[C#]使用C#部署yolov8的obb旋转框检测tensorrt模型

【测试通过环境】

win10 x
vs2019
cuda11.7+cudnn8.8.0
TensorRT-8.6.1.6
opencvsharp==4.9.0
.NET Framework4.7.2

NVIDIA GeForce RTX 2070 Super

Windows版 CUDA安装参考：

【特别注意】

tensorrt依赖不同硬件需要自己从onnx转换tensorrt，转换就是调用api实现，比如

TensorRtSharp.Custom.Nvinfer.OnnxToEngine(@"yolov8s-obb.onnx",1024);

【视频演示和解说】

【部分实现源码】

using System;
using System.Collections.Generic;
using System.ComponentModel;
using System.Data;
using System.Drawing;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
using System.Windows.Forms;
using FIRC;
using OpenCvSharp;
using TrtCommon;
using TensorRtSharp;
using TensorRtSharp.Custom;
using System.Diagnostics;

namespace WindowsFormsApp1
{
    public partial class Form1 : Form
    {
        public Form1()
        {
            InitializeComponent();
        }

        private void button1_Click(object sender, EventArgs e)
        {
            CocoOption.lables = Dotav1Option.lables;
            Yolov8Obb detector = new Yolov8Obb("yolov8s-obb.engine");
            Mat image1 = Cv2.ImRead(@"P0032.png");
            var Results = detector.Predict(new List<Mat> { image1 });
            Mat re_image1 = Visualize.DrawObbResult(Results[0], image1);
            Cv2.NamedWindow("result",WindowFlags.KeepRatio);
            Cv2.ImShow("result", re_image1);
            Cv2.WaitKey(0);
        }

        private void button2_Click(object sender, EventArgs e)
        {
            TensorRtSharp.Custom.Nvinfer.OnnxToEngine(@"yolov8s-obb.onnx",1024);
        }

        private void button3_Click(object sender, EventArgs e)
        {
            CocoOption.lables = Dotav1Option.lables;
            Yolov8Obb detector = new Yolov8Obb("yolov8s-obb.engine");
            VideoCapture capture = new VideoCapture(0);
            if (!capture.IsOpened())
            {
                Console.WriteLine("video not open!");
                return;
            }
            Mat frame = new Mat();
            var sw = new Stopwatch();
            int fps = 0;
            while (true)
            {

                capture.Read(frame);
                if (frame.Empty())
                {
                    Console.WriteLine("data is empty!");
                    break;
                }
                sw.Start();
                var results = detector.Predict(new List<Mat> { frame });
                Mat resultImg = Visualize.DrawObbResult(results[0], frame);
                sw.Stop();
                fps = Convert.ToInt32(1 / sw.Elapsed.TotalSeconds);
                sw.Reset();
                Cv2.PutText(resultImg, "FPS=" + fps, new OpenCvSharp.Point(30, 30), HersheyFonts.HersheyComplex, 1.0, new Scalar(255, 0, 0), 3);
                //显示结果
                Cv2.ImShow("Result", resultImg);
                int key = Cv2.WaitKey(10);
                if (key == 27)
                    break;
            }

            capture.Release();
        }
    }
}

注意源码提供上面对应环境的dll，只需要安装上面一样cuda+cudnn和tensorrt版本即可正常运行。如果您不安装一样版本不能正常运行。此时需要重新编译TensorRtExtern.dll，此外由于tensorrt依赖硬件不一样电脑可能无法共用tensorrt模型，所以必须要重新转换onnx模型到engine才可以运行。