OpenCV实现卡尔曼滤波

        本文将简要回顾一下卡尔曼滤波理论,然后详细介绍如何在OpenCV中使用卡尔曼滤波进行跟踪,最后给两个程序实例。

1. 卡尔曼滤波理论回顾

      对于一个动态系统,我们首先定义一组状态空间方程

     状态方程:     

     测量方程:      

        xk是状态向量,zk是测量向量,Ak是状态转移矩阵,uk是控制向量,Bk是控制矩阵,wk是系统误差(噪声),Hk是测量矩阵,vk是测量误差(噪声)。wk和vk都是高斯噪声,即

                             

    整个卡尔曼滤波的过程就是个递推计算的过程,不断的“预测——更新——预测——更新……”

预测

     预测状态值:              

     预测最小均方误差:   

更新

    测量误差:                   

    测量协方差:                

    最优卡尔曼增益:         

    修正状态值:                

    修正最小均方误差:     


2.OpenCV中的KalmanFilter详解

OpenCV中有两个版本的卡尔曼滤波方法KalmanFilter(C++)和CvKalman(C),用法差不太多,这里只介绍KalmanFilter。

C++版本中将KalmanFilter封装到一个类中,其结构如下所示:


 
 
  
  
  1. class CV_EXPORTS_W KalmanFilter
  2. {
  3. public:
  4. CV_WRAP KalmanFilter(); //构造默认KalmanFilter对象
  5. CV_WRAP KalmanFilter(int dynamParams, int measureParams, int controlParams=0, int type=CV_32F);   //完整构造KalmanFilter对象方法
  6. void init(int dynamParams, int measureParams, int controlParams=0, int type=CV_32F); //初始化KalmanFilter对象,会替换原来的KF对象
  7. CV_WRAP const Mat& predict(const Mat& control=Mat()); //计算预测的状态值
  8. CV_WRAP const Mat& correct(const Mat& measurement); //根据测量值更新状态值
  9. Mat statePre; //预测值 (x'(k)): x(k)=A*x(k-1)+B*u(k)
  10. Mat statePost; //状态值 (x(k)): x(k)=x'(k)+K(k)*(z(k)-H*x'(k))
  11. Mat transitionMatrix; //状态转移矩阵 (A)
  12. Mat controlMatrix; //控制矩阵 B
  13. Mat measurementMatrix; //测量矩阵 H
  14. Mat processNoiseCov; //系统误差 Q
  15. Mat measurementNoiseCov; //测量误差 R
  16. Mat errorCovPre; //最小均方误差 (P'(k)): P'(k)=A*P(k-1)*At + Q)
  17. Mat gain; //卡尔曼增益 (K(k)): K(k)=P'(k)*Ht*inv(H*P'(k)*Ht+R)
  18. Mat errorCovPost; //修正的最小均方误差 (P(k)): P(k)=(I-K(k)*H)*P'(k)
  19. // 临时矩阵
  20. Mat temp1;
  21. Mat temp2;
  22. Mat temp3;
  23. Mat temp4;
  24. Mat temp5;
  25. };
  26. enum
  27. {
  28. OPTFLOW_USE_INITIAL_FLOW = CV_LKFLOW_INITIAL_GUESSES,
  29. OPTFLOW_LK_GET_MIN_EIGENVALS = CV_LKFLOW_GET_MIN_EIGENVALS,
  30. OPTFLOW_FARNEBACK_GAUSSIAN = 256
  31. };


       函数原型见:…..\OpenCV2\sources\modules\ocl\src\kalman.cpp

       只有四个方法: 构造KF对象KalmanFilter(DP,MP,CP)、初始化KF对象init(DP,MP,CP)、预测predict( )、更新correct( )。除非你要重新构造KF对象,否则用不到init( )。

KalmanFilter(DP,MP,CP)和init( )就是赋值,没什么好说的。

      注意:KalmanFilter结构体中并没有测量值,测量值需要自己定义,而且一定要定义,因为后面要用。


编程步骤

step1:定义KalmanFilter类并初始化

    //构造KF对象

    KalmanFilter KF(DP, MP, 0);

    //初始化相关参数

    KF.transitionMatrix                         转移矩阵 A

    KF.measurementMatrix                  测量矩阵    H

    KF.processNoiseCov                     过程噪声 Q

    KF.measurementNoiseCov            测量噪声        R

    KF.errorCovPost                            最小均方误差 P

    KF.statePost                                系统初始状态 x(0) 

    Mat measurement                          定义初始测量值 z(0) 

step2:预测

    KF.predict( )                                                 //返回的是下一时刻的状态值KF.statePost (k+1) 

step3:更新

    更新measurement;                                     //注意measurement不能通过观测方程进行计算得到,要自己定义!

