博主创建了一个科研互助群Q：772356582，欢迎大家加入讨论。这是一个科研互助群，主要围绕机器人，无人驾驶，无人机方面的感知定位，决策规划，以及论文发表经验，以方便大家很好很快的科研，少走弯路。欢迎在群里积极提问与回答，相互交流共同学习。
一、简介
PL-VINS是基于最先进的基于点的VINS- mono，开发的一种基于点和线特征的实时、高效优化的单目VINS方法。原始的 PL-VINS 是在ubuntu18.04基于opencv3去开发的。源码的下载地址：https://github.com/cnqiangfu/PL-VINS
我的配置：ubuntu20.04+opencv4.2+eigen3.3.7
更改好的代码如下，可直接用，需要修改mage_node_b.cpp的main函数第一行的地址
huashu996/Ubuntu20.04PL_VINS · GitHub
二、编译
如果你使用官方的代码，编译时候可能会遇到如下问题
mkdir -p ~/catkin_plvins/src
cd catkin_plvins/src //进入创建的catkin_plvins/src文件夹下
catkin_init_workspace 进行空间创建cd ~/catkin_plvins //在文件夹catkin_plvins下建立终端输入
catkin_make //终端输入
source devel/setup.bash
echo $ROS_PACKAGE_PATH//将代码下载到src目录下 或者执行下面代码
cd ~/catkin_plvins/src
git clone https://github.com/cnqiangfu/PL-VINS.git
//编译
cd .. //回到文件夹catkin_plvins
catkin_make
source devel/setup.bash- 将feature_tracker里面的Cmakelist修改一下：主要修改以下2处的路径
- 打开catkin_plvins/src/PL-VINS/image_node_b文件下的CMakeLists.txt添加
set(CMAKE_CXX_STANDARD 14)如果你配置正确，应该能够编译顺利通过
三、opencv4适配
但运行往往会出现如下图问题，对于每个问题我们一次解决
- [image_node_b-7] 挂掉的原因是linefeature_tracker_node中只发布了归一化坐标，没有发布linefeature的startpoint和endpoint的像素坐标。
1.mage_node_b.cpp的main函数第一行的地址，修改原绝对地址。
2.加入project函数
void project(cv::Point2f& pt, cv::Mat const& k)
{pt.x=k.at<float>(0,0)*pt.x+k.at<float>(0,2);pt.y=k.at<float>(1,1)*pt.y+k.at<float>(1,2);
}3.替换
将下面两行
cv::Point startPoint = cv::Point(line_feature_msg->channels[3].values[i], line_feature_msg->channels[4].values[i]);
cv::Point endPoint = cv::Point(line_feature_msg->channels[5].values[i], line_feature_msg->channels[6].values[i]);
替换
cv::Point2f startPoint = cv::Point2f(line_feature_msg->points[i].x,line_feature_msg->points[i].y );project (startPoint,K_);
cv::Point2f endPoint = cv::Point2f(line_feature_msg->channels[1].values[i],line_feature_msg->channels[2].values[i]);- [linefeature_tracker-4]
此问题由于opencv和cv_bridge冲突的问题，因为在ubuntu20.04 cv_bridge是4，如果你使用opencv3那么将产生冲突，如果使用opencv4跑，代码本身又不支持opencv4所以需要更改。相应的解决方式也有两种：
1）把vins-mono代码全部改成opencv4的版本 或者 2）把cv_bridge改成opencv3版本。参考如下博客修改
【精选】Ubuntu20.04下成功运行VINS-mono_ubuntu vinsmono-CSDN博客
1.将所有包含opencv的Cmakelists.txt中opencv引入都换成4，如下
2.更新头文件使使用opencv4
(1)将camera_model包改成兼容opencv4在camera_model包中的头文件Chessboard.h中添加
#include <opencv2/imgproc/types_c.h>
#include <opencv2/calib3d/calib3d_c.h>
在CameraCalibration.h中添加
#include <opencv2/imgproc/types_c.h>
#include <opencv2/imgproc/imgproc_c.h>(2)将包中所有报错的头文件
#include <opencv/cv.h>
#include <opencv/highgui.h>
替换为
#include <opencv2/highgui.hpp>
#include <opencv2/cvconfig.h>
3.加入线特征检测函数
imgproc.hpp关于线特征的部分由于opencv4权限的问题无法使用了，这里只需要自己定义一个相同功能的头文件去把opencv3中的实现拷贝过去。在路径
PL-VINS/src/PL-VINS/feature_tracker/src/line_descriptor/src
创建一个my_lsd.hpp
//
// Created by fs on 2021/11/3.
