adabooost算法实时性

[jiangwenbo737373]

研究生毕业设计，检测人眼，开始用粒子群优化，后来效果不好，转成adaboost分类器，在网上看到的资料都是检测速度快，可以达到及时毫秒，但是在我实现看来，速度很慢，举个例子，简单识别人脸，用opencv自带的分类器.xml文件，摄像头获取图片为320*240大小，检测效果相当不好，卡顿现象很严重，求教这是为什么？

代码如下：
void adaboost_face(IplImage *dst)
{
CvPoint point1, point2;
CvRect* rect;
//CvScalar color = {0, 255, 0};
CvMemStorage* storage = cvCreateMemStorage(0);
CvHaarClassifierCascade* cascade = (CvHaarClassifierCascade*)cvLoad( "haarcascade_eye.xml", 0, 0, 0 );
cvClearMemStorage( storage );
if(cascade) {
CvSeq* faces = cvHaarDetectObjects( dst, cascade, storage, 1.2, 2, 1 , cvSize(30, 30) );
for(int i = 0; i < (faces ? faces->total : 0); i++ )
{
rect = (CvRect*)cvGetSeqElem( faces, i );
point1 = cvPoint(rect->x, rect->y);
point2 = cvPoint(rect->x + rect->width, rect->y + rect->height);
cvRectangle(dst, point1, point2, cvScalar(0, 255, 0), 2, 8, 0);
}
}
cvReleaseMemStorage(&storage);
}




int main()
{
int nFrmNum = 0;
IplImage* pFrame = 0,*pGray=0,*pResult = 0,*pRoi = 0;
CvCapture *capture = 0;

CvPoint pointx={0,0},pointy={0,0},point3={0,0},point4={0,0};
CvRect* rect_eye = 0,*rect_face = 0;
cvNamedWindow("video", 1);
capture = cvCaptureFromCAM(-1);
for(;;)
{
pFrame = cvQueryFrame(capture);
if (pFrame == 0)
{
continue;
}
nFrmNum++;
if (nFrmNum == 1)
{
pResult = cvCreateImage(cvGetSize(pFrame),pFrame->depth,pFrame->nChannels);
pGray = cvCreateImage(cvGetSize(pFrame),pFrame->depth,pFrame->nChannels);
}
pResult->width = pFrame->width/2;
pResult->height = pFrame->height/2;
cvResize(pFrame, pResult, CV_INTER_LINEAR);

adaboost_face(pResult);
//cvSetImageROI(pResult,cvRect(pointx.x,pointx.y,pointy.x,pointy.y));
//pRoi = cvCreateImage(cvGetSize(pResult),pResult->depth,pResult->nChannels);
//cvCopy(pResult,pRoi);
//cvResetImageROI(pResult);
//adaboost_eyes(pRoi,point3,point4);
cvShowImage("video",pResult);
cvWaitKey(50);
}
cvReleaseImage(&pResult);
//cvReleaseImage(&pRoi);
cvDestroyWindow("video");
return 0;
}

[青松2]

你不能每调用一次adaboost_face，就加载、开辟资源，这样太慢了，一次性加载吧

[Hope_f]

人才，最浪费时间的是： CvHaarClassifierCascade* cascade = (CvHaarClassifierCascade*)cvLoad( "haarcascade_eye.xml", 0, 0, 0 ); opencv解析xml文件的时候很费时的，你需要把cascade一次加载，然后每次直接使用就好了，haar已经可以做到实时检测人脸了，在1280*720的情况下都能实时，更何况你那点320*240的.