C++Yolov4目标检测实战

news/2024/11/14 11:56:27/

Introduction

  今年2月份，Yolo之父Joseph Redmon由于Yolo被用于军事和隐私窥探退出CV界表示抗议，就当我们以为Yolo系列就此终结的时候，4月24日，Yolov4横空出世，新的接棒者出现，而一作正是赫赫有名的AB大神。
  paper github
  在本篇文章里，我们先不急去探究Yolov4的原理，而是从工程的角度来使用Yolov4。首先我们来看一下，Yolov4的性能有多么强劲，下面是使用不同显卡的时候，主流目标检测器的性能，从下图我们发现，Yolov4真的比自己的前辈Yolov3强劲了很多。
yolov4和其他主流目标检测器性能对比

官方darknet

darknetAB
  官方工程的使用在AB大神的github里面已经讲的非常清楚，可以实现Yolov4网络的训练，检测等功能。但是使用官方darknet项目我们很难直接进行定制化的项目开发，因此本文利用官方项目提供的C++接口进行目标检测实战。

Yolov4实战

  首先我们从github下载yolov4.weights,不能上外网的同学下载应该很慢，我上传到百度云盘里，方便大家下载，链接如下：
链接: https://pan.baidu.com/s/14U9pkxJE3MHYj7KCWHnkNw 提取码: 54v4
  工程包括两个文件main.cpp, yolo_v2_class.hpp.我把自己的整个工程上传到github上，同学们可以直接去git clone下来，觉得有用的同学麻烦star一下，谢谢。

