knn背景建模优点与不足

kNN背景建模c++代码（包括比较）如下：

#include "opencv2/imgcodecs.hpp"
#include "opencv2/imgproc.hpp"
#include "opencv2/videoio.hpp"
#include <opencv2/highgui.hpp>
#include <opencv2/video.hpp>
#include <iostream>
#include <sstream>
using namespace cv;
using namespace std;
#define RATIO 2
const int HISTORY_NUM = 14;// 14;// 历史信息帧数
const int nKNN = 3;// KNN聚类后判断为背景的阈值
const float defaultDist2Threshold = 20.0f;// 灰度聚类阈值

struct PixelHistory
{
	unsigned char *gray;// 历史灰度值
	unsigned char *IsBG;// 对应灰度值的前景/背景判断，1代表判断为背景，0代表判断为前景
};


int main()
{
	PixelHistory* framePixelHistory = NULL;// 记录一帧图像中每个像素点的历史信息
	cv::Mat frame, FGMask, FGMask_KNN;
	int keyboard = 0;
	int rows, cols;
	rows = cols = 0;
	bool InitFlag = false;
	int frameCnt = 0;
	int gray = 0;
	char* file_path = "..//Data//in_man.avi";
	// Foreground mask generated by MOG2 method
	Mat fgMaskMOG2;
	// Background
	Mat bgImg;
	VideoCapture capture(file_path);
	Ptr<BackgroundSubtractorKNN> pBackgroundKnn =
		createBackgroundSubtractorKNN();
	pBackgroundKnn->setHistory(200);
	pBackgroundKnn->setDist2Threshold(600);
	pBackgroundKnn->setShadowThreshold(0.5);
	Ptr<BackgroundSubtractorMOG2> pMOG2 = createBackgroundSubtractorMOG2(200, 36.0, false);

	while ((char)keyboard != 'q' && (char)keyboard != 27)
	{
		// 读取当前帧
		if (!capture.read(frame))
			exit(EXIT_FAILURE);

		resize(frame, frame, Size(frame.size().width / RATIO, frame.size().height / RATIO));
		imshow("Frame", frame);
		pMOG2->apply(frame, fgMaskMOG2);
		pMOG2->getBackgroundImage(bgImg);
		medianBlur(fgMaskMOG2, fgMaskMOG2, 5);
		//			imshow("medianBlur", fgMaskMOG2);
					// Fill black holes
		morphologyEx(fgMaskMOG2, fgMaskMOG2, MORPH_CLOSE, getStructuringElement(MORPH_RECT, Size(5, 5)));
		// Fill white holes
		morphologyEx(fgMaskMOG2, fgMaskMOG2, MORPH_OPEN, getStructuringElement(MORPH_RECT, Size(5, 5)));
		imshow("morphologyEx", fgMaskMOG2);

		cvtColor(frame, frame, CV_BGR2GRAY);
		if (!InitFlag)
		{
			// 初始化一些变量
			rows = frame.rows;
			cols = frame.cols;
			FGMask.create(rows, cols, CV_8UC1);// 输出图像初始化

			// framePixelHistory分配空间
			framePixelHistory = (PixelHistory*)malloc(rows*cols * sizeof(PixelHistory));
			for (int i = 0; i < rows*cols; i++)
			{
				framePixelHistory[i].gray = (unsigned char*)malloc(HISTORY_NUM * sizeof(unsigned char));
				framePixelHistory[i].IsBG = (unsigned char*)malloc(HISTORY_NUM * sizeof(unsigned char));
				memset(framePixelHistory[i].gray, 0, HISTORY_NUM * sizeof(unsigned char));
				memset(framePixelHistory[i].IsBG, 0, HISTORY_NUM * sizeof(unsigned char));
			}

			InitFlag = true;
		}
		if (InitFlag)
		{
			FGMask.setTo(Scalar(255));
			for (int i = 0; i < rows; i++)
			{
				for (int j = 0; j < cols; j++)
				{
					gray = frame.at<unsigned char>(i, j);
					int fit = 0;
					int fit_bg = 0;
					// 比较确定前景/背景
					for (int n = 0; n < HISTORY_NUM; n++)
					{
						if (fabs(gray - framePixelHistory[i*cols + j].gray[n]) < defaultDist2Threshold)// 灰度差别是否位于设定阈值内
						{
							fit++;
							if (framePixelHistory[i*cols + j].IsBG[n])// 历史信息对应点之前被判断为背景
							{
								fit_bg++;
							}
						}
					}
					if (fit_bg >= nKNN)// 当前点判断为背景
					{
						FGMask.at<unsigned char>(i, j) = 0;
					}

					// 更新历史值
					int index = frameCnt % HISTORY_NUM;
					framePixelHistory[i*cols + j].gray[index] = gray;
					framePixelHistory[i*cols + j].IsBG[index] = fit >= nKNN ? 1 : 0;// 当前点作为背景点存入历史信息

				}
			}
		}
			pBackgroundKnn->apply(frame, FGMask_KNN);
			imshow("yuanFGMask", FGMask);
			medianBlur(FGMask, FGMask, 5);
			imshow("medianBlur", FGMask);
			cv::Mat structuringElement2x2 = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(2, 2));
			erode(FGMask, FGMask, structuringElement2x2);  //两次侵蚀处理，以消除噪音
			cv::threshold(FGMask, FGMask, 20, 255.0, CV_THRESH_BINARY);	//执行阈值处理并获得阈值掩码
			cv::dilate(FGMask, FGMask, structuringElement2x2);
			imshow("FGMask", FGMask);
			//构造各种尺寸的元素以用于形态学变换
			//cv::Mat structuringElement2x2 = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(2, 2));
			erode(FGMask_KNN, FGMask_KNN, structuringElement2x2);  //两次侵蚀处理，以消除噪音
			cv::threshold(FGMask_KNN, FGMask_KNN, 20, 255.0, CV_THRESH_BINARY);	//执行阈值处理并获得阈值掩码
			cv::dilate(FGMask_KNN, FGMask_KNN, structuringElement2x2);
			imshow("FGMask_KNN", FGMask_KNN);

		keyboard = waitKey(30);
		frameCnt++;
	}
	capture.release();

	return 0;

}

相比较而言：knn能够很快的检测出进入窗口的物体，检测到较小的物体
但有很严重的“尾巴”（特别是当k值较大（如5）的时候）

参考文章：OpenCV3学习（10.4）基于KNN的背景/前景分割算法BackgroundSubtractorKNN算法
image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaWppYW5jaGVuZzk5OQ==,size_16,color_FFFFFF,t_70