Added object tracking using optical flow

AlexeyAB

2017-12-25 7113d2cc7fceb0fb599326aec65771c33f5f4386

 src/yolo_v2_class.hpp

@@ -8,12 +8,25 @@
#include <opencv2/opencv.hpp>       // C++
#include "opencv2/highgui/highgui_c.h" // C
#include "opencv2/imgproc/imgproc_c.h" // C

#include <opencv2/cudaoptflow.hpp>
#include <opencv2/cudaimgproc.hpp>
#include <opencv2/cudaarithm.hpp>
#include "opencv2/core/cuda.hpp"
#endif   // OPENCV

#ifdef YOLODLL_EXPORTS
#if defined(_MSC_VER)
#define YOLODLL_API __declspec(dllexport) 
#else
#define YOLODLL_API __attribute__((visibility("default")))
#endif
#else
#if defined(_MSC_VER)
#define YOLODLL_API __declspec(dllimport) 
#else
#define YOLODLL_API
#endif
#endif

struct bbox_t {
@@ -33,35 +46,59 @@

class Detector {
   std::shared_ptr<void> detector_gpu_ptr;
   std::deque<std::vector<bbox_t>> prev_bbox_vec_deque;
public:
   float nms = .4;

   YOLODLL_API Detector(std::string cfg_filename, std::string weight_filename, int gpu_id = 0);
   YOLODLL_API ~Detector();

   YOLODLL_API std::vector<bbox_t> detect(std::string image_filename, float thresh = 0.2);
   YOLODLL_API std::vector<bbox_t> detect(image_t img, float thresh = 0.2);
   YOLODLL_API std::vector<bbox_t> detect(std::string image_filename, float thresh = 0.2, bool use_mean = false);
   YOLODLL_API std::vector<bbox_t> detect(image_t img, float thresh = 0.2, bool use_mean = false);
   static YOLODLL_API image_t load_image(std::string image_filename);
   static YOLODLL_API void free_image(image_t m);
   YOLODLL_API int get_net_width() const;
   YOLODLL_API int get_net_height() const;

   YOLODLL_API std::vector<bbox_t> tracking(std::vector<bbox_t> cur_bbox_vec, int const frames_story = 6);

#ifdef OPENCV
   std::vector<bbox_t> detect(cv::Mat mat, float thresh = 0.2) 
   std::vector<bbox_t> detect(cv::Mat mat, float thresh = 0.2, bool use_mean = false)
   {
      if(mat.data == NULL)
         throw std::runtime_error("file not found");
         throw std::runtime_error("Image is empty");
      auto image_ptr = mat_to_image_resize(mat);
      return detect_resized(*image_ptr, mat.size(), thresh, use_mean);
   }

   std::vector<bbox_t> detect_resized(image_t img, cv::Size init_size, float thresh = 0.2, bool use_mean = false)
   {
      if (img.data == NULL)
         throw std::runtime_error("Image is empty");
      auto detection_boxes = detect(img, thresh, use_mean);
      float wk = (float)init_size.width / img.w, hk = (float)init_size.height / img.h;
      for (auto &i : detection_boxes) i.x *= wk, i.w *= wk, i.y *= hk, i.h *= hk;
      return detection_boxes;
   }

   std::shared_ptr<image_t> mat_to_image_resize(cv::Mat mat) const
   {
      if (mat.data == NULL) return std::shared_ptr<image_t>(NULL);
      cv::Mat det_mat;
      cv::resize(mat, det_mat, cv::Size(get_net_width(), get_net_height()));
      return mat_to_image(det_mat);
   }

   static std::shared_ptr<image_t> mat_to_image(cv::Mat img)
   {
      std::shared_ptr<image_t> image_ptr(new image_t, [](image_t *img) { free_image(*img); delete img; });
      *image_ptr = mat_to_image(mat);
      return detect(*image_ptr, thresh);
      std::shared_ptr<IplImage> ipl_small = std::make_shared<IplImage>(img);
      *image_ptr = ipl_to_image(ipl_small.get());
      rgbgr_image(*image_ptr);
      return image_ptr;
   }

private:
   static image_t mat_to_image(cv::Mat img)
   {
      std::shared_ptr<IplImage> ipl_small = std::make_shared<IplImage>(img);
      image_t im_small = ipl_to_image(ipl_small.get());
      rgbgr_image(im_small);
      return im_small;
   }

   static image_t ipl_to_image(IplImage* src)
   {
@@ -112,59 +149,165 @@

#endif   // OPENCV

   std::deque<std::vector<bbox_t>> prev_bbox_vec_deque;

public:
   std::vector<bbox_t> tracking(std::vector<bbox_t> cur_bbox_vec, int const frames_story = 4)
   {
      bool prev_track_id_present = false;
      for (auto &i : prev_bbox_vec_deque)
         if (i.size() > 0) prev_track_id_present = true;

      static unsigned int track_id = 1;

      if(!prev_track_id_present) {
         //track_id = 1;
         for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
            cur_bbox_vec[i].track_id = track_id++;
         prev_bbox_vec_deque.push_front(cur_bbox_vec);
         if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();
         return cur_bbox_vec;
      }

      std::vector<unsigned int> dist_vec(cur_bbox_vec.size(), std::numeric_limits<unsigned int>::max());

      for (auto &prev_bbox_vec : prev_bbox_vec_deque) {
         for (auto &i : prev_bbox_vec) {
            int cur_index = -1;
            for (size_t m = 0; m < cur_bbox_vec.size(); ++m) {
               bbox_t const& k = cur_bbox_vec[m];
               if (i.obj_id == k.obj_id) {
                  unsigned int cur_dist = sqrt(((float)i.x - k.x)*((float)i.x - k.x) + ((float)i.y - k.y)*((float)i.y - k.y));
                  if (cur_dist < 100 && (k.track_id == 0 || dist_vec[m] > cur_dist)) {
                     dist_vec[m] = cur_dist;
                     cur_index = m;
                  }
               }
            }

            bool track_id_absent = !std::any_of(cur_bbox_vec.begin(), cur_bbox_vec.end(), [&](bbox_t const& b) { return b.track_id == i.track_id; });

            if (cur_index >= 0 && track_id_absent)
               cur_bbox_vec[cur_index].track_id = i.track_id;
         }
      }

      for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
         if (cur_bbox_vec[i].track_id == 0)
            cur_bbox_vec[i].track_id = track_id++;

      prev_bbox_vec_deque.push_front(cur_bbox_vec);
      if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();

      return cur_bbox_vec;
   }
};


#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<cv::cuda::SparsePyrLKOpticalFlow> 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<bbox_t> tracking_flow(cv::Mat dst_mat, std::vector<bbox_t> 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<bbox_t> result_bbox_vec;

      for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
      {
         cv::Point2f cur_key_pt = cur_pts_flow_cpu.at<cv::Point2f>(0, i);
         cv::Point2f prev_key_pt = prev_pts_flow_cpu.at<cv::Point2f>(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<float>(0, i) < 60 && status_cpu.at<unsigned char>(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)