本文转自:http://blog.csdn.net/yihaizhiyan/article/details/5973729
http://blog.csdn.net/Sway_2012/article/details/37765765
http://blog.csdn.net/zouwen198317/article/details/38494149
1.什么是RANSAC?
RANSAC是RANdom SAmple Consensus(随机抽样一致性)的缩写。它是从一个观察数据集合中,估计模型参数(模型拟合)的迭代方法。它是一种随机的不确定算法,每次运算求出的结果可能不相同,但总能给出一个合理的结果,为了提高概率必须提高迭代次数。
2.算法详解
给定两个点p1与p2的坐标,确定这两点所构成的直线,要求对于输入的任意点p3,都可以判断它是否在该直线上。初中解析几何知识告诉我们,判断一个点在直线上,只需其与直线上任意两点点斜率都相同即可。实际操作当中,往往会先根据已知的两点算出直线的表达式(点斜式、截距式等等),然后通过向量计算即可方便地判断p3是否在该直线上。
生产实践中的数据往往会有一定的偏差。例如我们知道两个变量X与Y之间呈线性关系,Y=aX+b,我们想确定参数a与b的具体值。通过实验,可以得到一组X与Y的测试值。虽然理论上两个未知数的方程只需要两组值即可确认,但由于系统误差的原因,任意取两点算出的a与b的值都不尽相同。我们希望的是,最后计算得出的理论模型与测试值的误差最小。大学的高等数学课程中,详细阐述了最小二乘法的思想。通过计算最小均方差关于参数a、b的偏导数为零时的值。事实上,在很多情况下,最小二乘法都是线性回归的代名词。
遗憾的是,最小二乘法只适合与误差较小的情况。试想一下这种情况,假使需要从一个噪音较大的数据集中提取模型(比方说只有20%的数据时符合模型的)时,最小二乘法就显得力不从心了。例如下图,肉眼可以很轻易地看出一条直线(模式),但算法却找错了。
RANSAC算法的输入是一组观测数据(往往含有较大的噪声或无效点),一个用于解释观测数据的参数化模型以及一些可信的参数。RANSAC通过反复选择数据中的一组随机子集来达成目标。被选取的子集被假设为局内点,并用下述方法进行验证:
- 有一个模型适应于假设的局内点,即所有的未知参数都能从假设的局内点计算得出。
- 用1中得到的模型去测试所有的其它数据,如果某个点适用于估计的模型,认为它也是局内点。
- 如果有足够多的点被归类为假设的局内点,那么估计的模型就足够合理。
- 然后,用所有假设的局内点去重新估计模型(譬如使用最小二乘法),因为它仅仅被初始的假设局内点估计过。
- 最后,通过估计局内点与模型的错误率来评估模型。
- 上述过程被重复执行固定的次数,每次产生的模型要么因为局内点太少而被舍弃,要么因为比现有的模型更好而被选用。
整个过程可参考下图:
3.代码实现
随机一致性采样RANSAC是一种鲁棒的模型拟合算法,能够从有外点的数据中拟合准确的模型。
RANSAC过程中用到的参数
N-- 拟合模型所需要的最少的样本个数
K--算法的迭代次数
t--用于判断数据是否是内点
d--判定模型是否符合使用于数据集,也就是判断是否是好的模型
RANSAC算法过程
1 for K 次迭代
2 从数据中均匀随机采样N个点
3 利用采样的N个点拟合你个模型
4 for 对于除采样点外的每一个样本点
5 利用t检测样本点到模型的距离,如果小于t则认为是一致,否则认为是外点
6 end
7 如果有d或者更多的一致点,则认为拟合的模型是好的
8 end
9 使用拟合误差作为标准,选择最好的拟合模型
迭代次数的计算
假设 r = 内点个数/所有点的个数
则:
p0 = pow(r, N) 表示采样的N个点全为内点,也就是是一次有效采样的概率
p1 = 1 - pow(r, N) 表示采样的N个点中至少有一个外点,即一次无效采样的概率
p2 = pow(p1, K) 表示K次无效采样的概率
假设p表示K次采样中至少一次采样是有效采样,则有1-p = pow(p1, K), 两边取对数
则有 K = log(1- p )/log(1-p1).