    更新KF   KF.correct(measurement)

最终的结果应该是更新后的statePost.


相关参数的确定

    对于系统状态方程,简记为Y=AX+B,X和Y是表示系统状态的列向量,A是转移矩阵,B是其他项。

    状态值(向量)只要能表示系统的状态即可,状态值的维数决定了转移矩阵A的维数,比如X和Y是N×1的,则A是N×N的。

    A的确定跟X有关,只要保证方程中不相干项的系数为0即可,看下面例子

      X和Y是二维的,

       X和Y是三维的,

          X和Y是三维的,但c和△ c是相关项

      上面的1也可以是其他值。



下面对predict( ) 和correct( )函数介绍下,可以不用看,不影响编程。


 
 
  
  
  1. CV_EXPORTS const oclMat& KalmanFilter::predict(const oclMat& control)
  2. {
  3. gemm(transitionMatrix, statePost, 1, oclMat(), 0, statePre);
  4. oclMat temp;
  5. if(control.data)
  6. gemm(controlMatrix, control, 1, statePre, 1, statePre);
  7. gemm(transitionMatrix, errorCovPost, 1, oclMat(), 0, temp1);
  8. gemm(temp1, transitionMatrix, 1, processNoiseCov, 1, errorCovPre, GEMM_2_T);
  9. statePre.copyTo(statePost);
  10. return statePre;
  11. }

gemm( )是矩阵的广义乘法

void gemm(const GpuMat& src1, constGpuMat& src2, double alpha, const GpuMat& src3, double beta,GpuMat& dst, int flags=0, Stream& stream=Stream::Null())

    dst = alpha · src1 · src2 +beta· src3

   上面,oclMat()其实是uk,只不过默认为0,所以没赋值。整个过程就计算了x'和P’。(用x'代表x的预测值,用P'代表P的预测值)。GEMM_2_T表示对第2个参数转置。

x’(k)=1·A·x(k-1)

如果B非空, x'(k) = 1·B·u + 1·x'(k-1)

temp1 = 1·A·P(k-1) + 0·u(k)

P’(k) = 1· temp1·AT + 1· Qk= A·P(k-1)·AT + 1· Qk

       可见,和第一部分的理论介绍完全一致。


 
 
  
  
  1. CV_EXPORTS const oclMat& KalmanFilter::correct(const oclMat& measurement)
  2. {
  3. CV_Assert(measurement.empty() == false);
  4. gemm(measurementMatrix, errorCovPre, 1, oclMat(), 0, temp2);
  5. gemm(temp2, measurementMatrix, 1, measurementNoiseCov, 1, temp3, GEMM_2_T);
  6. Mat temp;
  7. solve(Mat(temp3), Mat(temp2), temp, DECOMP_SVD);
  8. temp4.upload(temp);
  9. gain = temp4.t();
  10. gemm(measurementMatrix, statePre, -1, measurement, 1, temp5);
  11. gemm(gain, temp5, 1, statePre, 1, statePost);
  12. gemm(gain, temp2, -1, errorCovPre, 1, errorCovPost);
  13. return statePost;
  14. }
bool solve(InputArray src1, InputArray src2, OutputArray dst, int flags=DECOMP_LU)

求解线型最小二乘估计



temp2 = 1· H·P’ + 0·u(k)

temp3 = 1· temp2·HT + 1·R = H·P’·HT+ 1· R   也就是上面的Sk

temp = argmin||tem2- temp3||

K=temp

temp5 = -1· H·x’ + 1·zk        就是上面的y’。

x = 1·K·temp5 + 1·x’ = KT·y’ +x’

P =-1·K·temp2 + 1·P’ = -K·H·P’+P’ = (I- K·H) P’

也和第一部分的理论完全一致。


通过深入函数内部,学到了两个实用的函数哦。矩阵广义乘法gemm( )、最小二乘估计solve( )


