基於eigen實現 bundle adjustment

純粹是演示性代碼，用於展示BA過程，沒有做稀疏矩陣相關的優化

#include <iostream>
#include <opencv2/opencv.hpp>
#include <opencv2/features2d/features2d.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <Eigen/Core>
#include <Eigen/Dense>

#include "sophus/se3.h"

using namespace std;
using namespace Eigen;
using namespace cv;

typedef vector<Vector3d, Eigen::aligned_allocator<Vector3d>> VecVector3d;
typedef vector<Vector2d, Eigen::aligned_allocator<Vector3d>> VecVector2d;
typedef Matrix<double, 6, 1> Vector6d;

double fx = 520.9, fy = 521.0, cx = 325.1, cy = 249.7;

template<typename T>
Eigen::Matrix<T, 3, 3> RPY2mat(T roll, T pitch, T yaw)
{
    Eigen::Matrix<T, 3, 3> m;

    T cr = cos(roll);
    T sr = sin(roll);
    T cp = cos(pitch);
    T sp = sin(pitch);
    T cy = cos(yaw);
    T sy = sin(yaw);

    m(0,0) = cy * cp;
    m(0,1) = cy * sp * sr - sy * cr;
    m(0,2) = cy * sp * cr + sy * sr;
    m(1,0) = sy * cp;
    m(1,1) = sy * sp * sr + cy * cr;
    m(1,2) = sy * sp * cr - cy * sr;
    m(2,0) = - sp;
    m(2,1) = cp * sr;
    m(2,2) = cp * cr;
    return m;
}

void getPt3d(VecVector3d& p3d, int size)
{
	srand(time(NULL));
	
	for(int i = 0; i < size; i++)
	{
		double x = rand() % 100 - 100;
		double y = rand() % 100 - 100;
		double z = rand() % 100 + 500;
		
		p3d.push_back ( Vector3d( x, y, z) );
	}
}

void outputPt3d(const VecVector3d & p3ds)
{
	for(int i = 0; i < p3ds.size(); i++)
	{
		std::cout << p3ds[i](0) << " " << p3ds[i](0)  << " " << p3ds[i](0) << std::endl;
	}
}

void getUV(VecVector2d & p2ds,
        VecVector3d& p3d,
		Sophus::SE3 & SE3_RT);
		
int main()
{	
    // 1 生成位姿
	Eigen::Matrix3d R = Eigen::AngleAxisd(M_PI/4, Eigen::Vector3d(0, 0, 0.81)).toRotationMatrix();
	//Eigen::Matrix3d R = RPY2mat(0.1, 0.1, 0.0);
	Eigen::Vector3d t(1, 0, 0);
	

	// 從R, t 構造  SE3
	Sophus::SE3 SE3_RT1(R, t);
	
	Eigen::Matrix3d R2 = Eigen::AngleAxisd(M_PI/4, Eigen::Vector3d(0.81, 0, 1)).toRotationMatrix();
	Eigen::Vector3d t2(20, 10, 100);
	
	Sophus::SE3 SE3_RT2(R2, t2);
	
	Eigen::Matrix3d R3 = Eigen::AngleAxisd(M_PI/4, Eigen::Vector3d(0.52, 0, 1)).toRotationMatrix();
	Eigen::Vector3d t3(50, 30, 100);
	
	Sophus::SE3 SE3_RT3(R3, t3);
	
	vector<Sophus::SE3> poses = { SE3_RT1, SE3_RT2, SE3_RT3 };
	//vector<Sophus::SE3> poses = { SE3_RT1};
	
	// 2 生成3D 座標點
	const int POINT_SIZE = 20;
	
    VecVector3d p3d;
	getPt3d(p3d, POINT_SIZE);
	
	// 3 生成像素座標點
	vector<VecVector2d> pt2ds(poses.size());
	for(int i = 0; i < poses.size(); i++)
	{
		getUV(pt2ds[i], p3d, poses[i]);
	}
	
	//outputPt3d(p3d);
	
	//對座標點和誤差加 噪聲！
	for(int i = 0; i < POINT_SIZE; i++)
	{
		p3d[i](0) += 4;//((i % 2) ==0 ? -4 : 4);
	}
	
	//Eigen::Matrix3d R4 = Eigen::AngleAxisd(M_PI/4, Eigen::Vector3d(0.01, 0.0, 0.81)).toRotationMatrix();
	//Eigen::Matrix3d R4 = RPY2mat(0.15, 0.10, 0.0);
	//Eigen::Vector3d t4(1, 0, 0);
	//Sophus::SE3 SE3_RT4(R4, t4);
	//poses[0] = SE3_RT4;
	