附一份来自google 的RANSAC的代码框架
- #ifndef FVISION_RANSAC_H_
- #define FVISION_RANSAC_H_
- #include <fvision/utils/random_utils.h>
- #include <fvision/utils/misc.h>
- #include <vector>
- #include <iostream>
- #include <cassert>
- namespace fvision {
- class RANSAC_SamplesNumber {
- public:
- RANSAC_SamplesNumber(int modelSampleSize) {
- this->s = modelSampleSize;
- this->p = 0.99;
- }
- ~RANSAC_SamplesNumber(void) {}
- public:
- long calcN(int inliersNumber, int samplesNumber) {
- double e = 1 - (double)inliersNumber / samplesNumber;
- //cout<<"e: "<<e<<endl;
- if (e > 0.9) e = 0.9;
- //cout<<"pow: "<<pow((1 - e), s)<<endl;
- //cout<<log(1 - pow((1 - e), s))<<endl;
- long N = (long)(log(1 - p) / log(1 - pow((1 - e), s)));
- if (N < 0) return (long)1000000000;
- else return N;
- }
- private:
- int s; //samples size for fitting a model
- double p; //probability that at least one of the random samples if free from outliers
- //usually 0.99
- };
- //fit a model to a set of samples
- template <typename M, typename S>
- class GenericModelCalculator {
- public:
- typedef std::vector<S> Samples;
- virtual M compute(const Samples& samples) = 0;
- virtual ~GenericModelCalculator<M, S>() {}
- //the model calculator may only use a subset of the samples for computing
- //default return empty for both
- virtual const std::vector<int>& getInlierIndices() const { return defaultInlierIndices; };
- virtual const std::vector<int>& getOutlierIndices() const { return defaultOutlierIndices; };
- // if the subclass has a threshold parameter, it need to override the following three functions
- // this is used for algorithms which have a normalization step on input samples
- virtual bool hasThreshold() const { return false; }
- virtual void setThreshold(double threshold) {}
- virtual double getThreshold() const { return 0; }
- protected:
- std::vector<int> defaultInlierIndices;
- std::vector<int> defaultOutlierIndices;
- };
- //evaluate a model to samples
- //using a threshold to distinguish inliers and outliers
- template <typename M, typename S>
- class GenericErrorCaclculator {
- public:
- virtual ~GenericErrorCaclculator<M, S>() {}
- typedef std::vector<S> Samples;
- virtual double compute(const M& model, const S& sample) const = 0;
- double computeAverage(const M& model, const Samples& samples) const {
- int n = (int)samples.size();
- if (n == 0) return 0;
- double sum = 0;
- for (int i = 0; i < n; i++) {
- sum += compute(model, samples[i]);
- }
- return sum / n;
- }
- double computeInlierAverage(const M& model, const Samples& samples) const {
- int n = (int)samples.size();
- if (n == 0) return 0;
- double sum = 0;
- double error = 0;
- int inlierNum = 0;
- for (int i = 0; i < n; i++) {
- error = compute(model, samples[i]);
- if (error <= threshold) {
- sum += error;
- inlierNum++;
- }
- }
- if (inlierNum == 0) return 1000000;
- return sum / inlierNum;
- }
- public:
- /** set a threshold for classify inliers and outliers
- */
- void setThreshold(double v) { threshold = v; }
- double getThreshold() const { return threshold; }
- /** classify all samples to inliers and outliers
- *
- */
- void classify(const M& model, const Samples& samples, Samples& inliers, Samples& outliers) const {
- inliers.clear();
- outliers.clear();
- Samples::const_iterator iter = samples.begin();
- for (; iter != samples.end(); ++iter) {
- if (isInlier(model, *iter)) inliers.push_back(*iter);
- else outliers.push_back(*iter);
- }
- }
- /** classify all samples to inliers and outliers, output indices