补充

1)以例2为例,为什么状态值一般都设置成(x,y,△x,△y)?我们不妨设置成(x,y,△x),对应的转移矩阵也改成3×3的。可以看到仍能跟上,不过在x方向跟踪速度快,在y方向跟踪速度慢。进一步设置成(x,y)和2×2的转移矩阵,程序的跟踪速度简直是龟速。所以,简单理解,△x和△y严重影响对应方向上的跟踪速度。



3.实例

例1 OpenCV自带的示例程序


 
 
  
  
  1. #include "opencv2/video/tracking.hpp"
  2. #include "opencv2/highgui/highgui.hpp"
  3. #include <iostream>
  4. #include <stdio.h>
  5. using namespace std;
  6. using namespace cv;
  7. //计算相对窗口的坐标值,因为坐标原点在左上角,所以sin前有个负号
  8. static inline Point calcPoint(Point2f center, double R, double angle)
  9. {
  10. return center + Point2f(( float) cos(angle), ( float)- sin(angle))*( float)R;
  11. }
  12. static void help()
  13. {
  14. printf( "\nExamle of c calls to OpenCV's Kalman filter.\n"
  15. " Tracking of rotating point.\n"
  16. " Rotation speed is constant.\n"
  17. " Both state and measurements vectors are 1D (a point angle),\n"
  18. " Measurement is the real point angle + gaussian noise.\n"
  19. " The real and the estimated points are connected with yellow line segment,\n"
  20. " the real and the measured points are connected with red line segment.\n"
  21. " (if Kalman filter works correctly,\n"
  22. " the yellow segment should be shorter than the red one).\n"
  23. "\n"
  24. " Pressing any key (except ESC) will reset the tracking with a different speed.\n"
  25. " Pressing ESC will stop the program.\n"
  26. );
  27. }
  28. int main(int, char**)
  29. {
  30. help();
  31. Mat img(500, 500, CV_8UC3);
  32. KalmanFilter KF(2, 1, 0); //创建卡尔曼滤波器对象KF
  33. Mat state(2, 1, CV_32F); //state(角度,△角度)
  34. Mat processNoise(2, 1, CV_32F);
  35. Mat measurement = Mat::zeros( 1, 1, CV_32F); //定义测量值
  36. char code = ( char) -1;
  37. for(;;)
  38. {
  39. //1.初始化
  40. randn( state, Scalar::all( 0), Scalar::all( 0.1) ); //
  41. KF.transitionMatrix = *(Mat_< float>( 2, 2) << 1, 1, 0, 1); //转移矩阵A[1,1;0,1]
  42. //将下面几个矩阵设置为对角阵
  43. setIdentity(KF.measurementMatrix); //测量矩阵H
  44. setIdentity(KF.processNoiseCov, Scalar::all( 1e-5)); //系统噪声方差矩阵Q
  45. setIdentity(KF.measurementNoiseCov, Scalar::all( 1e-1)); //测量噪声方差矩阵R
  46. setIdentity(KF.errorCovPost, Scalar::all( 1)); //后验错误估计协方差矩阵P
  47. randn(KF.statePost, Scalar::all( 0), Scalar::all( 0.1)); //x(0)初始化
  48. for(;;)
  49. {
  50. Point2f center(img.cols*0.5f, img.rows*0.5f); //center图像中心点
  51. float R = img.cols/ 3.f; //半径
  52. double stateAngle = state.at< float>( 0); //跟踪点角度
  53. Point statePt = calcPoint(center, R, stateAngle); //跟踪点坐标statePt
  54. //2. 预测
  55. Mat prediction = KF.predict(); //计算预测值,返回x'
  56. double predictAngle = prediction.at< float>( 0); //预测点的角度
  57. Point predictPt = calcPoint(center, R, predictAngle); //预测点坐标predictPt
  58. //3.更新
  59. //measurement是测量值
  60. randn( measurement, Scalar::all( 0), Scalar::all(KF.measurementNoiseCov.at< float>( 0))); //给measurement赋值N(0,R)的随机值
  61. // generate measurement
  62. measurement += KF.measurementMatrix*state; //z = z + H*x;
  63. double measAngle = measurement.at< float>( 0);
  64. Point measPt = calcPoint(center, R, measAngle);
  65. // plot points
  66. //定义了画十字的方法,值得学习下
  67. #define drawCross( center, color, d ) \
  68. line( img, Point( center.x - d, center.y - d ), \
  69. Point( center.x + d, center.y + d ), color, 1, CV_AA, 0); \
  70. line( img, Point( center.x + d, center.y - d ), \
  71. Point( center.x - d, center.y + d ), color, 1, CV_AA, 0 )
  72. img = Scalar::all( 0);
  73. drawCross( statePt, Scalar( 255, 255, 255), 3 );
  74. drawCross( measPt, Scalar( 0, 0, 255), 3 );
  75. drawCross( predictPt, Scalar( 0, 255, 0), 3 );
  76. line( img, statePt, measPt, Scalar( 0, 0, 255), 3, CV_AA, 0 );
  77. line( img, statePt, predictPt, Scalar( 0, 255, 255), 3, CV_AA, 0 );
  78. //调用kalman这个类的correct方法得到加入观察值校正后的状态变量值矩阵
  79. if(theRNG().uniform( 0, 4) != 0)
  80. KF.correct(measurement);
  81. //不加噪声的话就是匀速圆周运动,加了点噪声类似匀速圆周运动,因为噪声的原因,运动方向可能会改变
  82. randn( processNoise, Scalar( 0), Scalar::all( sqrt(KF.processNoiseCov.at< float>( 0, 0)))); //vk
  83. state = KF.transitionMatrix*state + processNoise;
  84. imshow( "Kalman", img );
  85. code = ( char)waitKey( 100);
  86. if( code > 0 )
  87. break;
  88. }
  89. if( code == 27 || code == 'q' || code == 'Q' )
  90. break;
  91. }
  92. return 0;
  93. }
程序结果