main.cpp

#include <iostream>
#include <iomanip>
#include <string>
#include <vector>
#include <queue>
#include <fstream>
#include <thread>
#include <future>
#include <atomic>
#include <mutex>         // std::mutex, std::unique_lock
#include <cmath>// It makes sense only for video-Camera (not for video-File)
// To use - uncomment the following line. Optical-flow is supported only by OpenCV 3.x - 4.x
//#define TRACK_OPTFLOW
//#define GPU// To use 3D-stereo camera ZED - uncomment the following line. ZED_SDK should be installed.
//#define ZED_STEREO#include "yolo_v2_class.hpp"    // imported functions from DLL#ifdef OPENCV
#ifdef ZED_STEREO
#include <sl/Camera.hpp>
#if ZED_SDK_MAJOR_VERSION == 2
#define ZED_STEREO_2_COMPAT_MODE
#endif#undef GPU // avoid conflict with sl::MEM::GPU#ifdef ZED_STEREO_2_COMPAT_MODE
#pragma comment(lib, "sl_core64.lib")
#pragma comment(lib, "sl_input64.lib")
#endif
#pragma comment(lib, "sl_zed64.lib")float getMedian(std::vector<float> &v) {size_t n = v.size() / 2;std::nth_element(v.begin(), v.begin() + n, v.end());return v[n];
}std::vector<bbox_t> get_3d_coordinates(std::vector<bbox_t> bbox_vect, cv::Mat xyzrgba)
{bool valid_measure;int i, j;const unsigned int R_max_global = 10;std::vector<bbox_t> bbox3d_vect;for (auto &cur_box : bbox_vect) {const unsigned int obj_size = std::min(cur_box.w, cur_box.h);const unsigned int R_max = std::min(R_max_global, obj_size / 2);int center_i = cur_box.x + cur_box.w * 0.5f, center_j = cur_box.y + cur_box.h * 0.5f;std::vector<float> x_vect, y_vect, z_vect;for (int R = 0; R < R_max; R++) {for (int y = -R; y <= R; y++) {for (int x = -R; x <= R; x++) {i = center_i + x;j = center_j + y;sl::float4 out(NAN, NAN, NAN, NAN);if (i >= 0 && i < xyzrgba.cols && j >= 0 && j < xyzrgba.rows) {cv::Vec4f &elem = xyzrgba.at<cv::Vec4f>(j, i);  // x,y,z,wout.x = elem[0];out.y = elem[1];out.z = elem[2];out.w = elem[3];}valid_measure = std::isfinite(out.z);if (valid_measure){x_vect.push_back(out.x);y_vect.push_back(out.y);z_vect.push_back(out.z);}}}}if (x_vect.size() * y_vect.size() * z_vect.size() > 0){cur_box.x_3d = getMedian(x_vect);cur_box.y_3d = getMedian(y_vect);cur_box.z_3d = getMedian(z_vect);}else {cur_box.x_3d = NAN;cur_box.y_3d = NAN;cur_box.z_3d = NAN;}bbox3d_vect.emplace_back(cur_box);}return bbox3d_vect;
}cv::Mat slMat2cvMat(sl::Mat &input) {int cv_type = -1; // Mapping between MAT_TYPE and CV_TYPEif(input.getDataType() ==
#ifdef ZED_STEREO_2_COMPAT_MODEsl::MAT_TYPE_32F_C4
#elsesl::MAT_TYPE::F32_C4
#endif) {cv_type = CV_32FC4;} else cv_type = CV_8UC4; // sl::Mat used are either RGBA images or XYZ (4C) point cloudsreturn cv::Mat(input.getHeight(), input.getWidth(), cv_type, input.getPtr<sl::uchar1>(
#ifdef ZED_STEREO_2_COMPAT_MODEsl::MEM::MEM_CPU
#elsesl::MEM::CPU
#endif));
}cv::Mat zed_capture_rgb(sl::Camera &zed) {sl::Mat left;zed.retrieveImage(left);cv::Mat left_rgb;cv::cvtColor(slMat2cvMat(left), left_rgb, CV_RGBA2RGB);return left_rgb;
}cv::Mat zed_capture_3d(sl::Camera &zed) {sl::Mat cur_cloud;zed.retrieveMeasure(cur_cloud,
#ifdef ZED_STEREO_2_COMPAT_MODEsl::MEASURE_XYZ
#elsesl::MEASURE::XYZ
#endif);return slMat2cvMat(cur_cloud).clone();
}static sl::Camera zed; // ZED-camera#else   // ZED_STEREO
std::vector<bbox_t> get_3d_coordinates(std::vector<bbox_t> bbox_vect, cv::Mat xyzrgba) {return bbox_vect;
}
#endif  // ZED_STEREO#include <opencv2/opencv.hpp>            // C++
#include <opencv2/core/version.hpp>
#ifndef CV_VERSION_EPOCH     // OpenCV 3.x and 4.x
#include <opencv2/videoio/videoio.hpp>
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_MAJOR)"" CVAUX_STR(CV_VERSION_MINOR)"" CVAUX_STR(CV_VERSION_REVISION)
#ifndef USE_CMAKE_LIBS
#pragma comment(lib, "opencv_world" OPENCV_VERSION ".lib")
#ifdef TRACK_OPTFLOW
/*
#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")
*/
#endif    // TRACK_OPTFLOW
#endif    // USE_CMAKE_LIBS
#else     // OpenCV 2.x
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_EPOCH)"" CVAUX_STR(CV_VERSION_MAJOR)"" CVAUX_STR(CV_VERSION_MINOR)
#ifndef USE_CMAKE_LIBS
#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_video" OPENCV_VERSION ".lib")
#endif    // USE_CMAKE_LIBS
#endif    // CV_VERSION_EPOCHusing namespace std;
vector<string> split(const string&s,char sepeartor)
{vector<string> split_vector;int subinit=0;for (int id=0;id!=s.length();id++){if (s[id]==sepeartor){split_vector.push_back(s.substr(subinit,id-subinit));subinit=id+1;}}split_vector.push_back(s.substr(subinit,s.length()-subinit));return split_vector;
}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)
