Added linear extrapolation of coordinates

AlexeyAB

2018-02-12 14db8f3384e72d10952da882cd0dc111c582cc45

Added linear extrapolation of coordinates

 src/yolo_console_dll.cpp

@@ -14,7 +14,7 @@
#endif

// To use tracking - uncomment the following line. Tracking is supported only by OpenCV 3.x
//#define TRACK_OPTFLOW
#define TRACK_OPTFLOW

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

@@ -37,6 +37,140 @@
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
#endif

class track_kalman {
public:
    cv::KalmanFilter kf;
    int state_size, meas_size, contr_size;


    track_kalman(int _state_size = 10, int _meas_size = 10, int _contr_size = 0)
        : state_size(_state_size), meas_size(_meas_size), contr_size(_contr_size)
    {
        kf.init(state_size, meas_size, contr_size, CV_32F);

        cv::setIdentity(kf.measurementMatrix);
        cv::setIdentity(kf.measurementNoiseCov, cv::Scalar::all(1e-1));
        cv::setIdentity(kf.processNoiseCov, cv::Scalar::all(1e-5));
        cv::setIdentity(kf.errorCovPost, cv::Scalar::all(1e-2));
        cv::setIdentity(kf.transitionMatrix);
    }

    void set(std::vector<bbox_t> result_vec) {
        for (size_t i = 0; i < result_vec.size() && i < state_size*2; ++i) {
            kf.statePost.at<float>(i * 2 + 0) = result_vec[i].x;
            kf.statePost.at<float>(i * 2 + 1) = result_vec[i].y;
        }
    }

    // Kalman.correct() calculates: statePost = statePre + gain * (z(k)-measurementMatrix*statePre);
    // corrected state (x(k)): x(k)=x'(k)+K(k)*(z(k)-H*x'(k))
    std::vector<bbox_t> correct(std::vector<bbox_t> result_vec) {
        cv::Mat measurement(meas_size, 1, CV_32F);
        for (size_t i = 0; i < result_vec.size() && i < meas_size * 2; ++i) {
            measurement.at<float>(i * 2 + 0) = result_vec[i].x;
            measurement.at<float>(i * 2 + 1) = result_vec[i].y;
        }
        cv::Mat estimated = kf.correct(measurement);
        for (size_t i = 0; i < result_vec.size() && i < meas_size * 2; ++i) {
            result_vec[i].x = estimated.at<float>(i * 2 + 0);
            result_vec[i].y = estimated.at<float>(i * 2 + 1);
        }
        return result_vec;
    }

    // Kalman.predict() calculates: statePre = TransitionMatrix * statePost;
    // predicted state (x'(k)): x(k)=A*x(k-1)+B*u(k)
    std::vector<bbox_t> predict() {
        std::vector<bbox_t> result_vec;
        cv::Mat control;
        cv::Mat prediction = kf.predict(control);
        for (size_t i = 0; i < prediction.rows && i < state_size * 2; ++i) {
            result_vec[i].x = prediction.at<float>(i * 2 + 0);
            result_vec[i].y = prediction.at<float>(i * 2 + 1);
        }
        return result_vec;
    }

};




class extrapolate_coords_t {
public:
    std::vector<bbox_t> old_result_vec;
    std::vector<float> dx_vec, dy_vec, time_vec;
    std::vector<float> old_dx_vec, old_dy_vec;

    void new_result(std::vector<bbox_t> new_result_vec, float new_time) {
        old_dx_vec = dx_vec;
        old_dy_vec = dy_vec;
        if (old_dx_vec.size() != old_result_vec.size()) std::cout << "old_dx != old_res \n";
        dx_vec = std::vector<float>(new_result_vec.size(), 0);
        dy_vec = std::vector<float>(new_result_vec.size(), 0);
        update_result(new_result_vec, new_time, false);
        old_result_vec = new_result_vec;
        time_vec = std::vector<float>(new_result_vec.size(), new_time);
    }

