开源项目(5-2) yolo打包成库

Windows系统下YOLO动态链接库的封装和调用

Windows10+VS2015+OpenCV3.4.1+CUDA8.0+cuDNN8.0

参考教程 https://blog.csdn.net/stjuliet/article/details/87884976

承接上一篇文章所做工作，这篇文章进一步讲述如何将YOLO封装成动态链接库以方便后续目标检测时直接调用。
关于动态链接库的介绍：
https://www.cnblogs.com/chechen/p/8676226.html
https://www.jianshu.com/p/458f87251b3d?tdsourcetag=s_pctim_aiomsg

step1 运行环境和前期准备

与上一篇文章所需环境完全一致，具体可参考：
https://blog.csdn.net/stjuliet/article/details/87731998

配置opecv3.4.1   cuda8.0以及配套cudnn

step2 编译动态链接库

1、下载Darknet源代码：
https://github.com/AlexeyAB/darknet

2、
(1)下载解压后，进入darknet-master->build->darknet目录：

(2)打开yolo_cpp_dll.vcxproj文件，将具有CUDA的版本改成自己使用的版本（默认为10.0），一共有两处，分别在55行和302行

自己电脑装了cuda10和8,这里用8

(3)打开yolo_cpp_dll.sln文件，在属性管理器中配置包含目录、库目录、附加依赖项（和OpenCV环境配置一样），特别注意要将CUDA设备中的Generation改成自己显卡对应的计算能力（默认添加了35和75两项，可能不是你的显卡的计算能力，可以去英伟达显卡官网查询计算能力：https://developer.nvidia.com/cuda-gpus#collapseOne）
，否则接下来的生成会出错。

(4)分别设置Debug/Release - x64，右键项目->生成，成功后在darknet-master\build\darknet\x64目录下找到生成的yolo_cpp_dll.lib和yolo_cpp_dll.dll两个文件。

step3 调用动态链接库

一、至此所有准备工作已经完成，首先将调用所需的所有文件找出来：
1、动态链接库（均在darknet-master\build\darknet\x64目录下）
(1)yolo_cpp_dll.lib
(2)yolo_cpp_dll.dll
(3)pthreadGC2.dll
(4)pthreadVC2.dll

2、OpenCV库（取决于使用debug还是release模式）
(1)opencv_world340d.dll
(2)opencv_world340.dll

如果是扩展库需要

opencv_aruco341.lib
opencv_bgsegm341.lib
opencv_bioinspired341.lib
opencv_calib3d341.lib
opencv_ccalib341.lib
opencv_core341.lib
opencv_cudaarithm341.lib
opencv_cudabgsegm341.lib
opencv_cudacodec341.lib
opencv_cudafeatures2d341.lib
opencv_cudafilters341.lib
opencv_cudaimgproc341.lib
opencv_cudalegacy341.lib
opencv_cudaobjdetect341.lib
opencv_cudaoptflow341.lib
opencv_cudastereo341.lib
opencv_cudawarping341.lib
opencv_cudev341.lib
opencv_datasets341.lib
opencv_dnn341.lib
opencv_dnn_objdetect341.lib
opencv_dpm341.lib
opencv_face341.lib
opencv_features2d341.lib
opencv_flann341.lib
opencv_fuzzy341.lib
opencv_hfs341.lib
opencv_highgui341.lib
opencv_imgcodecs341.lib
opencv_imgproc341.lib
opencv_img_hash341.lib
opencv_line_descriptor341.lib
opencv_ml341.lib
opencv_objdetect341.lib
opencv_optflow341.lib
opencv_phase_unwrapping341.lib
opencv_photo341.lib
opencv_plot341.lib
opencv_reg341.lib
opencv_rgbd341.lib
opencv_saliency341.lib
opencv_shape341.lib
opencv_stereo341.lib
opencv_stitching341.lib
opencv_structured_light341.lib
opencv_superres341.lib
opencv_surface_matching341.lib
opencv_text341.lib
opencv_tracking341.lib
opencv_video341.lib
opencv_videoio341.lib
opencv_videostab341.lib
opencv_xfeatures2d341.lib
opencv_ximgproc341.lib
opencv_xobjdetect341.lib
opencv_xphoto341.lib

　　

3、YOLO模型文件（第一个文件在darknet-master\build\darknet\x64\data目录下，第二个文件在darknet-master\build\darknet\x64目录下，第三个文件需要自己下载，下载链接见前一篇文章）
(1)coco.names
(2)yolov3.cfg
(3)yolov3.weights