// This file is my_lsd.hpp
//
//#include "opencv2/../../src/precomp.hpp"
#include "opencv2/imgproc.hpp"
//#include "opencv2/core/private.hpp"
#include <vector>#define M_3_2_PI (3 * CV_PI) / 2 // 3/2 pi
#define M_2__PI (2 * CV_PI) // 2 pi#ifndef M_LN10
#define M_LN10 2.30258509299404568402
#endif#define NOTDEF double(-1024.0) // Label for pixels with undefined gradient.#define NOTUSED 0 // Label for pixels not used in yet.
#define USED 1 // Label for pixels already used in detection.#define RELATIVE_ERROR_FACTOR 100.0const double DEG_TO_RADS = CV_PI / 180;#define log_gamma(x) ((x)>15.0?log_gamma_windschitl(x):log_gamma_lanczos(x))struct edge
{cv::Point p;bool taken;
};inline double distSq(const double x1, const double y1,const double x2, const double y2)
{return (x2 - x1)*(x2 - x1) + (y2 - y1)*(y2 - y1);
}inline double dist(const double x1, const double y1,const double x2, const double y2)
{return sqrt(distSq(x1, y1, x2, y2));
}// Signed angle difference
inline double angle_diff_signed(const double& a, const double& b)
{double diff = a - b;while(diff <= -CV_PI) diff += M_2__PI;while(diff > CV_PI) diff -= M_2__PI;return diff;
}// Absolute value angle difference
inline double angle_diff(const double& a, const double& b)
{return std::fabs(angle_diff_signed(a, b));
}// Compare doubles by relative error.
inline bool double_equal(const double& a, const double& b)
{// trivial caseif(a == b) return true;double abs_diff = fabs(a - b);double aa = fabs(a);double bb = fabs(b);double abs_max = (aa > bb)? aa : bb;if(abs_max < DBL_MIN) abs_max = DBL_MIN;return (abs_diff / abs_max) <= (RELATIVE_ERROR_FACTOR * DBL_EPSILON);
}inline bool AsmallerB_XoverY(const edge& a, const edge& b)
{if (a.p.x == b.p.x) return a.p.y < b.p.y;else return a.p.x < b.p.x;
}/*** Computes the natural logarithm of the absolute value of* the gamma function of x using Windschitl method.* See http://www.rskey.org/gamma.htm*/
inline double log_gamma_windschitl(const double& x)
{return 0.918938533204673 + (x-0.5)*log(x) - x+ 0.5*x*log(x*sinh(1/x) + 1/(810.0*pow(x, 6.0)));
}/*** Computes the natural logarithm of the absolute value of* the gamma function of x using the Lanczos approximation.* See http://www.rskey.org/gamma.htm*/
inline double log_gamma_lanczos(const double& x)
{static double q[7] = { 75122.6331530, 80916.6278952, 36308.2951477,8687.24529705, 1168.92649479, 83.8676043424,2.50662827511 };double a = (x + 0.5) * log(x + 5.5) - (x + 5.5);double b = 0;for(int n = 0; n < 7; ++n){a -= log(x + double(n));b += q[n] * pow(x, double(n));}return a + log(b);
}
///namespace cv {class myLineSegmentDetectorImpl CV_FINAL : public LineSegmentDetector{public:/*** Create a LineSegmentDetectorImpl object. Specifying scale, number of subdivisions for the image, should the lines be refined and other constants as follows:** @param _refine How should the lines found be refined?* LSD_REFINE_NONE - No refinement applied.* LSD_REFINE_STD - Standard refinement is applied. E.g. breaking arches into smaller line approximations.* LSD_REFINE_ADV - Advanced refinement. Number of false alarms is calculated,* lines are refined through increase of precision, decrement in size, etc.* @param _scale The scale of the image that will be used to find the lines. Range (0..1].* @param _sigma_scale Sigma for Gaussian filter is computed as sigma = _sigma_scale/_scale.