{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 = obj_id_to_color(i.obj_id);cv::rectangle(mat_img, cv::Rect(i.x, i.y, i.w, i.h), color, 2);if (obj_names.size() > i.obj_id) {std::string obj_name = obj_names[i.obj_id];if (i.track_id > 0) obj_name += " - " + std::to_string(i.track_id);cv::Size const text_size = getTextSize(obj_name, cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, 2, 0);int max_width = (text_size.width > i.w + 2) ? text_size.width : (i.w + 2);max_width = std::max(max_width, (int)i.w + 2);//max_width = std::max(max_width, 283);std::string coords_3d;if (!std::isnan(i.z_3d)) {std::stringstream ss;ss << std::fixed << std::setprecision(2) << "x:" << i.x_3d << "m y:" << i.y_3d << "m z:" << i.z_3d << "m ";coords_3d = ss.str();cv::Size const text_size_3d = getTextSize(ss.str(), cv::FONT_HERSHEY_COMPLEX_SMALL, 0.8, 1, 0);int const max_width_3d = (text_size_3d.width > i.w + 2) ? text_size_3d.width : (i.w + 2);if (max_width_3d > max_width) max_width = max_width_3d;}cv::rectangle(mat_img, cv::Point2f(std::max((int)i.x - 1, 0), std::max((int)i.y - 35, 0)),cv::Point2f(std::min((int)i.x + max_width, mat_img.cols - 1), std::min((int)i.y, mat_img.rows - 1)),color, CV_FILLED, 8, 0);putText(mat_img, obj_name, cv::Point2f(i.x, i.y - 16), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(0, 0, 0), 2);if(!coords_3d.empty()) putText(mat_img, coords_3d, cv::Point2f(i.x, i.y-1), cv::FONT_HERSHEY_COMPLEX_SMALL, 0.8, cv::Scalar(0, 0, 0), 1);}}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);}
}
#endif    // OPENCVvoid show_console_result(std::vector<bbox_t> const result_vec, std::vector<std::string> const obj_names, int frame_id = -1) {if (frame_id >= 0) std::cout << " Frame: " << frame_id << std::endl;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<< ", w = " << i.w << ", h = " << i.h<< std::setprecision(3) << ", prob = " << i.prob << std::endl;}
}std::vector<std::string> objects_names_from_file(std::string const filename) {std::ifstream file(filename);std::vector<std::string> file_lines;if (!file.is_open()) return file_lines;for(std::string line; getline(file, line);) file_lines.push_back(line);std::cout << "object names loaded \n";return file_lines;
}template<typename T>
class send_one_replaceable_object_t {const bool sync;std::atomic<T *> a_ptr;
public:void send(T const& _obj) {T *new_ptr = new T;*new_ptr = _obj;if (sync) {while (a_ptr.load()) std::this_thread::sleep_for(std::chrono::milliseconds(3));}std::unique_ptr<T> old_ptr(a_ptr.exchange(new_ptr));}T receive() {std::unique_ptr<T> ptr;do {while(!a_ptr.load()) std::this_thread::sleep_for(std::chrono::milliseconds(3));ptr.reset(a_ptr.exchange(NULL));} while (!ptr);T obj = *ptr;return obj;}bool is_object_present() {return (a_ptr.load() != NULL);}send_one_replaceable_object_t(bool _sync) : sync(_sync), a_ptr(NULL){}
};int main(int argc, char *argv[])
{std::string  names_file = "coco.names";std::string  cfg_file = "cfg/yolov4.cfg";std::string  weights_file = "yolov4.weights";std::string filename;if (argc > 4) {    //voc.names yolo-voc.cfg yolo-voc.weights test.mp4names_file = argv[1];cfg_file = argv[2];weights_file = argv[3];filename = argv[4];}else if (argc > 1) filename = argv[1];float const thresh = (argc > 5) ? std::stof(argv[5]) : 0.2;Detector detector(cfg_file, weights_file);auto obj_names = objects_names_from_file(names_file);std::string out_videofile = "result.avi";bool const save_output_videofile = false;   // true - for historybool const send_network = false;        // true - for remote detectionbool const use_kalman_filter = false;   // true - for stationary camerabool detection_sync = true;             // true - for video-file
#ifdef TRACK_OPTFLOW    // for slow GPUdetection_sync = false;Tracker_optflow tracker_flow;//detector.wait_stream = true;
#endif  // TRACK_OPTFLOWwhile (true){std::cout << "input image or video filename: ";if(filename.size() == 0) std::cin >> filename;if (filename.size() == 0) break;try {
#ifdef OPENCVpreview_boxes_t large_preview(100, 150, false), small_preview(50, 50, true);bool show_small_boxes = false;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 fileprotocol == "rtmp://" || protocol == "rtsp://" || protocol == "http://" || protocol == "https:/" ||    // video network streamfilename == "zed_camera" || file_ext == "svo" || filename == "web_camera")   // ZED stereo camera{if (protocol == "rtsp://" || protocol == "http://" || protocol == "https:/" || filename == "zed_camera" || filename == "web_camera")detection_sync = false;cv::Mat cur_frame;std::atomic<int> fps_cap_counter(0), fps_det_counter(0);std::atomic<int> current_fps_cap(0), current_fps_det(0);std::atomic<bool> exit_flag(false);std::chrono::steady_clock::time_point steady_start, steady_end;int video_fps = 25;bool use_zed_camera = false;track_kalman_t track_kalman;#ifdef ZED_STEREOsl::InitParameters init_params;init_params.depth_minimum_distance = 0.5;#ifdef ZED_STEREO_2_COMPAT_MODEinit_params.depth_mode = sl::DEPTH_MODE_ULTRA;init_params.camera_resolution = sl::RESOLUTION_HD720;// sl::RESOLUTION_HD1080, sl::RESOLUTION_HD720init_params.coordinate_units = sl::UNIT_METER;init_params.camera_buffer_count_linux = 2;if (file_ext == "svo") init_params.svo_input_filename.set(filename.c_str());#elseinit_params.depth_mode = sl::DEPTH_MODE::ULTRA;init_params.camera_resolution = sl::RESOLUTION::HD720;// sl::RESOLUTION::HD1080, sl::RESOLUTION::HD720init_params.coordinate_units = sl::UNIT::METER;if (file_ext == "svo") init_params.input.setFromSVOFile(filename.c_str());#endif//init_params.sdk_cuda_ctx = (CUcontext)detector.get_cuda_context();init_params.sdk_gpu_id = detector.cur_gpu_id;if (filename == "zed_camera" || file_ext == "svo") {std::cout << "ZED 3D Camera " << zed.open(init_params) << std::endl;if (!zed.isOpened()) {std::cout << " Error: ZED Camera should be connected to USB 3.0. And ZED_SDK should be installed. \n";getchar();return 0;}cur_frame = zed_capture_rgb(zed);use_zed_camera = true;}
