalar color(60, 160, 260);
      int const offset = i.obj_id * 123457 % 6;
      int const color_scale = 150 + (i.obj_id * 123457) % 100;
      cv::Scalar color(colors[offset][0], colors[offset][1], colors[offset][2]);
      color *= color_scale;
      cv::rectangle(mat_img, cv::Rect(i.x, i.y, i.w, i.h), color, 5);
      if (obj_names.size() > i.obj_id) {
         std::string obj_name = obj_names[i.obj_id];
@@ -40,15 +59,17 @@
         putText(mat_img, obj_name, cv::Point2f(i.x, i.y - 10), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(0, 0, 0), 2);
      }
   }
   if(current_fps >= 0)
      putText(mat_img, "FPS: " + std::to_string(current_fps), cv::Point2f(10, 20), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(50, 255, 0), 2);
   if (current_det_fps >= 0 && current_cap_fps >= 0) {
      std::string fps_str = "FPS detection: " + std::to_string(current_det_fps) + "   FPS capture: " + std::to_string(current_cap_fps);
      putText(mat_img, fps_str, cv::Point2f(10, 20), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(50, 255, 0), 2);
   }
   cv::imshow("window name", mat_img);
   cv::waitKey(wait_msec);
}
#endif   // OPENCV


void show_result(std::vector<bbox_t> const result_vec, std::vector<std::string> const obj_names) {
void show_console_result(std::vector<bbox_t> const result_vec, std::vector<std::string> const obj_names) {
   for (auto &i : result_vec) {
      if (obj_names.size() > i.obj_id) std::cout << obj_names[i.obj_id] << " - ";
      std::cout << "obj_id = " << i.obj_id << ",  x = " << i.x << ", y = " << i.y 
@@ -61,7 +82,7 @@
   std::ifstream file(filename);
   std::vector<std::string> file_lines;
   if (!file.is_open()) return file_lines;
   for(std::string line; file >> line;) file_lines.push_back(line);
   for(std::string line; getline(file, line);) file_lines.push_back(line);
   std::cout << "object names loaded \n";
   return file_lines;
}
@@ -72,11 +93,15 @@
   std::string filename;
   if (argc > 1) filename = argv[1];

   Detector detector("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) 
   {     
@@ -89,43 +114,177 @@
         std::string const file_ext = filename.substr(filename.find_last_of(".") + 1);
         std::string const protocol = filename.substr(0, 7);
         if (file_ext == "avi" || file_ext == "mp4" || file_ext == "mjpg" || file_ext == "mov" ||  // video file
            protocol == "rtsp://" || protocol == "http://" || protocol == "https:/")   // video network stream
            protocol == "rtmp://" || protocol == "rtsp://" || protocol == "http://" || protocol == "https:/")  // video network stream
         {
            cv::Mat frame, prev_frame, det_frame;
            cv::Mat cap_frame, cur_frame, det_frame, write_frame;
            std::queue<cv::Mat> track_optflow_queue;
            int passed_flow_frames = 0;
            std::shared_ptr<image_t> det_image;
            std::vector<bbox_t> result_vec, thread_result_vec;
            detector.nms = 0.02; // comment it - if track_id is not required
            std::thread td([]() {});
            std::atomic<int> ready_flag;
            ready_flag = true;
            std::atomic<bool> consumed, videowrite_ready;
            consumed = true;
            videowrite_ready = true;
            std::atomic<int> fps_det_counter, fps_cap_counter;
            fps_det_counter = 0;
            fps_cap_counter = 0;
            int current_det_fps = 0, current_cap_fps = 0;
            std::thread t_detect, t_cap, t_videowrite;
            std::mutex mtx;
            std::condition_variable cv;
            std::chrono::steady_clock::time_point steady_start, steady_end;
            int fps_counter = 0, current_fps = 0;
            cv::VideoCapture cap(filename); cap >> cur_frame;
            int const video_fps = cap.get(CV_CAP_PROP_FPS);
            cv::Size const frame_size = cur_frame.size();
            cv::VideoWriter output_video;
            for (cv::VideoCapture cap(filename); cap >> frame, cap.isOpened();) {
               if(!output_video.isOpened() && save_output_videofile)
                  output_video.open(out_videofile, CV_FOURCC('D','I','V','X'), cap.get(CV_CAP_PROP_FPS), frame.size(), true);
               if (ready_flag || (protocol != "rtsp://" && protocol != "http://" && protocol != "https:/")) {
                  td.join();
                  ++fps_counter;
                  ready_flag = false;
                  result_vec = thread_result_vec;
                  result_vec = detector.tracking(result_vec);  // comment it - if track_id is not required
                  det_frame = frame.clone();						
                  td = std::thread([&]() { thread_result_vec = detector.detect(det_frame, 0.24, true); ready_flag = true; });
            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()) 
            {
               // always sync
               if (t_cap.joinable()) {
                  t_cap.join();
                  ++fps_cap_counter;
                  cur_frame = cap_frame.clone();
               }
               if (!prev_frame.empty()) {	
               t_cap = std::thread([&]() { cap >> cap_frame; });

