使用ONNX Runtime在C++中部署YOLOv5模型

 引言

        介绍如何在C++环境中使用ONNX Runtime部署YOLOv5模型实现目标检测。由于笔记本电脑配置问题，没有实现cuda加速。下图是图片检测结果，代码内容是视频检测。

一、技术栈概述

1. YOLOv5模型
        YOLOv5是Ultralytics公司开发的单阶段目标检测器，相比前代有更优的速度-精度平衡。其核心特点包括：

• 使用CSPDarknet53作为骨干网络
• PANet特征金字塔结构
• 自适应锚框计算
• 输出格式为1x25200x85的张量

2. ONNX Runtime
ONNX Runtime是一个高性能推理引擎，支持多种硬件平台，提供统一的API接口。主要优势：

• 跨平台支持(Windows/Linux/macOS)
• 多硬件后端(CPU/GPU/TPU等)
• 图优化和量化支持
• 低延迟推理

二、部署代码详解

#include "onnxruntime_cxx_api.h"
#include "cpu_provider_factory.h" //引入了ONNX Runtime的C++ API和CPU提供者工厂
#include<opencv2/opencv.hpp>
#include<fstream> //用于读取标签文件
#include <iostream>

using namespace std;
using namespace cv;
using namespace Ort;

/*
yolov5
输入: 1x3x640x640的归一化图像
输出: 1x25200x85的检测结果矩阵
*/

//读取类别标签
string labels_txt_file = "E:/PycharmProjects/YOLO/yolov5-7.0/yolov5-7.0/classes.txt";
vector<string> readClassNames()
{
	vector<string> classNames;

	//打开 labels_txt_file 指定的文件，并创建一个文件输入流对象 fp，后续可以通过 fp 操作该文件
	ifstream fp(labels_txt_file);
	if (!fp.is_open())
	{
		printf("文件打开失败!\n");
		exit(-1);
	}
	string name;
	while (!fp.eof())
	{
		getline(fp, name);
		if (name.length())
			classNames.push_back(name);
	}
	fp.close();
	return classNames;

}

int main()
{
	//初始化
	float x_factor = 0.0;
	float y_factor = 0.0;
	vector<string> labels = readClassNames();
	//Mat frame = imread("E:/PycharmProjects/YOLO/yolov5-7.0/yolov5-7.0/data/images/bus.jpg");//检测图片
	string videopath = "E:/face_test.mp4";//检测视频
	//VideoCapture cap(0);//检测本机摄像头
	VideoCapture cap(videopath);
	if (!cap.isOpened())
	{
		cerr << "无法打开摄像头" << endl;
		return -1;
	}

	//创建onnx runtime会话
	string onnxpath = "E:/PycharmProjects/YOLO/yolov5-7.0/yolov5-7.0/yolov5s.onnx";
	wstring modelpath = wstring(onnxpath.begin(), onnxpath.end());//逐个字符转换为宽字符字符串
	SessionOptions session_options;
	Env env = Env(ORT_LOGGING_LEVEL_ERROR, "env_yolov5s_onnx");

	session_options.SetGraphOptimizationLevel(ORT_ENABLE_BASIC);//优化计算图，提高推理速度
	OrtSessionOptionsAppendExecutionProvider_CPU(session_options, 0);
	Session session_(env, modelpath.c_str(), session_options);//创建会话

	
	//获取输入输出信息
	int input_nodes_num = session_.GetInputCount();
	int output_nodes_num = session_.GetOutputCount();
	vector<string> input_node_names;
	vector<string> output_node_names;
	AllocatorWithDefaultOptions allocator;
	int input_h = 0;
	int input_w = 0;

	for (int i = 0; i < input_nodes_num; i++)
	{
		auto input_name = session_.GetInputNameAllocated(i, allocator);
		input_node_names.push_back(input_name.get());
		auto inputShapeInfo = session_.GetInputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
		int ch = inputShapeInfo[1];
		input_h = inputShapeInfo[2];
		input_w = inputShapeInfo[3];
		cout << "输入格式：" << ch << "x" << input_h << "x" << input_w << endl;

	}

	int num = 0;
	int nc = 0;
	for (int i = 0; i < output_nodes_num; i++)
	{
		auto output_name = session_.GetOutputNameAllocated(i, allocator);
		output_node_names.push_back(output_name.get());
		auto outShapeInfo = session_.GetOutputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
		num = outShapeInfo[1];
		nc = outShapeInfo[2];
		cout << "输出格式：" << num << "x" << nc << endl;

