加权迭代重建背景，测试常规循环，SSE和OpenCV的速度

mk_background.cpp

#include <opencv2/opencv.hpp>
//#include <stdlib.h>
#include <mmintrin.h>
//#include <xmmintrin.h>
#include <time.h>
using namespace cv;


#define CALC_RUN_TIME(fun_line, func_name) \
{\
	clock_t start,stop;\
	float temp;\
	start = clock();\
	fun_line;\
	stop = clock();\
	temp=((double)(stop-start)*1000.0)/CLOCKS_PER_SEC;\
	char format_str[64];\
	strcpy(format_str, (func_name));\
	strcat(format_str, ": %.4fms\n");\
	printf(format_str, temp);\
};

void mk_bg_sse(Mat &gray, Mat &bg, Mat &gray32, Mat &bg32, float alpha){ //release:3.5ms
	gray.convertTo(gray32, CV_32FC1);
	bg.convertTo(bg32, CV_32FC1);

	int Ni = gray.rows * gray.cols / 4;
	float *pGray = (float*)gray32.data;
	float *pBg = (float*)bg32.data;
	__m128 _m_gray, _m_bg;
	__m128 *_p_m_gray = (__m128*)gray32.data;
	__m128 *_p_m_bg = (__m128*)bg32.data;
	__m128 _m_alpha_1, _m_alpha_2;
	_m_alpha_1 = _mm_set_ps1(alpha);
	_m_alpha_2 = _mm_set_ps1(1 - alpha);

	for (int i = 0; i < Ni; i++)
	{
		/*_m_gray = _mm_load_ps(pGray);
		_m_bg = _mm_load_ps(pBg);
		_m_bg = _mm_add_ps(_mm_mul_ps(_m_gray, _m_alpha_1), _mm_mul_ps(_m_bg, _m_alpha_2));
		_mm_store_ps(pBg, _m_bg);

		pGray += 4;
		pBg += 4;*/
		*_p_m_bg = _mm_add_ps(_mm_mul_ps(*_p_m_gray, _m_alpha_1), _mm_mul_ps(*_p_m_bg, _m_alpha_2));

		_p_m_bg++;
		_p_m_gray++;
	}

	bg32.convertTo(bg, CV_8UC1);
}


void mk_bg_cv(Mat &gray, Mat &bg, Mat &gray32, Mat &bg32, float alpha){ //release:2.5ms
	//gray.convertTo(gray32, CV_32FC1);
	bg.convertTo(bg32, CV_32FC1);
	accumulateWeighted(gray, bg32, alpha);
	bg32.convertTo(bg, CV_8UC1);
}

void mk_bg_normal(Mat &gray, Mat &bg, Mat &gray32, Mat &bg32, float alpha){ //release:5ms
	int Ni = gray.rows * gray.cols;
	uchar *pGray = gray.data;
	uchar *pBg = bg.data;
	float alpha1 = 1 - alpha;
	for (int i = 0; i < Ni; i++)
	{
		*pBg = *pBg * alpha1 + *pGray * alpha;

		pGray++;
		pBg++;
	}

}

int main(int argc, char **argv){
	VideoCapture capture("test.mov");
	Mat colorMat;
	capture >> colorMat;

	Mat grayMat;
	grayMat.create(colorMat.rows, colorMat.cols, CV_8UC1); //768*576
	Mat backgroundMat;
	backgroundMat.create(colorMat.rows, colorMat.cols, CV_8UC1);
	backgroundMat.setTo(Scalar(0));

	Mat grayMat32;
	grayMat32.create(colorMat.rows, colorMat.cols, CV_32FC1);
	Mat backgroundMat32;
	backgroundMat32.create(colorMat.rows, colorMat.cols, CV_32FC1);
	
	char keyborad = 0;
	namedWindow("Gray");
	namedWindow("BG");
	while (keyborad != 27)
	{
		capture >> colorMat;
		if (colorMat.empty())
		{
			break;
		}
		cvtColor(colorMat, grayMat, CV_BGR2GRAY);

		CALC_RUN_TIME(mk_bg_normal(grayMat, backgroundMat, grayMat32, backgroundMat32, 0.1), "mk_bg_normal");
		//mk_bg_sse(grayMat, backgroundMat, grayMat32, backgroundMat32, 0.1);
		
		imshow("Gray", grayMat);
		imshow("BG", backgroundMat);
		keyborad = waitKey(3);
	}
	return 0;
}