#endif  // ZED_STEREOcv::VideoCapture cap;if (filename == "web_camera") {cap.open(0);cap >> cur_frame;} else if (!use_zed_camera) {cap.open(filename);cap >> cur_frame;}
#ifdef CV_VERSION_EPOCH // OpenCV 2.xvideo_fps = cap.get(CV_CAP_PROP_FPS);
#elsevideo_fps = cap.get(cv::CAP_PROP_FPS);
#endifcv::Size const frame_size = cur_frame.size();//cv::Size const frame_size(cap.get(CV_CAP_PROP_FRAME_WIDTH), cap.get(CV_CAP_PROP_FRAME_HEIGHT));std::cout << "\n Video size: " << frame_size << std::endl;cv::VideoWriter output_video;if (save_output_videofile)
#ifdef CV_VERSION_EPOCH // OpenCV 2.xoutput_video.open(out_videofile, CV_FOURCC('D', 'I', 'V', 'X'), std::max(35, video_fps), frame_size, true);
#elseoutput_video.open(out_videofile, cv::VideoWriter::fourcc('D', 'I', 'V', 'X'), std::max(35, video_fps), frame_size, true);
#endifstruct detection_data_t {cv::Mat cap_frame;std::shared_ptr<image_t> det_image;std::vector<bbox_t> result_vec;cv::Mat draw_frame;bool new_detection;uint64_t frame_id;bool exit_flag;cv::Mat zed_cloud;std::queue<cv::Mat> track_optflow_queue;detection_data_t() : exit_flag(false), new_detection(false) {}};const bool sync = detection_sync; // sync data exchangesend_one_replaceable_object_t<detection_data_t> cap2prepare(sync), cap2draw(sync),prepare2detect(sync), detect2draw(sync), draw2show(sync), draw2write(sync), draw2net(sync);std::thread t_cap, t_prepare, t_detect, t_post, t_draw, t_write, t_network;// capture new video-frameif (t_cap.joinable()) t_cap.join();t_cap = std::thread([&](){uint64_t frame_id = 0;detection_data_t detection_data;do {detection_data = detection_data_t();
#ifdef ZED_STEREOif (use_zed_camera) {while (zed.grab() !=#ifdef ZED_STEREO_2_COMPAT_MODEsl::SUCCESS#elsesl::ERROR_CODE::SUCCESS#endif) std::this_thread::sleep_for(std::chrono::milliseconds(2));detection_data.cap_frame = zed_capture_rgb(zed);detection_data.zed_cloud = zed_capture_3d(zed);}else
#endif   // ZED_STEREO{cap >> detection_data.cap_frame;}fps_cap_counter++;detection_data.frame_id = frame_id++;if (detection_data.cap_frame.empty() || exit_flag) {std::cout << " exit_flag: detection_data.cap_frame.size = " << detection_data.cap_frame.size() << std::endl;detection_data.exit_flag = true;detection_data.cap_frame = cv::Mat(frame_size, CV_8UC3);}if (!detection_sync) {cap2draw.send(detection_data);       // skip detection}cap2prepare.send(detection_data);} while (!detection_data.exit_flag);std::cout << " t_cap exit \n";});// pre-processing video frame (resize, convertion)t_prepare = std::thread([&](){std::shared_ptr<image_t> det_image;detection_data_t detection_data;do {detection_data = cap2prepare.receive();det_image = detector.mat_to_image_resize(detection_data.cap_frame);detection_data.det_image = det_image;prepare2detect.send(detection_data);    // detection} while (!detection_data.exit_flag);std::cout << " t_prepare exit \n";});// detection by Yoloif (t_detect.joinable()) t_detect.join();t_detect = std::thread([&](){std::shared_ptr<image_t> det_image;detection_data_t detection_data;do {detection_data = prepare2detect.receive();det_image = detection_data.det_image;std::vector<bbox_t> result_vec;if(det_image)result_vec = detector.detect_resized(*det_image, frame_size.width, frame_size.height, thresh, true);  // truefps_det_counter++;//std::this_thread::sleep_for(std::chrono::milliseconds(150));detection_data.new_detection = true;detection_data.result_vec = result_vec;detect2draw.send(detection_data);} while (!detection_data.exit_flag);std::cout << " t_detect exit \n";});// draw rectangles (and track objects)t_draw = std::thread([&](){std::queue<cv::Mat> track_optflow_queue;detection_data_t detection_data;do {// for Video-fileif (detection_sync) {detection_data = detect2draw.receive();}// for Video-cameraelse{// get new Detection result if presentif (detect2draw.is_object_present()) {cv::Mat old_cap_frame = detection_data.cap_frame;   // use old captured framedetection_data = detect2draw.receive();if (!old_cap_frame.empty()) detection_data.cap_frame = old_cap_frame;}// get new Captured frameelse {std::vector<bbox_t> old_result_vec = detection_data.result_vec; // use old detectionsdetection_data = cap2draw.receive();detection_data.result_vec = old_result_vec;}}cv::Mat cap_frame = detection_data.cap_frame;cv::Mat draw_frame = detection_data.cap_frame.clone();std::vector<bbox_t> result_vec = detection_data.result_vec;#ifdef TRACK_OPTFLOWif (detection_data.new_detection) {tracker_flow.update_tracking_flow(detection_data.cap_frame, detection_data.result_vec);while (track_optflow_queue.size() > 0) {draw_frame = track_optflow_queue.back();result_vec = tracker_flow.tracking_flow(track_optflow_queue.front(), false);track_optflow_queue.pop();}}else {track_optflow_queue.push(cap_frame);result_vec = tracker_flow.tracking_flow(cap_frame, false);}detection_data.new_detection = true;    // to correct kalman filter