* @param _quant Bound to the quantization error on the gradient norm.* @param _ang_th Gradient angle tolerance in degrees.* @param _log_eps Detection threshold: -log10(NFA) > _log_eps* @param _density_th Minimal density of aligned region points in rectangle.* @param _n_bins Number of bins in pseudo-ordering of gradient modulus.*/myLineSegmentDetectorImpl(int _refine = LSD_REFINE_STD, double _scale = 0.8,double _sigma_scale = 0.6, double _quant = 2.0, double _ang_th = 22.5,double _log_eps = 0, double _density_th = 0.7, int _n_bins = 1024);/*** Detect lines in the input image.** @param _image A grayscale(CV_8UC1) input image.* If only a roi needs to be selected, use* lsd_ptr->detect(image(roi), ..., lines);* lines += Scalar(roi.x, roi.y, roi.x, roi.y);* @param _lines Return: A vector of Vec4i or Vec4f elements specifying the beginning and ending point of a line.* Where Vec4i/Vec4f is (x1, y1, x2, y2), point 1 is the start, point 2 - end.* Returned lines are strictly oriented depending on the gradient.* @param width Return: Vector of widths of the regions, where the lines are found. E.g. Width of line.* @param prec Return: Vector of precisions with which the lines are found.* @param nfa Return: Vector containing number of false alarms in the line region, with precision of 10%.* The bigger the value, logarithmically better the detection.* * -1 corresponds to 10 mean false alarms* * 0 corresponds to 1 mean false alarm* * 1 corresponds to 0.1 mean false alarms* This vector will be calculated _only_ when the objects type is REFINE_ADV*/void detect(InputArray _image, OutputArray _lines,OutputArray width = noArray(), OutputArray prec = noArray(),OutputArray nfa = noArray()) CV_OVERRIDE;/*** Draw lines on the given canvas.** @param image The image, where lines will be drawn.* Should have the size of the image, where the lines were found* @param lines The lines that need to be drawn*/void drawSegments(InputOutputArray _image, InputArray lines) CV_OVERRIDE;/*** Draw both vectors on the image canvas. Uses blue for lines 1 and red for lines 2.** @param size The size of the image, where lines1 and lines2 were found.* @param lines1 The first lines that need to be drawn. Color - Blue.* @param lines2 The second lines that need to be drawn. Color - Red.* @param image An optional image, where lines will be drawn.* Should have the size of the image, where the lines were found* @return The number of mismatching pixels between lines1 and lines2.*/int compareSegments(const Size& size, InputArray lines1, InputArray lines2, InputOutputArray _image = noArray()) CV_OVERRIDE;private:Mat image;Mat scaled_image;Mat_<double> angles; // in radsMat_<double> modgrad;Mat_<uchar> used;int img_width;int img_height;double LOG_NT;bool w_needed;bool p_needed;bool n_needed;const double SCALE;const int doRefine;const double SIGMA_SCALE;const double QUANT;const double ANG_TH;const double LOG_EPS;const double DENSITY_TH;const int N_BINS;struct RegionPoint {int x;int y;uchar* used;double angle;double modgrad;};struct normPoint{Point2i p;int norm;};std::vector<normPoint> ordered_points;struct rect{double x1, y1, x2, y2; // first and second point of the line segmentdouble width; // rectangle widthdouble x, y; // center of the rectangledouble theta; // angledouble dx,dy; // (dx,dy) is vector oriented as the line segmentdouble prec; // tolerance angledouble p; // probability of a point with angle within 'prec'};myLineSegmentDetectorImpl& operator= (const myLineSegmentDetectorImpl&); // to quiet MSVC/*** Detect lines in the whole input image.