例2  跟踪鼠标位置

在我介绍粒子滤波的博文“学习Opencv2——粒子滤波Condensation算法”里,有个例3,是跟踪鼠标位置。现在我们用卡尔曼滤波来实现。


 
 
  
  
  1. #include "opencv2/video/tracking.hpp"
  2. #include "opencv2/highgui/highgui.hpp"
  3. #include <stdio.h>
  4. using namespace cv;
  5. using namespace std;
  6. const int winHeight= 600;
  7. const int winWidth= 800;
  8. Point mousePosition= Point(winWidth>> 1,winHeight>> 1);
  9. //mouse event callback
  10. void mouseEvent(int event, int x, int y, int flags, void *param )
  11. {
  12. if (event==CV_EVENT_MOUSEMOVE) {
  13. mousePosition = Point(x,y);
  14. }
  15. }
  16. int main (void)
  17. {
  18. RNG rng;
  19. //1.kalman filter setup
  20. const int stateNum= 4; //状态值4×1向量(x,y,△x,△y)
  21. const int measureNum= 2; //测量值2×1向量(x,y)
  22. KalmanFilter KF(stateNum, measureNum, 0);
  23. KF.transitionMatrix = *(Mat_< float>( 4, 4) << 1, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 1); //转移矩阵A
  24. setIdentity(KF.measurementMatrix); //测量矩阵H
  25. setIdentity(KF.processNoiseCov, Scalar::all( 1e-5)); //系统噪声方差矩阵Q
  26. setIdentity(KF.measurementNoiseCov, Scalar::all( 1e-1)); //测量噪声方差矩阵R
  27. setIdentity(KF.errorCovPost, Scalar::all( 1)); //后验错误估计协方差矩阵P
  28. rng.fill(KF.statePost,RNG::UNIFORM, 0,winHeight>winWidth?winWidth:winHeight); //初始状态值x(0)
  29. Mat measurement = Mat::zeros(measureNum, 1, CV_32F); //初始测量值x'(0),因为后面要更新这个值,所以必须先定义
  30. namedWindow( "kalman");
  31. setMouseCallback( "kalman",mouseEvent);
  32. Mat image(winHeight,winWidth,CV_8UC3,Scalar(0));
  33. while ( 1)
  34. {
  35. //2.kalman prediction
  36. Mat prediction = KF.predict();
  37. Point predict_pt = Point(prediction.at< float>( 0),prediction.at< float>( 1) ); //预测值(x',y')
  38. //3.update measurement
  39. measurement.at< float>( 0) = ( float)mousePosition.x;
  40. measurement.at< float>( 1) = ( float)mousePosition.y;
  41. //4.update
  42. KF.correct(measurement);
  43. //draw
  44. image.setTo(Scalar( 255, 255, 255, 0));
  45. circle(image,predict_pt, 5,Scalar( 0, 255, 0), 3); //predicted point with green
  46. circle(image,mousePosition, 5,Scalar( 255, 0, 0), 3); //current position with red
  47. char buf[ 256];
  48. sprintf_s(buf, 256, "predicted position:(%3d,%3d)",predict_pt.x,predict_pt.y);
  49. putText(image,buf,Point( 10, 30),CV_FONT_HERSHEY_SCRIPT_COMPLEX, 1,Scalar( 0, 0, 0), 1, 8);
  50. sprintf_s(buf, 256, "current position :(%3d,%3d)",mousePosition.x,mousePosition.y);
  51. putText(image,buf,cvPoint( 10, 60),CV_FONT_HERSHEY_SCRIPT_COMPLEX, 1,Scalar( 0, 0, 0), 1, 8);
  52. imshow( "kalman", image);
  53. int key=waitKey( 3);
  54. if (key== 27){ //esc
  55. break;
  56. }
  57. }
  58. }

