【发布时间】:2018-07-31 23:30:26
【问题描述】:
我有一辆带有漂亮 IMU 和声纳的车辆。我正在使用点云库线性 ICP 对声纳点云进行精细配准。我想将ICP变换的结果与IMU数据进行比较,但我不明白如何从最终的齐次4x4变换矩阵中提取平移。
类似的question和other sources我发现都说翻译只是表格中的第4列。
我遇到的问题是我获得的翻译值是不可能的,似乎旋转分量越大,得到的值越荒谬,这导致我相信我不能简单地提取最后一列。滚动、俯仰和偏航值在合理范围内且有意义,但在任何方向上的偏移量都不可能超过 1 米。矩阵在应用时确实按预期执行,所以我知道矩阵是正确的我只是不明白如何解释或提取 x、y、z 线性平移。
测量原始云的质心与最终云之间的距离会得到更合理的结果,但我不知道这是否是一种可接受的方法。看起来有点老套。
代码:
myCloud::Ptr target, source, output; // PCL clouds
myPoint cInit, cRough, cFinal; // centroid points
Eigen::Matrix4f estimation, icpResult, finalTransform; // transforms
// load vectors of sonar data into point clouds
target = pointVector_to_pointCloud(verbose, tgtPoints);
source = pointVector_to_pointCloud(verbose, srcPoints);
pcl::computeCentroid(*source, cInit);
// x, y, z offsets come from a previous rough alignment
Eigen::Affine3f fromIMU(Eigen::Translation3f(x, y, z));
estimation = fromIMU.matrix();
pcl::transformPointCloud(*cloud, *cloud, estimation);
pcl::computeCentroid(*source, cRough);
// create new empty cloud in the output pointer, set up ICP
output.reset(new myCloud);
icp.setInputSource(source);
icp.setInputTarget(target);
/**** Set ICP parameters, omitted ****/
icp.align(*output);
icpResult = icp.getFinalTransformation();
finalTransform = estimation * icpResult;
pcl::computeCentroid(*source, cFinal);
// Output Results
Eigen::Affine3f roughT(estimation);
Eigen::Affine3f fineT(icpResult);
float tx, ty, tz, rx, ry, rz;
pcl::getTranslationAndEulerAngles(roughT, tx, ty, tz, rx, ry, rz);
std::cerr << "********* ICP RESULTS **********\n";
std::cerr << "Rough Transform Matrix:\n" << transform << endl;
std::cerr << "Translation (x, y, z) : " << tx << ", " << ty << ", " << tz << endl;
std::cerr << "Rotation (roll, pitch, yaw) : " << rx << ", " << ry << ", " << rz << endl;
pcl::getTranslationAndEulerAngles(fineT, tx, ty, tz, rx, ry, rz);
std::cerr << "\nFine Transform Matrix:\n" << icpResult << endl;
std::cerr << "Translation (x, y, z) : " << tx << ", " << ty << ", " << tz << endl;
std::cerr << "Rotation (roll, pitch, yaw) : " << rx << ", " << ry << ", " << rz << endl << endl;
std::cerr << "\nFinal Transformation Matrix:\n" << finalTransform << endl;
std::cerr << "\n\tCentroid after Rough Alignment: " << cRough << " ... Distance From Start: " << pcl::geometry::distance(cInit, cRough) << endl;
std::cerr << "\tCentroid after ICP: " << cFinal << " ... Distance From Start: " << pcl::geometry::distance(cInit, cFinal) << endl;
哪些输出(对于一个示例数据集):
********* INSIDE ICP TRANSFORM STATS **********
Rough Transform Matrix:
1 0 0 0.612095
0 1 0 -0.211855
0 0 1 0
0 0 0 1
Translation (x, y, z) : 0.612095, -0.211855, 0
Rotation (roll, pitch, yaw) : 0, -0, 0
Fine Transform Matrix:
0.999992 -0.00257317 0.00361636 2.92558
0.00256172 0.999995 0.00328003 2.66182
-0.00362478 -0.00327113 0.999988 0.0578782
0 0 0 1
Translation (x, y, z) : 2.92558, 2.66182, 0.0578782
Rotation (roll, pitch, yaw) : -0.00327116, 0.00362479, 0.00256174
Final Transformation Matrix:
0.999992 -0.00257317 0.00361636 3.53767
0.00256172 0.999995 0.00328003 2.44996
-0.00362478 -0.00327113 0.999988 0.0578782
0 0 0 1
Centroid after Rough Alignment: (8.8218,9.12704,-807.301 - 0,126,255) ... Distance From Start: 0.647709
Centroid after ICP: (8.8068,9.1658,-807.3 - 0,126,255) ... Distance From Start: 0.621667
【问题讨论】:
-
可以接受什么?你不是在制定规则吗?平移将取决于点云的原点在哪里。如果它在外面很远,旋转部分已经将点云移动了很多。然后,翻译只会添加到这个动作中。使用质心听起来很合理。
标签: c++ math matrix point-cloud-library point-clouds