** @param lines Return: A vector of Vec4f elements specifying the beginning and ending point of a line.* Where Vec4f is (x1, y1, x2, y2), point 1 is the start, point 2 - end.* Returned lines are strictly oriented depending on the gradient.* @param widths Return: Vector of widths of the regions, where the lines are found. E.g. Width of line.* @param precisions Return: Vector of precisions with which the lines are found.* @param nfas Return: Vector containing number of false alarms in the line region, with precision of 10%.* The bigger the value, logarithmically better the detection.* * -1 corresponds to 10 mean false alarms* * 0 corresponds to 1 mean false alarm* * 1 corresponds to 0.1 mean false alarms*/void flsd(std::vector<Vec4f>& lines,std::vector<double>& widths, std::vector<double>& precisions,std::vector<double>& nfas);/*** Finds the angles and the gradients of the image. Generates a list of pseudo ordered points.** @param threshold The minimum value of the angle that is considered defined, otherwise NOTDEF* @param n_bins The number of bins with which gradients are ordered by, using bucket sort.* @param ordered_points Return: Vector of coordinate points that are pseudo ordered by magnitude.* Pixels would be ordered by norm value, up to a precision given by max_grad/n_bins.*/void ll_angle(const double& threshold, const unsigned int& n_bins);/*** Grow a region starting from point s with a defined precision,* returning the containing points size and the angle of the gradients.** @param s Starting point for the region.* @param reg Return: Vector of points, that are part of the region* @param reg_angle Return: The mean angle of the region.* @param prec The precision by which each region angle should be aligned to the mean.*/void region_grow(const Point2i& s, std::vector<RegionPoint>& reg,double& reg_angle, const double& prec);/*** Finds the bounding rotated rectangle of a region.** @param reg The region of points, from which the rectangle to be constructed from.* @param reg_angle The mean angle of the region.* @param prec The precision by which points were found.* @param p Probability of a point with angle within 'prec'.* @param rec Return: The generated rectangle.*/void region2rect(const std::vector<RegionPoint>& reg, const double reg_angle,const double prec, const double p, rect& rec) const;/*** Compute region's angle as the principal inertia axis of the region.* @return Regions angle.*/double get_theta(const std::vector<RegionPoint>& reg, const double& x,const double& y, const double& reg_angle, const double& prec) const;/*** An estimation of the angle tolerance is performed by the standard deviation of the angle at points* near the region's starting point. Then, a new region is grown starting from the same point, but using the* estimated angle tolerance. If this fails to produce a rectangle with the right density of region points,* 'reduce_region_radius' is called to try to satisfy this condition.*/bool refine(std::vector<RegionPoint>& reg, double reg_angle,const double prec, double p, rect& rec, const double& density_th);/*** Reduce the region size, by elimination the points far from the starting point, until that leads to* rectangle with the right density of region points or to discard the region if too small.