#endif //TRACK_OPTFLOW// track ID by using kalman filterif (use_kalman_filter) {if (detection_data.new_detection) {result_vec = track_kalman.correct(result_vec);}else {result_vec = track_kalman.predict();}}// track ID by using custom functionelse {int frame_story = std::max(5, current_fps_cap.load());result_vec = detector.tracking_id(result_vec, true, frame_story, 40);}if (use_zed_camera && !detection_data.zed_cloud.empty()) {result_vec = get_3d_coordinates(result_vec, detection_data.zed_cloud);}//small_preview.set(draw_frame, result_vec);//large_preview.set(draw_frame, result_vec);draw_boxes(draw_frame, result_vec, obj_names, current_fps_det, current_fps_cap);//show_console_result(result_vec, obj_names, detection_data.frame_id);//large_preview.draw(draw_frame);//small_preview.draw(draw_frame, true);detection_data.result_vec = result_vec;detection_data.draw_frame = draw_frame;draw2show.send(detection_data);if (send_network) draw2net.send(detection_data);if (output_video.isOpened()) draw2write.send(detection_data);} while (!detection_data.exit_flag);std::cout << " t_draw exit \n";});// write frame to videofilet_write = std::thread([&](){if (output_video.isOpened()) {detection_data_t detection_data;cv::Mat output_frame;do {detection_data = draw2write.receive();if(detection_data.draw_frame.channels() == 4) cv::cvtColor(detection_data.draw_frame, output_frame, CV_RGBA2RGB);else output_frame = detection_data.draw_frame;output_video << output_frame;} while (!detection_data.exit_flag);output_video.release();}std::cout << " t_write exit \n";});// send detection to the networkt_network = std::thread([&](){if (send_network) {detection_data_t detection_data;do {detection_data = draw2net.receive();detector.send_json_http(detection_data.result_vec, obj_names, detection_data.frame_id, filename);} while (!detection_data.exit_flag);}std::cout << " t_network exit \n";});// show detectiondetection_data_t detection_data;do {steady_end = std::chrono::steady_clock::now();float time_sec = std::chrono::duration<double>(steady_end - steady_start).count();if (time_sec >= 1) {current_fps_det = fps_det_counter.load() / time_sec;current_fps_cap = fps_cap_counter.load() / time_sec;steady_start = steady_end;fps_det_counter = 0;fps_cap_counter = 0;}detection_data = draw2show.receive();cv::Mat draw_frame = detection_data.draw_frame;//if (extrapolate_flag) {//    cv::putText(draw_frame, "extrapolate", cv::Point2f(10, 40), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.0, cv::Scalar(50, 50, 0), 2);//}cv::imshow("window name", draw_frame);filename.replace(filename.end()-4, filename.end(), "_yolov4_out.jpg");int key = cv::waitKey(3);    // 3 or 16msif (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 (key == 27) { exit_flag = true;}//std::cout << " current_fps_det = " << current_fps_det << ", current_fps_cap = " << current_fps_cap << std::endl;} while (!detection_data.exit_flag);std::cout << " show detection exit \n";cv::destroyWindow("window name");// wait for all threadsif (t_cap.joinable()) t_cap.join();if (t_prepare.joinable()) t_prepare.join();if (t_detect.joinable()) t_detect.join();if (t_post.joinable()) t_post.join();if (t_draw.joinable()) t_draw.join();if (t_write.joinable()) t_write.join();if (t_network.joinable()) t_network.join();break;}else if (file_ext == "txt") {    // list of image filesstd::ifstream file(filename);if (!file.is_open()) std::cout << "File not found! \n";elsefor (std::string line; file >> line;) {std::cout << line << std::endl;cv::Mat mat_img = cv::imread(line);std::vector<bbox_t> result_vec = detector.detect(mat_img);show_console_result(result_vec, obj_names);//draw_boxes(mat_img, result_vec, obj_names);//cv::imwrite("res_" + line, mat_img);}}else {    // image file// to achive high performance for multiple images do these 2 lines in another threadcv::Mat mat_img = cv::imread(filename);auto det_image = detector.mat_to_image_resize(mat_img);auto start = std::chrono::steady_clock::now();std::vector<bbox_t> result_vec = detector.detect_resized(*det_image, mat_img.size().width, mat_img.size().height);auto end = std::chrono::steady_clock::now();std::chrono::duration<double> spent = end - start;std::cout << " Time: " << spent.count() << " sec \n";//result_vec = detector.tracking_id(result_vec);    // comment it - if track_id is not requireddraw_boxes(mat_img, result_vec, obj_names);cv::imshow("window name", mat_img);vector<string> filenamesplit=split(filename,'/');string endname=filenamesplit[filenamesplit.size()-1];endname.replace(endname.end()-4,endname.end(),"_yolov4_out.jpg");std::string outputfile="detect_result/"+endname;imwrite(outputfile, mat_img);show_console_result(result_vec, obj_names);cv::waitKey(0);}
#else   // OPENCV//std::vector<bbox_t> result_vec = detector.detect(filename);auto img = detector.load_image(filename);std::vector<bbox_t> result_vec = detector.detect(img);detector.free_image(img);show_console_result(result_vec, obj_names);
#endif  // OPENCV}catch (std::exception &e) { std::cerr << "exception: " << e.what() << "\n"; getchar(); }catch (...) { std::cerr << "unknown exception \n"; getchar(); }filename.clear();}return 0;
}