               // swap result bouned-boxes and input-frame
               if(consumed)
               {
                  {
                     std::unique_lock<std::mutex> 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<bbox_t> 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) {
                     //show_flow = track_optflow_queue.front().clone();
                     //draw_boxes(show_flow, result_vec, obj_names, 3, current_det_fps, current_cap_fps);

                     std::queue<cv::Mat> new_track_optflow_queue;
                     //std::cout << "\n !!!! all = " << track_optflow_queue.size() << ", cur = " << passed_flow_frames << std::endl;
                     if (result_vec.size() > 0) {
                        draw_boxes(track_optflow_queue.front().clone(), result_vec, obj_names, 3, current_det_fps, current_cap_fps);
                        std::cout << "\n frame_size = " << track_optflow_queue.size() << std::endl;
                        cv::waitKey(1000);
                     }
                     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) {
                        if (result_vec.size() > 0) {
                           draw_boxes(track_optflow_queue.front().clone(), result_vec, obj_names, 3, current_det_fps, current_cap_fps);
                           std::cout << "\n frame_size = " << track_optflow_queue.size() << std::endl;
                           cv::waitKey(1000);
                        }
                        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<cv::Mat>());
                     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);
                     //if (result_vec.size() > 0) {
                     // cv::waitKey(1000);
                     //}
                  }
#endif

               }
               // 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, false);  // true
                        //Sleep(200);
                        Sleep(50);
                        ++fps_det_counter;
                        std::unique_lock<std::mutex> lock(mtx);
                        thread_result_vec = result;
                        current_image = det_image;
                        consumed = true;
                        cv.notify_all();
                     }
                  });
               }

               if (!cur_frame.empty()) {
                  steady_end = std::chrono::steady_clock::now();
                  if (std::chrono::duration<double>(steady_end - steady_start).count() >= 1) {
                     current_fps = fps_counter;
                     current_det_fps = fps_det_counter;
                     current_cap_fps = fps_cap_counter;
                     steady_start = steady_end;
                     fps_counter = 0;
                     fps_det_counter = 0;
                     fps_cap_counter = 0;
                  }
                  draw_boxes(prev_frame, result_vec, obj_names, 3, current_fps);
                  show_result(result_vec, obj_names);
                  if (output_video.isOpened())
                     output_video << prev_frame;

#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);

                  if (output_video.isOpened() && videowrite_ready) {
                     if (t_videowrite.joinable()) t_videowrite.join();
                     write_frame = cur_frame.clone();
                     videowrite_ready = false;
                     t_videowrite = std::thread([&]() { 
                         output_video << write_frame; videowrite_ready = true;
                     });
                  }
               }
               prev_frame = frame.clone();					

               // wait detection result for video-file only (not for net-cam)
               //if (protocol != "rtsp://" && protocol != "http://" && protocol != "https:/") {
               // std::unique_lock<std::mutex> lock(mtx);
               // while (!consumed) cv.wait(lock);
               //}
            }
            if (t_cap.joinable()) t_cap.join();
            if (t_detect.joinable()) t_detect.join();
            if (t_videowrite.joinable()) t_videowrite.join();
            std::cout << "Video ended \n";
         }
         else if (file_ext == "txt") { // list of image files
            std::ifstream file(filename);
@@ -135,7 +294,7 @@
                  std::cout << line << std::endl;
                  cv::Mat mat_img = cv::imread(line);
                  std::vector<bbox_t> result_vec = detector.detect(mat_img);
                  show_result(result_vec, obj_names);
                  show_console_result(result_vec, obj_names);
                  //draw_boxes(mat_img, result_vec, obj_names);
                  //cv::imwrite("res_" + line, mat_img);
               }
@@ -146,7 +305,7 @@
            std::vector<bbox_t> result_vec = detector.detect(mat_img);
            result_vec = detector.tracking(result_vec);  // comment it - if track_id is not required
            draw_boxes(mat_img, result_vec, obj_names);
            show_result(result_vec, obj_names);
            show_console_result(result_vec, obj_names);
         }
#else
         //std::vector<bbox_t> result_vec = detector.detect(filename);
@@ -154,7 +313,7 @@
         auto img = detector.load_image(filename);
         std::vector<bbox_t> result_vec = detector.detect(img);
         detector.free_image(img);
         show_result(result_vec, obj_names);
         show_console_result(result_vec, obj_names);
#endif         
      }
      catch (std::exception &e) { std::cerr << "exception: " << e.what() << "\n"; getchar(); }