*/bool reduce_region_radius(std::vector<RegionPoint>& reg, double reg_angle,const double prec, double p, rect& rec, double density, const double& density_th);/*** Try some rectangles variations to improve NFA value. Only if the rectangle is not meaningful (i.e., log_nfa <= log_eps).* @return The new NFA value.*/double rect_improve(rect& rec) const;/*** Calculates the number of correctly aligned points within the rectangle.* @return The new NFA value.*/double rect_nfa(const rect& rec) const;/*** Computes the NFA values based on the total number of points, points that agree.* n, k, p are the binomial parameters.* @return The new NFA value.*/double nfa(const int& n, const int& k, const double& p) const;/*** Is the point at place 'address' aligned to angle theta, up to precision 'prec'?* @return Whether the point is aligned.*/bool isAligned(int x, int y, const double& theta, const double& prec) const;public:// Compare normstatic inline bool compare_norm( const normPoint& n1, const normPoint& n2 ){return (n1.norm > n2.norm);}};CV_EXPORTS Ptr<LineSegmentDetector> createLineSegmentDetector(int _refine, double _scale, double _sigma_scale, double _quant, double _ang_th,double _log_eps, double _density_th, int _n_bins){return makePtr<myLineSegmentDetectorImpl>(_refine, _scale, _sigma_scale, _quant, _ang_th,_log_eps, _density_th, _n_bins);}myLineSegmentDetectorImpl::myLineSegmentDetectorImpl(int _refine, double _scale, double _sigma_scale, double _quant,double _ang_th, double _log_eps, double _density_th, int _n_bins):img_width(0), img_height(0), LOG_NT(0), w_needed(false), p_needed(false), n_needed(false),SCALE(_scale), doRefine(_refine), SIGMA_SCALE(_sigma_scale), QUANT(_quant),ANG_TH(_ang_th), LOG_EPS(_log_eps), DENSITY_TH(_density_th), N_BINS(_n_bins){CV_Assert(_scale > 0 && _sigma_scale > 0 && _quant >= 0 &&_ang_th > 0 && _ang_th < 180 && _density_th >= 0 && _density_th < 1 &&_n_bins > 0);
// CV_UNUSED(_refine); CV_UNUSED(_log_eps);
// CV_Error(Error::StsNotImplemented, "Implementation has been removed due original code license issues");}void myLineSegmentDetectorImpl::detect(InputArray _image, OutputArray _lines,OutputArray _width, OutputArray _prec, OutputArray _nfa){// CV_INSTRUMENT_REGION();image = _image.getMat();CV_Assert(!image.empty() && image.type() == CV_8UC1);std::vector<Vec4f> lines;std::vector<double> w, p, n;w_needed = _width.needed();p_needed = _prec.needed();if (doRefine < LSD_REFINE_ADV)n_needed = false;elsen_needed = _nfa.needed();flsd(lines, w, p, n);Mat(lines).copyTo(_lines);if(w_needed) Mat(w).copyTo(_width);if(p_needed) Mat(p).copyTo(_prec);if(n_needed) Mat(n).copyTo(_nfa);// Clear used structuresordered_points.clear();// CV_UNUSED(_image); CV_UNUSED(_lines);
// CV_UNUSED(_width); CV_UNUSED(_prec); CV_UNUSED(_nfa);
// CV_Error(Error::StsNotImplemented, "Implementation has been removed due original code license issues");}void myLineSegmentDetectorImpl::drawSegments(InputOutputArray _image, InputArray lines){// CV_INSTRUMENT_REGION();CV_Assert(!