结果



例3 


  
  
   
   
  1. #include "opencv2/video/tracking.hpp"
  2. #include <opencv2/legacy/legacy.hpp> //#include "cvAux.h"
  3. #include <opencv2/highgui/highgui.hpp>
  4. #include <opencv2/core/core.hpp>
  5. #include <stdio.h>
  6. using namespace cv;
  7. using namespace std;
  8. int main( )
  9. {
  10. float A[ 10][ 3] =
  11. {
  12. 10, 50, 15.6,
  13. 12, 49, 16,
  14. 11, 52, 15.8,
  15. 13, 52.2, 15.8,
  16. 12.9, 50, 17,
  17. 14, 48, 16.6,
  18. 13.7, 49, 16.5,
  19. 13.6, 47.8, 16.4,
  20. 12.3, 46, 15.9,
  21. 13.1, 45, 16.2
  22. };
  23. const int stateNum= 3;
  24. const int measureNum= 3;
  25. KalmanFilter KF(stateNum, measureNum, 0);
  26. KF.transitionMatrix = *(Mat_< float>( 3, 3) << 1, 0, 0, 0, 1, 0, 0, 0, 1); //转移矩阵A
  27. setIdentity(KF.measurementMatrix); //测量矩阵H
  28. setIdentity(KF.processNoiseCov, Scalar::all( 1e-5)); //系统噪声方差矩阵Q
  29. setIdentity(KF.measurementNoiseCov, Scalar::all( 1e-1)); //测量噪声方差矩阵R
  30. setIdentity(KF.errorCovPost, Scalar::all( 1));
  31. Mat measurement = Mat::zeros(measureNum, 1, CV_32F);
  32. //初始状态值
  33. KF.statePost = *(Mat_< float>( 3, 1) <<A[ 0][ 0],A[ 0][ 1],A[ 0][ 2]);
  34. cout<< "state0="<<KF.statePost<< endl;
  35. for( int i= 1;i<= 9;i++)
  36. {
  37. //预测
  38. Mat prediction = KF.predict();
  39. //计算测量值
  40. measurement.at< float>( 0) = ( float)A[i][ 0];
  41. measurement.at< float>( 1) = ( float)A[i][ 1];
  42. measurement.at< float>( 2) = ( float)A[i][ 2];
  43. //更新
  44. KF.correct(measurement);
  45. //输出结果
  46. cout<< "predict ="<< "\t"<<prediction.at< float>( 0)<< "\t"<<prediction.at< float>( 1)<< "\t"<<prediction.at< float>( 2)<< endl;
  47. cout<< "measurement="<< "\t"<<measurement.at< float>( 0)<< "\t"<<measurement.at< float>( 1)<< "\t"<<measurement.at< float>( 2)<< endl;
  48. cout<< "correct ="<< "\t"<<KF.statePost.at< float>( 0)<< "\t"<<KF.statePost.at< float>( 1)<< "\t"<<KF.statePost.at< float>( 2)<< endl;
  49. }
  50. system( "pause");
  51. }
结果如下


这里预测值和上一个状态值一样,原因是转移矩阵A是单位阵,如果改成非单位阵,结果就不一样了。


猜你喜欢

转载自blog.csdn.net/ganbelieve/article/details/108076944