4、头文件
(1)yolo_v2_class.hpp
头文件包含了动态链接库中具体的类的定义，调用时需要引用，这个文件在darknet-master\build\darknet目录下的yolo_console_dll.sln中，将其复制到记事本保存成.hpp文件即可。

二、在VS2015中新建一个空项目，在源文件中添加main.cpp，将第一步中所有文件全部放入与main.cpp同路径的文件夹中，并且放入一个目标检测的测试视频test0.mp4，在main.cpp中添加如下代码：

#include <iostream>

#ifdef _WIN32
#define OPENCV
#define GPU
#endif

#include "yolo_v2_class.hpp" //引用动态链接库中的头文件
#include <opencv2/opencv.hpp>
#include "opencv2/highgui/highgui.hpp"

//#pragma comment(lib, "opencv_world340d.lib") //引入OpenCV链接库
#pragma comment(lib, "yolo_cpp_dll.lib") //引入YOLO动态链接库

//以下两段代码来自yolo_console_dll.sln
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 const max_width = (text_size.width > i.w + 2) ? text_size.width : (i.w + 2);
			cv::rectangle(mat_img, cv::Point2f(std::max((int)i.x - 1, 0), std::max((int)i.y - 30, 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 - 10), cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, cv::Scalar(0, 0, 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);
	}
}

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;
}

int main()
{
	std::string names_file = "../../yolo权重/coco.names";
	std::string cfg_file = "../../yolo权重/yolov3.cfg";
	std::string weights_file = "../../yolo权重/yolov3.weights";
	Detector detector(cfg_file, weights_file, 0); //初始化检测器
												  //std::vector<std::string> obj_names = objects_names_from_file(names_file); //调用获得分类对象名称
												  //或者使用以下四行代码也可实现读入分类对象文件
	std::vector<std::string> obj_names;
	std::ifstream ifs(names_file.c_str());
	std::string line;
	while (getline(ifs, line)) obj_names.push_back(line);
	//测试是否成功读入分类对象文件
	for (size_t i = 0; i < obj_names.size(); i++)
	{
		std::cout << obj_names[i] << std::endl;
	}

	cv::VideoCapture capture;
	capture.open("DJI_0002.MP4");
	if (!capture.isOpened())
	{
		printf("文件打开失败");
	}
	cv::Mat frame;
	while (true)
	{
		capture >> frame;
		std::vector<bbox_t> result_vec = detector.detect(frame);
		draw_boxes(frame, result_vec, obj_names);
		cv::namedWindow("test", CV_WINDOW_NORMAL);
		cv::imshow("test", frame);
		cv::waitKey(3);
	}
	return 0;
}

　　

工程配置

包含目录

opencv

cuda

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\include
F:\dongdong\0tool\navidia_cuda_opencv\opencv3.4.1\include
F:\dongdong\0tool\navidia_cuda_opencv\opencv3.4.1\include\opencv2
F:\dongdong\0tool\navidia_cuda_opencv\opencv3.4.1\include\opencv

　　

库目录

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\lib\x64
F:\dongdong\0tool\navidia_cuda_opencv\opencv3.4.1\x64\vc14\lib

　　

输入附加依赖项

增加 cuda

cublas.lib

cuda.lib

cudadevrt.lib

cudart.lib

cudart_static.lib
nvcuvid.lib

OpenCL.lib

cudnn.lib

　　增加yolo

yolo_cpp_dll.lib

　　增加opencv

opencv_aruco341.lib
opencv_bgsegm341.lib
opencv_bioinspired341.lib
opencv_calib3d341.lib
opencv_ccalib341.lib
opencv_core341.lib
opencv_cudaarithm341.lib
opencv_cudabgsegm341.lib
opencv_cudacodec341.lib
opencv_cudafeatures2d341.lib
opencv_cudafilters341.lib
opencv_cudaimgproc341.lib
opencv_cudalegacy341.lib
opencv_cudaobjdetect341.lib
opencv_cudaoptflow341.lib
opencv_cudastereo341.lib
opencv_cudawarping341.lib
opencv_cudev341.lib
opencv_datasets341.lib
opencv_dnn341.lib
opencv_dnn_objdetect341.lib
opencv_dpm341.lib
opencv_face341.lib
opencv_features2d341.lib
opencv_flann341.lib
opencv_fuzzy341.lib
opencv_hfs341.lib
opencv_highgui341.lib
opencv_imgcodecs341.lib
opencv_imgproc341.lib
opencv_img_hash341.lib
opencv_line_descriptor341.lib
opencv_ml341.lib
opencv_objdetect341.lib
opencv_optflow341.lib
opencv_phase_unwrapping341.lib
opencv_photo341.lib
opencv_plot341.lib
opencv_reg341.lib
opencv_rgbd341.lib
opencv_saliency341.lib
opencv_shape341.lib
opencv_stereo341.lib
opencv_stitching341.lib
opencv_structured_light341.lib
opencv_superres341.lib
opencv_surface_matching341.lib
opencv_text341.lib
opencv_tracking341.lib
opencv_video341.lib
opencv_videoio341.lib
opencv_videostab341.lib
opencv_xfeatures2d341.lib
opencv_ximgproc341.lib
opencv_xobjdetect341.lib
opencv_xphoto341.lib