- *
- */
- void classify(const M& model, const Samples& samples, std::vector<int>& inlierIndices, std::vector<int>& outlierIndices) const {
- inlierIndices.clear();
- outlierIndices.clear();
- Samples::const_iterator iter = samples.begin();
- int i = 0;
- for (; iter != samples.end(); ++iter, ++i) {
- if (isInlier(model, *iter)) inlierIndices.push_back(i);
- else outlierIndices.push_back(i);
- }
- }
- /** classify all samples to inliers and outliers
- *
- */
- void classify(const M& model, const Samples& samples,
- std::vector<int>& inlierIndices, std::vector<int>& outlierIndices,
- Samples& inliers, Samples& outliers) const {
- inliers.clear();
- outliers.clear();
- inlierIndices.clear();
- outlierIndices.clear();
- Samples::const_iterator iter = samples.begin();
- int i = 0;
- for (; iter != samples.end(); ++iter, ++i) {
- if (isInlier(model, *iter)) {
- inliers.push_back(*iter);
- inlierIndices.push_back(i);
- }
- else {
- outliers.push_back(*iter);
- outlierIndices.push_back(i);
- }
- }
- }
- int calcInliersNumber(const M& model, const Samples& samples) const {
- int n = 0;
- for (int i = 0; i < (int)samples.size(); i++) {
- if (isInlier(model, samples[i])) ++n;
- }
- return n;
- }
- bool isInlier(const M& model, const S& sample) const {
- return (compute(model, sample) <= threshold);
- }
- private:
- double threshold;
- };
- /** generic RANSAC framework
- * make use of a model calculator and an error calculator
- * M is the model type, need to support copy assignment operator and default constructor
- * S is the sample type.
- *
- * Interface:
- * M compute(samples); input a set of samples, output a model.
- * after compute, inliers and outliers can be retrieved
- *
- */
- template <typename M, typename S>
- class Ransac : public GenericModelCalculator<M, S> {
- public:
- typedef std::vector<S> Samples;
- /** Constructor
- *
- * @param pmc a GenericModelCalculator object
- * @param modelSampleSize how much samples are used to fit a model
- * @param pec a GenericErrorCaclculator object
- */
- Ransac(GenericModelCalculator<M, S>* pmc, int modelSampleSize, GenericErrorCaclculator<M, S>* pec) {
- this->pmc = pmc;
- this->modelSampleSize = modelSampleSize;
- this->pec = pec;
- this->maxSampleCount = 500;
- this->minInliersNum = 1000000;
- this->verbose = false;
- }
- const GenericErrorCaclculator<M, S>* getErrorCalculator() const { return pec; }
- virtual ~Ransac() {
- delete pmc;
- delete pec;
- }
- void setMaxSampleCount(int n) {
- this->maxSampleCount = n;
- }
- void setMinInliersNum(int n) {
- this->minInliersNum = n;
- }
- virtual bool hasThreshold() const { return true; }
- virtual void setThreshold(double threshold) {
- pec->setThreshold(threshold);
- }
- virtual double getThreshold() const {
- return pec->getThreshold();
- }
- public:
- /** Given samples, compute a model that has most inliers. Assume the samples size is larger or equal than model sample size
- * inliers, outliers, inlierIndices and outlierIndices are stored
- *
- */
- M compute(const Samples& samples) {
- clear();
- int pointsNumber = (int)samples.size();
- assert(pointsNumber >= modelSampleSize);
- long N = 100000;
- int sampleCount = 0;
- RANSAC_SamplesNumber ransac(modelSampleSize);
- M bestModel;
- int maxInliersNumber = 0;
- bool stop = false;
- while (sampleCount < N && sampleCount < maxSampleCount && !stop) {
- Samples nsamples;
- randomlySampleN(samples, nsamples, modelSampleSize);
- M sampleModel = pmc->compute(nsamples);