    void update_result(std::vector<bbox_t> new_result_vec, float new_time, bool update = true) {
        for (size_t i = 0; i < new_result_vec.size(); ++i) {
            for (size_t k = 0; k < old_result_vec.size(); ++k) {
                if (old_result_vec[k].track_id == new_result_vec[i].track_id && old_result_vec[k].obj_id == new_result_vec[i].obj_id) {
                    float const delta_time = new_time - time_vec[k];
                    if (abs(delta_time) < 1) break;
                    size_t index = (update) ? k : i;
                    float dx = ((float)new_result_vec[i].x - (float)old_result_vec[k].x) / delta_time;
                    float dy = ((float)new_result_vec[i].y - (float)old_result_vec[k].y) / delta_time;
                    float old_dx = dx, old_dy = dy;

                    // if it's shaking
                    if (update) {
                        if (dx * dx_vec[i] < 0) dx = dx / 2;
                        if (dy * dy_vec[i] < 0) dy = dy / 2;
                    } else {
                        if (dx * old_dx_vec[k] < 0) dx = dx / 2;
                        if (dy * old_dy_vec[k] < 0) dy = dy / 2;
                    }
                    dx_vec[index] = dx;
                    dy_vec[index] = dy;

                    //if (old_dx == dx && old_dy == dy) std::cout << "not shakin \n";
                    //else std::cout << "shakin \n";

                    if (dx_vec[index] > 1000 || dy_vec[index] > 1000) {
                        //std::cout << "!!! bad dx or dy, dx = " << dx_vec[index] << ", dy = " << dy_vec[index] << 
                        //  ", delta_time = " << delta_time << ", update = " << update << std::endl;
                        dx_vec[index] = 0;
                        dy_vec[index] = 0;						
                    }
                    old_result_vec[k].x = new_result_vec[i].x;
                    old_result_vec[k].y = new_result_vec[i].y;
                    time_vec[k] = new_time;
                    break;
                }
            }
        }
    }

    std::vector<bbox_t> predict(float cur_time) {
        std::vector<bbox_t> result_vec = old_result_vec;
        for (size_t i = 0; i < old_result_vec.size(); ++i) {
            float const delta_time = cur_time - time_vec[i];
            auto &bbox = result_vec[i];
            float new_x = (float) bbox.x + dx_vec[i] * delta_time;
            float new_y = (float) bbox.y + dy_vec[i] * delta_time;
            if (new_x > 0) bbox.x = new_x;
            else bbox.x = 0;
            if (new_y > 0) bbox.y = new_y;
            else bbox.y = 0;
        }
        return result_vec;
    }

};


void draw_boxes(cv::Mat mat_img, std::vector<bbox_t> result_vec, std::vector<std::string> obj_names, 
    int current_det_fps = -1, int current_cap_fps = -1)
{
@@ -104,7 +238,7 @@

    auto obj_names = objects_names_from_file(names_file);
    std::string out_videofile = "result.avi";
    bool const save_output_videofile = false;
    bool const save_output_videofile = true;
#ifdef TRACK_OPTFLOW
    Tracker_optflow tracker_flow;
    detector.wait_stream = true;
@@ -118,6 +252,9 @@
        
        try {
#ifdef OPENCV
            extrapolate_coords_t extrapolate_coords;
            bool extrapolate_flag = false;
            float cur_time_extrapolate = 0, old_time_extrapolate = 0;
            preview_boxes_t large_preview(100, 150, false), small_preview(50, 50, true);
            bool show_small_boxes = false;

@@ -159,6 +296,7 @@
                        cur_frame = cap_frame.clone();
                    }
                    t_cap = std::thread([&]() { cap >> cap_frame; });
                    ++cur_time_extrapolate;

                    // swap result bouned-boxes and input-frame
                    if(consumed)
@@ -177,18 +315,23 @@