　　预处理器

_CRT_SECURE_NO_WARNINGS

_WINSOCK_DEPRECATED_NO_WARNINGS

　　

工程配置完毕

4 配置代码

代码修改：

1包含yolo文件

#include "yolo_v2_class.hpp" //引用动态链接库中的头文件

　　

由于找不到库文件，把文件拷贝到工程main.cpp函数下

2修改权重文件路径

上一层

再上一层

进入

为了省事也可以直接放在工程里同级目录。

运行代码

贴一张原来教程的作者图

main测试代码

、

#include <iostream>

#ifdef _WIN32
#define OPENCV
#define GPU
#endif

#include "yolo_v2_class.hpp" //引用动态链接库中的头文件
#include <opencv2/opencv.hpp>
#include "opencv2/highgui/highgui.hpp"

//#pragma comment(lib, "opencv_world340d.lib") //引入OpenCV链接库
#pragma comment(lib, "yolo_cpp_dll.lib") //引入YOLO动态链接库

//以下两段代码来自yolo_console_dll.sln
/*
输入：
cv::Mat mat_img,                          目标图像
std::vector<bbox_t> result_vec,           所有目标框信息   位置 大小
std::vector<std::string> obj_names        所有目标名字列表
*/
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) {                     //遍历目标框
		cv::Scalar color = obj_id_to_color(i.obj_id);//根据目标框ID转换颜色
		cv::rectangle(mat_img, cv::Rect(i.x, i.y, i.w, i.h), color, 2);  // 在图像上画目标框
		if (obj_names.size() > i.obj_id) {            //如果目标ID小于名字最大ID，证明事先赋予了名字
			std::string obj_name = obj_names[i.obj_id];  //根据目标ID获取名字，所以训练的时候直接是分配ID了，根据ID在获取名字
			if (i.track_id > 0) obj_name += " - " + std::to_string(i.track_id);// 啥意思？如果有追踪ID?? 加上编号？？
			cv::Size const text_size = getTextSize(obj_name, cv::FONT_HERSHEY_COMPLEX_SMALL, 1.2, 2, 0);// 名字转化为text
			int const max_width = (text_size.width > i.w + 2) ? text_size.width : (i.w + 2);
			//画矩形
			cv::rectangle(mat_img, cv::Point2f(std::max((int)i.x - 1, 0), std::max((int)i.y - 30, 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);
			//画文字
			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_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);
	}
}

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;
}

int main()
{
	std::string names_file = "../../yolo权重/coco.names";
	std::string cfg_file = "../../yolo权重/yolov3.cfg";
	std::string weights_file = "../../yolo权重/yolov3.weights";
	Detector detector(cfg_file, weights_file, 0); //初始化检测器
	//std::vector<std::string> obj_names = objects_names_from_file(names_file); //调用获得分类对象名称
	//或者使用以下四行代码也可实现读入分类对象文件

	//将标签名字从文件逐条读取出来
	std::vector<std::string> obj_names;
	std::ifstream ifs(names_file.c_str());
	std::string line;
	while (getline(ifs, line)) obj_names.push_back(line);//读取成功一条
	//测试是否成功读入分类对象文件
	for (size_t i = 0; i < obj_names.size(); i++)
	{
		std::cout << obj_names[i] << std::endl;          //输出标签名字
	}

	cv::VideoCapture capture;
	capture.open("DJI_0002.MP4");						 //打开测试视频
	if (!capture.isOpened())
	{
		printf("文件打开失败");
	}
	cv::Mat frame;
	while (true)
	{
		capture >> frame;
		std::vector<bbox_t> result_vec = detector.detect(frame);  // 检测一帧，输出目标框信息容器
		draw_boxes(frame, result_vec, obj_names);                 // 目标图像 所有目标检测框 所有目标总分类名称
		cv::namedWindow("test", CV_WINDOW_NORMAL);
		cv::imshow("test", frame);
		cv::waitKey(3);
	}
	return 0;
}