yolo_v2_class.hpp直接使用github的代码
Cmakelist.txt如下：

cmake_minimum_required(VERSION 3.5)
project(yolov4)
find_package( OpenCV 3 REQUIRED )
set(CMAKE_CXX_STANDARD 14)
include_directories(${OpenCV_INCLUDE_DIRS})
add_executable(yolov4 main.cpp yolo_v2_class.hpp)
target_link_libraries(yolov4 ${OpenCV_LIBS} libdarknet.so libpthread.so.0)

  注意project的yolov4改成自己工程的名字，另外需要链接两个动态库，第一个库是darknet的所有网络层的定义，第二个库是linux的多线程库。两个库都放在/usr/lib目录下，因为这个目录是可以被C++自动检索到的。
  libpthread.so.0默认情况下在/lib/x86_64-linux-gnu下，因此可以自动被检索到，只需要把它写在Cmakelist里面就可以。

/usr/bin/ld: CMakeFiles/yolov4.dir/main.cpp.o: undefined reference to symbol 'pthread_create@@GLIBC_2.2.5'
//lib/x86_64-linux-gnu/libpthread.so.0: error adding symbols: DSO missing from command line
collect2: error: ld returned 1 exit status
CMakeFiles/yolov4.dir/build.make:111: recipe for target 'yolov4' failed
make[2]: *** [yolov4] Error 1
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/yolov4.dir/all' failed
make[1]: *** [CMakeFiles/yolov4.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

  libdarknet.so则可以通过编译AB大神的darknet工程得到，修改工程的makefile如下：

GPU=0
CUDNN=0
CUDNN_HALF=0
OPENCV=1
AVX=0
OPENMP=1
LIBSO=1
ZED_CAMERA=0 # ZED SDK 3.0 and above
ZED_CAMERA_v2_8=0 # ZED SDK 2.X

  然后make，编译成功以后，就在目录里就出现了libdarknet.so动态库，将其copy至/usr/lib文件夹。
  切换到我们的Yolov4文件夹，然后make，通过编译以后，可以运行可执行程序。