	//outputPt3d(p3d);
	
    //高斯牛頓迭代
    int iterations = 10;

	// Jij 的維度爲 2 * (cols)
	// H 的維度爲 (cols) * (cols)
	// b 的維度爲 cols
	int cols = poses.size() * 6 + p3d.size() * 3;
	
	std::cout << "cols is " << cols << std::endl;
	
	double lastCost = 0;
	double cost = 0;
	
	const int POSE_SIZE = poses.size();
	
	for (int iter = 0; iter < iterations; iter++) 
	{
		MatrixXd H = Eigen::MatrixXd::Zero(cols, cols);
		MatrixXd b = Eigen::MatrixXd::Zero(cols, 1);
		
		cost = 0;		
				
		for(int i = 0; i < poses.size(); i++)
		{
			for(int j = 0; j < p3d.size(); j++)
			{
				// 遍歷每一個觀測, 每次觀測 對 Jij 中的兩個塊進行賦值
				
				Vector2d p2 = pt2ds[i][j];
				
				Vector3d p3 = p3d[j];
				
				Vector3d P = poses[i] * p3;
				double x = P[0];
				double y = P[1];
				double z = P[2];
				
				Vector2d p2_ = { fx * ( x/z ) + cx, fy * ( y/z ) + cy };
				Vector2d e = p2 - p2_;    //誤差error = 觀測值 - 估計值
				cost += (e[0]*e[0]+e[1]*e[1]);
				
				MatrixXd J = Eigen::MatrixXd::Zero(2, cols);
				
				// 和位姿相關的塊
				// 第一個位姿 要固定
				if(i != 0)
				{					
					J(0,6*i + 0) = -(fx/z);
					J(0,6*i + 1) = 0;
					J(0,6*i + 2) = (fx*x/(z*z));
					J(0,6*i + 3) = (fx*x*y/(z*z));
					J(0,6*i + 4) = -(fx*x*x/(z*z)+fx);
					J(0,6*i + 5) = (fx*y/z);
					J(1,6*i + 0) = 0;
					J(1,6*i + 1) = -(fy/z);
					J(1,6*i + 2) = (fy*y/(z*z));
					J(1,6*i + 3) = (fy*y*y/(z*z)+fy);
					J(1,6*i + 4) = -(fy*x*y/(z*z));
					J(1,6*i + 5) = -(fy*x/z);	
				}				

				
				// 和 3D 座標相關的塊				
				Eigen::Matrix<double, 2, 3> K;
				K << -(fx / z), 0, (fx *x / (z*z)), 0, -(fy / z), fy * y / (z*z);
				Matrix3d R = poses[i].rotation_matrix();
				
				J.block(0, 6 * POSE_SIZE + 3 * j, 2,3) = K * R;				
				
				H += J.transpose() * J;
				b += -J.transpose() * e;

			}			
		}
		
		// solve dx
        VectorXd dx(cols);
        dx = H.ldlt().solve(b);
		
		if (std::isnan(dx[0])) {
            cout << "result is nan!" << endl;
            break;
        }
	
		// 更新 解向量
		// 先更新位姿
		//std::cout << " dx " << dx << std::endl;
		for(int i = 1; i < poses.size(); i++)
		{
			poses[i] = Sophus::SE3::exp(dx.block(i*6, 0, 6, 1)) * poses[i];
		}		
		 
		int startIdx = poses.size() * 6;
		// 再更新 座標點
		for(int j = 0; j < p3d.size(); j++)
		{
			p3d[j] = dx.block(startIdx + j*3, 0, 3, 1) + p3d[j];
		}
		
		lastCost = cost;
        cout << "iteration " << iter << " cost=" << cost << endl;
	} 
	
	// TODO: 輸出結果	
	
	
    return 0;
}

// p2ds 是觀察到的像素點, 得保存對應的3D點的信息纔行
// 記錄每個 p3d 的所有觀測數據
// p3d.size()== p2ds.size()
void getUV(VecVector2d & p2ds,
        VecVector3d& p3d,
		Sophus::SE3 & SE3_RT)
{
	//先假設每個點能被每個相機觀察到，這樣比較簡單一點
	for(auto & p3 : p3d)
	{
		Vector3d P = SE3_RT * p3;
		double x = P[0];
		double y = P[1];
		double z = P[2];

		Vector2d p2_ = { fx * ( x/z ) + cx, fy * ( y/z ) + cy };	

		p2ds.push_back(p2_);
	}
	
	return ;
}