Visual C++ darknet opencv cuda yolo4測試即時影像、小孩玩車、柴犬影片識別

 

 Visual C++ darknet opencv cuda yolo4測試即時影像、柴犬影片識別。「Vcpkg安裝darknet cuda並以C++ yolo v3 tiny辨識柴犬白狗動作影片」https://www.youtube.com/watch?v=ZODKCQABcKM

#include <iostream>
#include <fstream>
#include <opencv2/opencv.hpp>
#include <vector>
using namespace cv;
using namespace std;
using namespace cv::dnn;
vector<string> coco;
void name_list() {
	fstream fin("coco.names", ios::in);
	string tmp = "";
	while (getline(fin, tmp))
		coco.push_back(tmp);
	fin.close();
	cout << coco.size() << endl;
}
int main()
{
	cout << "參考https://pjreddie.com/darknet/yolo/ \n";
	name_list();
	cv::setUseOptimized(1);
	VideoCapture cap("shiba_turns_around.mp4");
	cap.set(CAP_PROP_FRAME_WIDTH, 1280);
	cap.set(CAP_PROP_FRAME_HEIGHT, 720);
	string model = "yolov4.weights";
	string config = "yolov4.cfg";
	Net network = readNet(model, config, "Darknet");
	network.setPreferableBackend(DNN_BACKEND_CUDA);
	network.setPreferableTarget(DNN_TARGET_CUDA);
	string wname = "Darknet yolo4 ";
	namedWindow(wname);
	Vec3b colors[] = { Vec3b(255,255,0), Vec3b(0, 255,0), Vec3b(0,255,255), Vec3b(255,0,255) };
	int i = 0;
	while (1)
	{
		Mat im;
		cap >> im;
		if (!im.data) break;
        
		Mat blobFromIm;
		bool swapRB = true;
        	blobFromImage(im, blobFromIm, 1, Size(416, 416), Scalar(), swapRB, false);
        	//Size(608, 608) or (416, 416)
      
		float scale = 1.0 / 255.0;
		Scalar mean = 0;
		network.setInput(blobFromIm, "", scale, mean);
 
		Mat detection = network.forward();
		// rows represent number of detected objects 
		int rowsNoOfDetection = detection.rows;
 
		// [x ; y ; w; h; class 0 ; class 1 ; class 2 ;....]
		//  
		int xywh_scores = detection.cols;
 
		// Loop over number of detected object. 
		for (int j = 0; j < rowsNoOfDetection; ++j)
		{
			// for each row, the score is from element 5 up to end
			Mat scores = detection.row(j).colRange(5, xywh_scores);
 
			Point PositionOfMax;
			double confidence;
 
			//This function find indexes of min and max confidence and related index.
			minMaxLoc(scores, 0, &confidence, 0, &PositionOfMax);
 
			if (confidence > 0.01)
			{
				// [x ; y ; w; h;
				int x = (int)(detection.at<float>(j, 0) * im.cols);
				int y = (int)(detection.at<float>(j, 1) * im.rows);
				int w = (int)(detection.at<float>(j, 2) * im.cols);
				int h = (int)(detection.at<float>(j, 3) * im.rows);
 
				int left = x - w / 2;
				int up = y - h / 2;
 
				stringstream ss;
				int i = PositionOfMax.x;
				ss << coco[i] << ",prob=" << 100 * confidence << "%";
				string what = ss.str();
				cout << '#' << j << ":[" << x << ',' << y << ',' << w << ',' << h << "]:";
				cout<< what << endl;
				Vec3b color = confidence * colors[i % 4];
				rectangle(im, Rect(left, up, w, h), color, 2, LINE_8, 0);
 
				int baseLine = 0;
				Size labelSize = getTextSize(what, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
				rectangle(im, Rect(Point(left, up - labelSize.height),
					Size(labelSize.width, labelSize.height + baseLine)), color, -1);
				putText(im, what, Point(left, up), FONT_HERSHEY_SIMPLEX, 0.5, Scalar());
			}
		}
		imshow(wname, im);
		if (waitKey(1) == 27) {
			imwrite("doggy.png", im);
			break;
		}
	}
	destroyAllWindows();
	cap.release();
	return 0;
}