./yolov4 image/kite.jpg     //检测图片
./yolov4 test.mp4             //检测视频
./yolov4 web_camera      //检测摄像头

mini_batch = 1, batch = 1, time_steps = 1, train = 0 layer   filters  size/strd(dil)      input                output0 conv     32       3 x 3/ 1    576 x 576 x   3 ->  576 x 576 x  32 0.573 BF1 conv     64       3 x 3/ 2    576 x 576 x  32 ->  288 x 288 x  64 3.058 BF2 conv     64       1 x 1/ 1    288 x 288 x  64 ->  288 x 288 x  64 0.679 BF3 route  1 		                           ->  288 x 288 x  64 4 conv     64       1 x 1/ 1    288 x 288 x  64 ->  288 x 288 x  64 0.679 BF5 conv     32       1 x 1/ 1    288 x 288 x  64 ->  288 x 288 x  32 0.340 BF6 conv     64       3 x 3/ 1    288 x 288 x  32 ->  288 x 288 x  64 3.058 BF7 Shortcut Layer: 4,  wt = 0, wn = 0, outputs: 288 x 288 x  64 0.005 BF8 conv     64       1 x 1/ 1    288 x 288 x  64 ->  288 x 288 x  64 0.679 BF9 route  8 2 	                           ->  288 x 288 x 128 10 conv     64       1 x 1/ 1    288 x 288 x 128 ->  288 x 288 x  64 1.359 BF11 conv    128       3 x 3/ 2    288 x 288 x  64 ->  144 x 144 x 128 3.058 BF12 conv     64       1 x 1/ 1    144 x 144 x 128 ->  144 x 144 x  64 0.340 BF13 route  11 		                           ->  144 x 144 x 128 14 conv     64       1 x 1/ 1    144 x 144 x 128 ->  144 x 144 x  64 0.340 BF15 conv     64       1 x 1/ 1    144 x 144 x  64 ->  144 x 144 x  64 0.170 BF16 conv     64       3 x 3/ 1    144 x 144 x  64 ->  144 x 144 x  64 1.529 BF17 Shortcut Layer: 14,  wt = 0, wn = 0, outputs: 144 x 144 x  64 0.001 BF18 conv     64       1 x 1/ 1    144 x 144 x  64 ->  144 x 144 x  64 0.170 BF19 conv     64       3 x 3/ 1    144 x 144 x  64 ->  144 x 144 x  64 1.529 BF20 Shortcut Layer: 17,  wt = 0, wn = 0, outputs: 144 x 144 x  64 0.001 BF21 conv     64       1 x 1/ 1    144 x 144 x  64 ->  144 x 144 x  64 0.170 BF22 route  21 12 	                           ->  144 x 144 x 128 23 conv    128       1 x 1/ 1    144 x 144 x 128 ->  144 x 144 x 128 0.679 BF24 conv    256       3 x 3/ 2    144 x 144 x 128 ->   72 x  72 x 256 3.058 BF25 conv    128       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 128 0.340 BF26 route  24 		                           ->   72 x  72 x 256 27 conv    128       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 128 0.340 BF28 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF29 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF30 Shortcut Layer: 27,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF31 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF32 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF33 Shortcut Layer: 30,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF34 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF35 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF36 Shortcut Layer: 33,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF37 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF38 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF39 Shortcut Layer: 36,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF40 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF41 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF42 Shortcut Layer: 39,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF43 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF44 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF45 Shortcut Layer: 42,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF46 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF47 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF48 Shortcut Layer: 45,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF49 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF50 conv    128       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 128 1.529 BF51 Shortcut Layer: 48,  wt = 0, wn = 0, outputs:  72 x  72 x 128 0.001 BF52 conv    128       1 x 1/ 1     72 x  72 x 128 ->   72 x  72 x 128 0.170 BF53 route  52 25 	                           ->   72 x  72 x 256 54 conv    256       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 256 0.679 BF55 conv    512       3 x 3/ 2     72 x  72 x 256 ->   36 x  36 x 512 3.058 BF56 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF57 route  55 		                           ->   36 x  36 x 512 58 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF59 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF60 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF61 Shortcut Layer: 58,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF62 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF63 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF64 Shortcut Layer: 61,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF65 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF66 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF67 Shortcut Layer: 64,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF68 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF69 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF70 Shortcut Layer: 67,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF71 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF72 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF73 Shortcut Layer: 70,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF74 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF75 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF76 Shortcut Layer: 73,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF77 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF78 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF79 Shortcut Layer: 76,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF80 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF81 conv    256       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 256 1.529 BF82 Shortcut Layer: 79,  wt = 0, wn = 0, outputs:  36 x  36 x 256 0.000 BF83 conv    256       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 256 0.170 BF84 route  83 56 	                           ->   36 x  36 x 512 85 conv    512       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 512 0.679 BF86 conv   1024       3 x 3/ 2     36 x  36 x 512 ->   18 x  18 x1024 3.058 BF87 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF88 route  86 		                           ->   18 x  18 x1024 89 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF90 conv    512       1 x 1/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.170 BF91 conv    512       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x 512 1.529 BF92 Shortcut Layer: 89,  wt = 0, wn = 0, outputs:  18 x  18 x 512 0.000 BF93 conv    512       1 x 1/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.170 BF94 conv    512       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x 512 1.529 BF95 Shortcut Layer: 92,  wt = 0, wn = 0, outputs:  18 x  18 x 512 0.000 BF96 conv    512       1 x 1/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.170 BF97 conv    512       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x 512 1.529 BF98 Shortcut Layer: 95,  wt = 0, wn = 0, outputs:  18 x  18 x 512 0.000 BF99 conv    512       1 x 1/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.170 BF100 conv    512       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x 512 1.529 BF101 Shortcut Layer: 98,  wt = 0, wn = 0, outputs:  18 x  18 x 512 0.000 BF102 conv    512       1 x 1/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.170 BF103 route  102 87 	                           ->   18 x  18 x1024 104 conv   1024       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x1024 0.679 BF105 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF106 conv   1024       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x1024 3.058 BF107 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF108 max                5x 5/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.004 BF109 route  107 		                           ->   18 x  18 x 512 110 max                9x 9/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.013 BF111 route  107 		                           ->   18 x  18 x 512 112 max               13x13/ 1     18 x  18 x 512 ->   18 x  18 x 512 0.028 BF113 route  112 110 108 107 	                   ->   18 x  18 x2048 114 conv    512       1 x 1/ 1     18 x  18 x2048 ->   18 x  18 x 512 0.679 BF115 conv   1024       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x1024 3.058 BF116 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF117 conv    256       1 x 1/ 1     18 x  18 x 512 ->   18 x  18 x 256 0.085 BF118 upsample                 2x    18 x  18 x 256 ->   36 x  36 x 256119 route  85 		                           ->   36 x  36 x 512 120 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF121 route  120 118 	                           ->   36 x  36 x 512 122 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF123 conv    512       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 512 3.058 BF124 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF125 conv    512       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 512 3.058 BF126 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF127 conv    128       1 x 1/ 1     36 x  36 x 256 ->   36 x  36 x 128 0.085 BF128 upsample                 2x    36 x  36 x 128 ->   72 x  72 x 128129 route  54 		                           ->   72 x  72 x 256 130 conv    128       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 128 0.340 BF131 route  130 128 	                           ->   72 x  72 x 256 132 conv    128       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 128 0.340 BF133 conv    256       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 256 3.058 BF134 conv    128       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 128 0.340 BF135 conv    256       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 256 3.058 BF136 conv    128       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 128 0.340 BF137 conv    256       3 x 3/ 1     72 x  72 x 128 ->   72 x  72 x 256 3.058 BF138 conv    255       1 x 1/ 1     72 x  72 x 256 ->   72 x  72 x 255 0.677 BF139 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, cls_norm: 1.00, scale_x_y: 1.20
nms_kind: greedynms (1), beta = 0.600000 140 route  136 		                           ->   72 x  72 x 128 141 conv    256       3 x 3/ 2     72 x  72 x 128 ->   36 x  36 x 256 0.764 BF142 route  141 126 	                           ->   36 x  36 x 512 143 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF144 conv    512       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 512 3.058 BF145 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF146 conv    512       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 512 3.058 BF147 conv    256       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 256 0.340 BF148 conv    512       3 x 3/ 1     36 x  36 x 256 ->   36 x  36 x 512 3.058 BF149 conv    255       1 x 1/ 1     36 x  36 x 512 ->   36 x  36 x 255 0.338 BF150 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, cls_norm: 1.00, scale_x_y: 1.10
nms_kind: greedynms (1), beta = 0.600000 151 route  147 		                           ->   36 x  36 x 256 152 conv    512       3 x 3/ 2     36 x  36 x 256 ->   18 x  18 x 512 0.764 BF153 route  152 116 	                           ->   18 x  18 x1024 154 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF155 conv   1024       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x1024 3.058 BF156 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF157 conv   1024       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x1024 3.058 BF158 conv    512       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 512 0.340 BF159 conv   1024       3 x 3/ 1     18 x  18 x 512 ->   18 x  18 x1024 3.058 BF160 conv    255       1 x 1/ 1     18 x  18 x1024 ->   18 x  18 x 255 0.169 BF161 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, cls_norm: 1.00, scale_x_y: 1.05
nms_kind: greedynms (1), beta = 0.600000 
Total BFLOPS 115.293 
avg_outputs = 958892 
Loading weights from yolov4.weights...seen 64, trained: 32032 K-images (500 Kilo-batches_64) 
Done! Loaded 162 layers from weights-file 
object names loaded 
input image or video filename:  Time: 7.29136 sec 
surfboard - obj_id = 37,  x = 521, y = 517, w = 32, h = 15, prob = 0.393
surfboard - obj_id = 37,  x = 497, y = 521, w = 39, h = 10, prob = 0.255
surfboard - obj_id = 37,  x = 814, y = 567, w = 30, h = 10, prob = 0.248
kite - obj_id = 33,  x = 591, y = 79, w = 75, h = 71, prob = 0.994
kite - obj_id = 33,  x = 279, y = 235, w = 23, h = 45, prob = 0.979
kite - obj_id = 33,  x = 575, y = 343, w = 25, h = 25, prob = 0.951
kite - obj_id = 33,  x = 1082, y = 393, w = 14, h = 28, prob = 0.943
kite - obj_id = 33,  x = 464, y = 339, w = 16, h = 18, prob = 0.855
kite - obj_id = 33,  x = 300, y = 375, w = 23, h = 32, prob = 0.68
kite - obj_id = 33,  x = 760, y = 379, w = 7, h = 9, prob = 0.634
person - obj_id = 0,  x = 110, y = 610, w = 51, h = 151, prob = 0.994
person - obj_id = 0,  x = 213, y = 698, w = 53, h = 159, prob = 0.993
person - obj_id = 0,  x = 1204, y = 450, w = 9, h = 12, prob = 0.872
person - obj_id = 0,  x = 37, y = 509, w = 16, h = 51, prob = 0.871
person - obj_id = 0,  x = 345, y = 487, w = 9, h = 14, prob = 0.866
person - obj_id = 0,  x = 176, y = 539, w = 11, h = 32, prob = 0.832
person - obj_id = 0,  x = 21, y = 529, w = 14, h = 26, prob = 0.801
person - obj_id = 0,  x = 82, y = 506, w = 25, h = 57, prob = 0.697
person - obj_id = 0,  x = 518, y = 506, w = 16, h = 18, prob = 0.606
person - obj_id = 0,  x = 692, y = 462, w = 7, h = 6, prob = 0.552
person - obj_id = 0,  x = 460, y = 471, w = 7, h = 6, prob = 0.394
person - obj_id = 0,  x = 537, y = 514, w = 14, h = 17, prob = 0.381