- if (maxInliersNumber == 0) bestModel = sampleModel; //init bestModel
- int inliersNumber = pec->calcInliersNumber(sampleModel, samples);
- if (verbose) std::cout<<"inliers number: "<<inliersNumber<<std::endl;
- if (inliersNumber > maxInliersNumber) {
- bestModel = sampleModel;
- maxInliersNumber = inliersNumber;
- N = ransac.calcN(inliersNumber, pointsNumber);
- if (maxInliersNumber > minInliersNum) stop = true;
- }
- if (verbose) std::cout<<"N: "<<N<<std::endl;
- sampleCount ++;
- }
- if (verbose) std::cout<<"sampleCount: "<<sampleCount<<std::endl;
- finalModel = computeUntilConverge(bestModel, maxInliersNumber, samples);
- pec->classify(finalModel, samples, inlierIndices, outlierIndices, inliers, outliers);
- inliersRate = (double)inliers.size() / samples.size();
- return finalModel;
- }
- const Samples& getInliers() const { return inliers; }
- const Samples& getOutliers() const { return outliers; }
- const std::vector<int>& getInlierIndices() const { return inlierIndices; }
- const std::vector<int>& getOutlierIndices() const { return outlierIndices; }
- double getInliersAverageError() const {
- return pec->computeAverage(finalModel, inliers);
- }
- double getInliersRate() const {
- return inliersRate;
- }
- void setVerbose(bool v) {
- verbose = v;
- }
- private:
- void randomlySampleN(const Samples& samples, Samples& nsamples, int sampleSize) {
- std::vector<int> is = ranis((int)samples.size(), sampleSize);
- for (int i = 0; i < sampleSize; i++) {
- nsamples.push_back(samples[is[i]]);
- }
- }
- /** from initial model, iterate to find the best model.
- *
- */
- M computeUntilConverge(M initModel, int initInliersNum, const Samples& samples) {
- if (verbose) {
- std::cout<<"iterate until converge...."<<std::endl;
- std::cout<<"init inliers number: "<<initInliersNum<<std::endl;
- }
- M bestModel = initModel;
- M newModel = initModel;
- int lastInliersNum = initInliersNum;
- Samples newInliers, newOutliers;
- pec->classify(initModel, samples, newInliers, newOutliers);
- double lastInlierAverageError = pec->computeAverage(initModel, newInliers);
- if (verbose) std::cout<<"init inlier average error: "<<lastInlierAverageError<<std::endl;
- while (true && (int)newInliers.size() >= modelSampleSize) {
- //update new model with new inliers, the new model does not necessarily have more inliers
- newModel = pmc->compute(newInliers);
- pec->classify(newModel, samples, newInliers, newOutliers);
- int newInliersNum = (int)newInliers.size();
- double newInlierAverageError = pec->computeAverage(newModel, newInliers);
- if (verbose) {
- std::cout<<"new inliers number: "<<newInliersNum<<std::endl;
- std::cout<<"new inlier average error: "<<newInlierAverageError<<std::endl;
- }
- if (newInliersNum < lastInliersNum) break;
- if (newInliersNum == lastInliersNum && newInlierAverageError >= lastInlierAverageError) break;
- //update best model with the model has more inliers
- bestModel = newModel;
- lastInliersNum = newInliersNum;
- lastInlierAverageError = newInlierAverageError;
- }
- return bestModel;
- }
- void clear() {
- inliers.clear();
- outliers.clear();
- inlierIndices.clear();
- outlierIndices.clear();
- }
- private:
- GenericModelCalculator<M, S>* pmc;
- GenericErrorCaclculator<M, S>* pec;
- int modelSampleSize;
- int maxSampleCount;
- int minInliersNum;
- M finalModel;
- Samples inliers;
- Samples outliers;
- std::vector<int> inlierIndices;
- std::vector<int> outlierIndices;
- double inliersRate;
- private:
- bool verbose;
- };
- }
- #endif // FVISION_RANSAC_H_