                            while (track_optflow_queue.size() > 1) {
                                track_optflow_queue.pop();
                                result_vec = tracker_flow.tracking_flow(track_optflow_queue.front(), false);
                                result_vec = tracker_flow.tracking_flow(track_optflow_queue.front(), true);
                            }
                            track_optflow_queue.pop();
                            passed_flow_frames = 0;

                            result_vec = detector.tracking_id(result_vec);
                            auto tmp_result_vec = detector.tracking_id(detected_result_vec);
                            auto tmp_result_vec = detector.tracking_id(detected_result_vec, false);
                            small_preview.set(first_frame, tmp_result_vec);
                        }
#endif
                        result_vec = detector.tracking_id(result_vec);  // comment it - if track_id is not required					

                            extrapolate_coords.new_result(tmp_result_vec, old_time_extrapolate);
                            old_time_extrapolate = cur_time_extrapolate;
                            extrapolate_coords.update_result(result_vec, cur_time_extrapolate - 1);
                        }
#else
                        result_vec = detector.tracking_id(result_vec);  // comment it - if track_id is not required					
                        extrapolate_coords.new_result(result_vec, cur_time_extrapolate - 1);
#endif
                        // add old tracked objects
                        for (auto &i : old_result_vec) {
                            auto it = std::find_if(result_vec.begin(), result_vec.end(),
@@ -204,7 +347,7 @@
                        }
#ifdef TRACK_OPTFLOW
                        tracker_flow.update_cur_bbox_vec(result_vec);
                        result_vec = tracker_flow.tracking_flow(cur_frame, false);  // track optical flow
                        result_vec = tracker_flow.tracking_flow(cur_frame, true);   // track optical flow
#endif
                        consumed = false;
                        cv_pre_tracked.notify_all();
@@ -245,16 +388,23 @@
                        ++passed_flow_frames;
                        track_optflow_queue.push(cur_frame.clone());
                        result_vec = tracker_flow.tracking_flow(cur_frame); // track optical flow
                        extrapolate_coords.update_result(result_vec, cur_time_extrapolate);
                        small_preview.draw(cur_frame, show_small_boxes);
#endif                      

                        draw_boxes(cur_frame, result_vec, obj_names, current_det_fps, current_cap_fps);
                        auto result_vec_draw = result_vec;
                        if (extrapolate_flag) {
                            result_vec_draw = extrapolate_coords.predict(cur_time_extrapolate);
                            cv::putText(cur_frame, "extrapolate", cv::Point2f(10, 40), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.0, cv::Scalar(50, 50, 0), 2);
                        }
                        draw_boxes(cur_frame, result_vec_draw, obj_names, current_det_fps, current_cap_fps);
                        //show_console_result(result_vec, obj_names);
                        large_preview.draw(cur_frame);

                        cv::imshow("window name", cur_frame);
                        int key = cv::waitKey(3);   // 3 or 16ms
                        if (key == 'f') show_small_boxes = !show_small_boxes;
                        if (key == 'p') while (true) if(cv::waitKey(100) == 'p') break;
                        if (key == 'e') extrapolate_flag = !extrapolate_flag;

                        if (output_video.isOpened() && videowrite_ready) {
                            if (t_videowrite.joinable()) t_videowrite.join();

 src/yolo_v2_class.hpp

@@ -297,7 +297,8 @@
                float moved_y = cur_key_pt.y - prev_key_pt.y;

                if (abs(moved_x) < 100 && abs(moved_y) < 100 && good_bbox_vec_flags[i])
                    if (err_cpu.at<float>(0, i) < flow_error && status_cpu.at<unsigned char>(0, i) != 0)
                    if (err_cpu.at<float>(0, i) < flow_error && status_cpu.at<unsigned char>(0, i) != 0 &&
                        ((float)cur_bbox_vec[i].x + moved_x) > 0 && ((float)cur_bbox_vec[i].y + moved_y) > 0)
                    {
                        cur_bbox_vec[i].x += moved_x + 0.5;
                        cur_bbox_vec[i].y += moved_y + 0.5;