测试图片结果
kite_yolov4_out.jpg
person_yolov4_out.jpg
supermarket_yolov4_out.jpg

总结

  在本文中，我们使用C++调用了作者在COCO数据集上的训练结果进行了图片测试，并且可以进行视频测试和网络摄像头测试（自己笔记本显卡太弱，视频跑不起来），如果我们想要自己训练数据集，仍然可以参考github，可以通过训练得到weights文件，然后按照本文所讲的进行测试。


http://www.ppmy.cn/news/674378.html

相关文章

YOLO系列（v1~v3）的学习及YOLO-Fastest在海思平台的部署（下）

YOLO系列（v1~v3）的学习及YOLO-Fastest在海思平台的部署（下）

YOLO系列&#xff08;v1~v3&#xff09;的学习及YOLO-Fastest在海思平台的部署&#xff08;上&#xff09; YOLO系列&#xff08;v1~v3&#xff09;的学习及YOLO-Fastest在海思平台的部署&#xff08;中&#xff09; YOLO系列&#xff08;v1~v3&#xff09;的学习及YOLO-Fastes…
阅读更多...
YOLOV4 论文原理 模型分析 win10 vs2015 cuda9 opencv3.3 代码测试 网盘权重yolov4.conv.137 yolov4.weight下载

YOLOV4 论文原理 模型分析 win10 vs2015 cuda9 opencv3.3 代码测试 网盘权重yolov4.conv.137 yolov4.weight下载

提出目标&#xff1a;相比低计算量(BFLOP)&#xff0c;更着重于优化并行计算&#xff0c;在production system中实现快速计算。 YOLOv4的作者阵容里并没有Joe Redmon, 一作为俄罗斯 Alexey Bochkovskiy &#xff0c;是 YOLO 的 windows 版本github的作者。 并得到YOLO官方githu…
阅读更多...
DarkNet框架下YOLO-v4算法推理过程的特征图可视化

DarkNet框架下YOLO-v4算法推理过程的特征图可视化

说明 本篇文章针对AB版DarkNet源代码进行了修改&#xff0c;添加了一些函数&#xff0c;进行全通道特征图可视化。主要针对YOLO-v4推理过程中&#xff0c;对中间计算结果的特征图数据进行转换&#xff0c;将转换结果用表示成图片进行保存。 DarkNet源代码修改 在network_ker…
阅读更多...
Ubuntu18.04下安装NVIDIA显卡驱动、docker、nvidia-docker；容器中编译安装opencv-4.4.0与darknet-yolov4并完成测试；容器封装镜像转移。2022

Ubuntu18.04下安装NVIDIA显卡驱动、docker、nvidia-docker；容器中编译安装opencv-4.4.0与darknet-yolov4并完成测试；容器封装镜像转移。2022

记录一下第一次在CSDN发博客&#xff0c;欢迎大家光临~ 文章目录 前言一、宿主机配置1.安装Ubunntu18.04 64位系统2.为宿主机系统更换国内软件源Ubuntu 官方源服务器在欧洲&#xff0c;国内访问很慢。所以这里有必要将软件源更换为国内的源,2.1-复制以下内容2.2-备份文件2.3-更…
阅读更多...
Java实现内存分配算法 FF(首次适应算法) BF(最佳适应算法)

Java实现内存分配算法 FF(首次适应算法) BF(最佳适应算法)

一、概述 因为这次os作业对用户在控制台的输入输出有要求&#xff0c;所以我花了挺多的代码来完善控制台的显示。 MemoryAlgorithm类里只是和控制台输入输出有关的操作&#xff0c;而对内存的所有逻辑操作都是用Memory类里对应的方法实现的。 因为不同内存分配算法&#xff0c;…
阅读更多...
Ubuntu18.04 搭建YOLOV4环境

Ubuntu18.04 搭建YOLOV4环境

Darknet是一个轻型的深度学习和训练框架，从这一点上，它和tensorflow以及pytorch这种没有什么不同，特点在轻型二字，它主要对卷集神经网络进行了底层实现，并且主要用于YOLO的目标检测，特点主要有： C语言实现没有依赖项，除了opencv进行视频和UVC摄像头处理容易安装，可移植…
阅读更多...
【darknet】2、yolov4模型训练之模型训练

【darknet】2、yolov4模型训练之模型训练

文章目录 1、进行模型训练数据准备1.1 划分训练和验证集1.2 将数据标注格式转换为YOLO格式 2、修改配置文件2.1 新建cfg/vechle.names2.2 新建cfg/vechle.data 2.3 根据所选模型的不同&#xff0c;设置不同的配置文件2.3.1 新建cfg/yolov4_vechle.cfg2.3.2 新建cfg/yolov4_tiny…
阅读更多...
数据结构之平衡二叉树的平衡因子BF 的计算

数据结构之平衡二叉树的平衡因子BF 的计算

在书上看了平衡二叉树的代码后&#xff0c;发现前人的智慧真是无限。但是因为一次性给出的最完美的代码让人有时候看不太懂... 后来经过仔细推敲&#xff0c;才慢慢发现了其中的奥秘。一开始并不知道关于平衡二叉树的平衡因子BF是怎么修改的&#xff0c;后来才发现关于平衡二叉…
阅读更多...
最新文章

Copyright @ 2022~2023