实例2
- #include <math.h>
- #include "LineParamEstimator.h"
- LineParamEstimator::LineParamEstimator(double delta) : m_deltaSquared(delta*delta) {}
- /*****************************************************************************/
- /*
- * Compute the line parameters [n_x,n_y,a_x,a_y]
- * 通过输入的两点来确定所在直线,采用法线向量的方式来表示,以兼容平行或垂直的情况
- * 其中n_x,n_y为归一化后,与原点构成的法线向量,a_x,a_y为直线上任意一点
- */
- void LineParamEstimator::estimate(std::vector<Point2D *> &data,
- std::vector<double> ¶meters)
- {
- parameters.clear();
- if(data.size()<2)
- return;
- double nx = data[1]->y - data[0]->y;
- double ny = data[0]->x - data[1]->x;// 原始直线的斜率为K,则法线的斜率为-1/k
- double norm = sqrt(nx*nx + ny*ny);
- parameters.push_back(nx/norm);
- parameters.push_back(ny/norm);
- parameters.push_back(data[0]->x);
- parameters.push_back(data[0]->y);
- }
- /*****************************************************************************/
- /*
- * Compute the line parameters [n_x,n_y,a_x,a_y]
- * 使用最小二乘法,从输入点中拟合出确定直线模型的所需参量
- */
- void LineParamEstimator::leastSquaresEstimate(std::vector<Point2D *> &data,
- std::vector<double> ¶meters)
- {
- double meanX, meanY, nx, ny, norm;
- double covMat11, covMat12, covMat21, covMat22; // The entries of the symmetric covarinace matrix
- int i, dataSize = data.size();
- parameters.clear();
- if(data.size()<2)
- return;
- meanX = meanY = 0.0;
- covMat11 = covMat12 = covMat21 = covMat22 = 0;
- for(i=0; i<dataSize; i++) {
- meanX +=data[i]->x;
- meanY +=data[i]->y;
- covMat11 +=data[i]->x * data[i]->x;
- covMat12 +=data[i]->x * data[i]->y;
- covMat22 +=data[i]->y * data[i]->y;
- }
- meanX/=dataSize;
- meanY/=dataSize;
- covMat11 -= dataSize*meanX*meanX;
- covMat12 -= dataSize*meanX*meanY;
- covMat22 -= dataSize*meanY*meanY;
- covMat21 = covMat12;
- if(covMat11<1e-12) {
- nx = 1.0;
- ny = 0.0;
- }
- else { //lamda1 is the largest eigen-value of the covariance matrix
- //and is used to compute the eigne-vector corresponding to the smallest
- //eigenvalue, which isn't computed explicitly.
- double lamda1 = (covMat11 + covMat22 + sqrt((covMat11-covMat22)*(covMat11-covMat22) + 4*covMat12*covMat12)) / 2.0;
- nx = -covMat12;
- ny = lamda1 - covMat22;
- norm = sqrt(nx*nx + ny*ny);
- nx/=norm;
- ny/=norm;
- }
- parameters.push_back(nx);
- parameters.push_back(ny);
- parameters.push_back(meanX);
- parameters.push_back(meanY);
- }
- /*****************************************************************************/
- /*
- * Given the line parameters [n_x,n_y,a_x,a_y] check if
- * [n_x, n_y] dot [data.x-a_x, data.y-a_y] < m_delta
- * 通过与已知法线的点乘结果,确定待测点与已知直线的匹配程度;结果越小则越符合,为
- * 零则表明点在直线上
- */
- bool LineParamEstimator::agree(std::vector<double> ¶meters, Point2D &data)
- {
- double signedDistance = parameters[0]*(data.x-parameters[2]) + parameters[1]*(data.y-parameters[3]);
- return ((signedDistance*signedDistance) < m_deltaSquared);
- }
#include <math.h> #include "LineParamEstimator.h" LineParamEstimator::LineParamEstimator(double delta) : m_deltaSquared(delta*delta) {} /*****************************************************************************/ /* * Compute the line parameters [n_x,n_y,a_x,a_y] * 通过输入的两点来确定所在直线,采用法线向量的方式来表示,以兼容平行或垂直的情况 * 其中n_x,n_y为归一化后,与原点构成的法线向量,a_x,a_y为直线上任意一点 */ void LineParamEstimator::estimate(std::vector<Point2D *> &data, std::vector<double> ¶meters) { parameters.clear(); if(data.size()<2) return; double nx = data[1]->y - data[0]->y; double ny = data[0]->x - data[1]->x;// 原始直线的斜率为K,则法线的斜率为-1/k double norm = sqrt(nx*nx + ny*ny); parameters.push_back(nx/norm); parameters.push_back(ny/norm); parameters.push_back(data[0]->x); parameters.push_back(data[0]->y); } /*****************************************************************************/ /* * Compute the line parameters [n_x,n_y,a_x,a_y] * 使用最小二乘法,从输入点中拟合出确定直线模型的所需参量 */ void LineParamEstimator::leastSquaresEstimate(std::vector<Point2D *> &data, std::vector<double> ¶meters) { double meanX, meanY, nx, ny, norm; double covMat11, covMat12, covMat21, covMat22; // The entries of the symmetric covarinace matrix int i, dataSize = data.size(); parameters.clear(); if(data.size()<2) return; meanX = meanY = 0.0; covMat11 = covMat12 = covMat21 = covMat22 = 0; for(i=0; i<dataSize; i++) { meanX +=data[i]->x; meanY +=data[i]->y; covMat11 +=data[i]->x * data[i]->x; covMat12 +=data[i]->x * data[i]->y; covMat22 +=data[i]->y * data[i]->y; } meanX/=dataSize; meanY/=dataSize; covMat11 -= dataSize*meanX*meanX; covMat12 -= dataSize*meanX*meanY; covMat22 -= dataSize*meanY*meanY; covMat21 = covMat12; if(covMat11<1e-12) { nx = 1.0; ny = 0.0; } else { //lamda1 is the largest eigen-value of the covariance matrix //and is used to compute the eigne-vector corresponding to the smallest //eigenvalue, which isn't computed explicitly. double lamda1 = (covMat11 + covMat22 + sqrt((covMat11-covMat22)*(covMat11-covMat22) + 4*covMat12*covMat12)) / 2.0; nx = -covMat12; ny = lamda1 - covMat22; norm = sqrt(nx*nx + ny*ny); nx/=norm; ny/=norm; } parameters.push_back(nx); parameters.push_back(ny); parameters.push_back(meanX); parameters.push_back(meanY); } /*****************************************************************************/ /* * Given the line parameters [n_x,n_y,a_x,a_y] check if * [n_x, n_y] dot [data.x-a_x, data.y-a_y] < m_delta * 通过与已知法线的点乘结果,确定待测点与已知直线的匹配程度;结果越小则越符合,为 * 零则表明点在直线上 */ bool LineParamEstimator::agree(std::vector<double> ¶meters, Point2D &data) { double signedDistance = parameters[0]*(data.x-parameters[2]) + parameters[1]*(data.y-parameters[3]); return ((signedDistance*signedDistance) < m_deltaSquared); }
RANSAC寻找匹配的代码如下:
- /*****************************************************************************/
- template<class T, class S>
- double Ransac<T,S>::compute(std::vector<S> ¶meters,
- ParameterEsitmator<T,S> *paramEstimator ,
- std::vector<T> &data,
- int numForEstimate)
- {
- std::vector<T *> leastSquaresEstimateData;
- int numDataObjects = data.size();
- int numVotesForBest = -1;
- int *arr = new int[numForEstimate];// numForEstimate表示拟合模型所需要的最少点数,对本例的直线来说,该值为2
- short *curVotes = new short[numDataObjects]; //one if data[i] agrees with the current model, otherwise zero
- short *bestVotes = new short[numDataObjects]; //one if data[i] agrees with the best model, otherwise zero
- //there are less data objects than the minimum required for an exact fit
- if(numDataObjects < numForEstimate)
- return 0;
- // 计算所有可能的直线,寻找其中误差最小的解。对于100点的直线拟合来说,大约需要100*99*0.5=4950次运算,复杂度无疑是庞大的。一般采用随机选取子集的方式。
- computeAllChoices(paramEstimator,data,numForEstimate,
- bestVotes, curVotes, numVotesForBest, 0, data.size(), numForEstimate, 0, arr);
- //compute the least squares estimate using the largest sub set
- for(int j=0; j<numDataObjects; j++) {
- if(bestVotes[j])
- leastSquaresEstimateData.push_back(&(data[j]));
- }
- // 对局内点再次用最小二乘法拟合出模型
- paramEstimator->leastSquaresEstimate(leastSquaresEstimateData,parameters);
- delete [] arr;
- delete [] bestVotes;
- delete [] curVotes;
- return (double)leastSquaresEstimateData.size()/(double)numDataObjects;
- }