	}

	while(true)
	{ 
		Mat frame;
		cap >> frame;
		if (frame.empty())
		{
			cerr << "无法获取帧！" << endl;
			break;
		}

		//图像预处理
		int64 start = getTickCount();//记录当前时间戳，用于后续计算 FPS（帧率）
		//图像填充，将图像放在一个正方形背景中，保持宽高比不变，避免变形
		int w = frame.cols;
		int h = frame.rows;
		int max_ = max(h, w);
		Mat image = Mat::zeros(Size(max_, max_), CV_8UC3);
		Rect roi(0, 0, w, h);
		frame.copyTo(image(roi));
		x_factor = image.cols / static_cast<float>(640);//记录图像从实际尺寸到模型输入尺寸640x640的缩放比例
		y_factor = image.rows / static_cast<float>(640);
		//归一化
		Mat blob = dnn::blobFromImage(image, 1.0 / 255.0, Size(input_w, input_h), Scalar(0, 0, 0), true, false);
		size_t tpixels = input_h * input_w * 3; // 计算总像素数（640*640*3），表示输入张量的元素总数
		array<int64_t, 4> input_shape_info{ 1,3,input_h,input_w };//显式指定输入形状为 [1, 3, 640, 640]

		//创建输入张量并推理
		auto allocator_info = MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU);
		Value input_tensor_ = Value::CreateTensor<float>(allocator_info, blob.ptr<float>(), tpixels, input_shape_info.data(), input_shape_info.size());
		const array<const char*, 1>inputNames = { input_node_names[0].c_str() };
		const std::array<const char*, 1> outNames = { output_node_names[0].c_str() };

		vector<Value> ort_outputs;
		try
		{
			ort_outputs = session_.Run(RunOptions{ nullptr }, inputNames.data(), &input_tensor_, 1, outNames.data(), outNames.size());
		}
		catch (exception e)
		{
			cout << e.what() << endl;
		}

		const float* pdata = ort_outputs[0].GetTensorMutableData<float>();

		//后处理
		vector<Rect> boxes;
		vector<int> classIds;
		vector<float> confidences;
		Mat det_output(num, nc, CV_32F, (float*)pdata);

		for (int i = 0; i < det_output.rows; i++)
		{
			float confidence = det_output.at<float>(i, 4);
			if (confidence < 0.45) {
				continue;
			}
			Mat classes_scores = det_output.row(i).colRange(5, nc);
			Point classIdPoint;
			double score;
			minMaxLoc(classes_scores, 0, &score, 0, &classIdPoint);

			// 置信度 0～1之间
			if (score > 0.25)
			{
				float cx = det_output.at<float>(i, 0);
				float cy = det_output.at<float>(i, 1);
				float ow = det_output.at<float>(i, 2);
				float oh = det_output.at<float>(i, 3);
				int x = static_cast<int>((cx - 0.5 * ow) * x_factor);
				int y = static_cast<int>((cy - 0.5 * oh) * y_factor);
				int width = static_cast<int>(ow * x_factor);
				int height = static_cast<int>(oh * y_factor);
				cv::Rect box;
				box.x = x;
				box.y = y;
				box.width = width;
				box.height = height;

				boxes.push_back(box);
				classIds.push_back(classIdPoint.x);
				confidences.push_back(score);
			}
		}

		// 非极大值抑制
		vector<int> indexes;
		dnn::NMSBoxes(boxes, confidences, 0.25, 0.45, indexes);
		for (size_t i = 0; i < indexes.size(); i++)
		{
			int idx = indexes[i];
			int cid = classIds[idx];
			rectangle(frame, boxes[idx], cv::Scalar(0, 0, 255), 2, 8, 0);
			putText(frame, labels[cid].c_str(), boxes[idx].tl(), FONT_HERSHEY_PLAIN, 1.0, Scalar(255, 0, 0), 1, 8);
		}
		// 计算FPS
		float t = (cv::getTickCount() - start) / static_cast<float>(cv::getTickFrequency());
		putText(frame, cv::format("FPS: %.2f", 1.0 / t), Point(20, 40), FONT_HERSHEY_PLAIN, 2.0, Scalar(255, 0, 0), 1, 8);

		imshow("YOLOv5实时推理演示", frame);
		//按ESC退出
		if (waitKey(1) == 27)
		{
			break;
		}
	}
	// relase resource
	session_options.release();
	session_.release();
	return 0;
}