_image.empty() && (_image.channels() == 1 || _image.channels() == 3));if (_image.channels() == 1){cvtColor(_image, _image, COLOR_GRAY2BGR);}Mat _lines = lines.getMat();const int N = _lines.checkVector(4);CV_Assert(_lines.depth() == CV_32F || _lines.depth() == CV_32S);// Draw segmentsif (_lines.depth() == CV_32F){for (int i = 0; i < N; ++i){const Vec4f& v = _lines.at<Vec4f>(i);const Point2f b(v[0], v[1]);const Point2f e(v[2], v[3]);line(_image, b, e, Scalar(0, 0, 255), 1);}}else{for (int i = 0; i < N; ++i){const Vec4i& v = _lines.at<Vec4i>(i);const Point2i b(v[0], v[1]);const Point2i e(v[2], v[3]);line(_image, b, e, Scalar(0, 0, 255), 1);}}}int myLineSegmentDetectorImpl::compareSegments(const Size& size, InputArray lines1, InputArray lines2, InputOutputArray _image){// CV_INSTRUMENT_REGION();Size sz = size;if (_image.needed() && _image.size() != size) sz = _image.size();CV_Assert(!sz.empty());Mat_<uchar> I1 = Mat_<uchar>::zeros(sz);Mat_<uchar> I2 = Mat_<uchar>::zeros(sz);Mat _lines1 = lines1.getMat();Mat _lines2 = lines2.getMat();const int N1 = _lines1.checkVector(4);const int N2 = _lines2.checkVector(4);CV_Assert(_lines1.depth() == CV_32F || _lines1.depth() == CV_32S);CV_Assert(_lines2.depth() == CV_32F || _lines2.depth() == CV_32S);if (_lines1.depth() == CV_32S)_lines1.convertTo(_lines1, CV_32F);if (_lines2.depth() == CV_32S)_lines2.convertTo(_lines2, CV_32F);// Draw segmentsfor(int i = 0; i < N1; ++i){const Point2f b(_lines1.at<Vec4f>(i)[0], _lines1.at<Vec4f>(i)[1]);const Point2f e(_lines1.at<Vec4f>(i)[2], _lines1.at<Vec4f>(i)[3]);line(I1, b, e, Scalar::all(255), 1);}for(int i = 0; i < N2; ++i){const Point2f b(_lines2.at<Vec4f>(i)[0], _lines2.at<Vec4f>(i)[1]);const Point2f e(_lines2.at<Vec4f>(i)[2], _lines2.at<Vec4f>(i)[3]);line(I2, b, e, Scalar::all(255), 1);}// Count the pixels that don't agreeMat Ixor;bitwise_xor(I1, I2, Ixor);int N = countNonZero(Ixor);if (_image.needed()){CV_Assert(_image.channels() == 3);Mat img = _image.getMatRef();CV_Assert(img.isContinuous() && I1.isContinuous() && I2.isContinuous());for (unsigned int i = 0; i < I1.total(); ++i){uchar i1 = I1.ptr()[i];uchar i2 = I2.ptr()[i];if (i1 || i2){unsigned int base_idx = i * 3;if (i1) img.ptr()[base_idx] = 255;else img.ptr()[base_idx] = 0;img.ptr()[base_idx + 1] = 0;if (i2) img.ptr()[base_idx + 2] = 255;else img.ptr()[base_idx + 2] = 0;}}}return N;}void myLineSegmentDetectorImpl::flsd(std::vector<Vec4f>& lines,std::vector<double>& widths, std::vector<double>& precisions,std::vector<double>& nfas){// Angle toleranceconst double prec = CV_PI * ANG_TH / 180;const double p = ANG_TH / 180;const double rho = QUANT / sin(prec); // gradient magnitude thresholdif(SCALE != 1){Mat gaussian_img;const double sigma = (SCALE < 1)?(SIGMA_SCALE / SCALE):(SIGMA_SCALE);const double sprec = 3;const unsigned int h = (unsigned int)(ceil(sigma * sqrt(2 * sprec * log(10.0))));Size ksize(1 + 2 * h, 1 + 2 * h); // kernel sizeGaussianBlur(image, gaussian_img, ksize, sigma);// Scale image to needed sizeresize(gaussian_img, scaled_image, Size(), SCALE, SCALE, INTER_LINEAR_EXACT);ll_angle(rho, N_BINS);}else{scaled_image = image;ll_angle(rho, N_BINS);}LOG_NT = 5 * (log10(double(img_width)) + log10(double(img_height))) / 2 + log10(11.0);const size_t min_reg_size = size_t(-LOG_NT/log10(p)); // minimal number of points in region that can give a meaningful event// // Initialize region only when needed// Mat region = Mat::zeros(scaled_image.size(), CV_8UC1);used = Mat_<uchar>::zeros(scaled_image.size()); // zeros = NOTUSEDstd::vector<RegionPoint> reg;// Search for line segmentsfor(size_t i = 0, points_size = ordered_points.size(); i < points_size; ++i){const Point2i& point = ordered_points[i].p;if((used.at<uchar>(point) == NOTUSED) && (angles.at<double>(point) != NOTDEF)){double reg_angle;region_grow(ordered_points[i].p, reg, reg_angle, prec);// Ignore small regionsif(reg.size() < min_reg_size) { continue; }// Construct rectangular approximation for the regionrect rec;region2rect(reg, reg_angle, prec, p, rec);double log_nfa = -1;if(doRefine > LSD_REFINE_NONE){// At least REFINE_STANDARD lvl.if(!refine(reg, reg_angle, prec, p, rec, DENSITY_TH)) { continue; }if(doRefine >= LSD_REFINE_ADV){// Compute NFAlog_nfa = rect_improve(rec);if(log_nfa <= LOG_EPS) { continue; }}}// Found new line// Add the offsetrec.x1 += 0.5; rec.y1 += 0.5;rec.x2 += 0.5; rec.y2 += 0.5;// scale the result values if a sub-sampling was performedif(SCALE != 1){rec.x1 /= SCALE; rec.y1 /= SCALE;rec.x2 /= SCALE; rec.y2 /= SCALE;rec.width /= SCALE;}//Store the relevant datalines.push_back(Vec4f(float(rec.x1), float(rec.y1), float(rec.x2), float(rec.y2)));if(w_needed) widths.push_back(rec.width);if(p_needed) precisions.push_back(rec.p);if(n_needed && doRefine >= LSD_REFINE_ADV) nfas.push_back(log_nfa);}}}void myLineSegmentDetectorImpl::ll_angle(const double& threshold,const unsigned int& n_bins){//Initialize dataangles = Mat_<double>(scaled_image.size());modgrad = Mat_<double>(scaled_image.size());img_width = scaled_image.cols;img_height = scaled_image.rows;// Undefined the down and right boundariesangles.row(img_height - 1).setTo(NOTDEF);angles.col(img_width - 1).setTo(NOTDEF);// Computing gradient for remaining pixelsdouble max_grad = -1;for(int y = 0; y < img_height - 1; ++y){const uchar* scaled_image_row = scaled_image.ptr<uchar>(y);const uchar* next_scaled_image_row = scaled_image.ptr<uchar>(y+1);double* angles_row = angles.ptr<double>(y);double* modgrad_row = modgrad.ptr<double>(y);for(int x = 0; x < img_width-1; ++x){int DA = next_scaled_image_row[x + 1] - scaled_image_row[x];int BC = scaled_image_row[x + 1] - next_scaled_image_row[x];int gx = DA + BC; // gradient x componentint gy = DA - BC; // gradient y componentdouble norm = std::sqrt((gx * gx + gy * gy) / 4.0); // gradient normmodgrad_row[x] = norm; // store gradientif (norm <= threshold) // norm too small, gradient no defined{angles_row[x] = NOTDEF;}else{angles_row[x] = fastAtan2(float(gx), float(-gy)) * DEG_TO_RADS; // gradient angle computationif (norm > max_grad) { max_grad = norm; }}}}// Compute histogram of gradient valuesdouble bin_coef = (max_grad > 0) ? double(n_bins - 1) / max_grad : 0; // If all image is smooth, max_grad <= 0for(int y = 0; y < img_height - 1; ++y){const double* modgrad_row = modgrad.ptr<double>(y);for(int x = 0; x < img_width - 1; ++x){normPoint _point;int i = int(modgrad_row[x] * bin_coef);_point.p = Point(x, y);_point.norm = i;ordered_points.push_back(_point);}}// Sortstd::sort(ordered_points.begin(), ordered_points.end(), compare_norm);}void myLineSegmentDetectorImpl::region_grow(const Point2i& s, std::vector<RegionPoint>& reg,double& reg_angle, const double& prec){reg.clear();// Point to this regionRegionPoint seed;seed.x = s.x;seed.y = s.y;seed.used = &used.at<uchar>(s);reg_angle = angles.at<double>(s);seed.angle = reg_angle;seed.modgrad = modgrad.at<double>(s);reg.push_back(seed);float sumdx = float(std::cos(reg_angle));float sumdy = float(std::sin(reg_angle));*seed.used = USED;//Try neighboring regionsfor (size_t i = 0;i<reg.size();i++){const RegionPoint& rpoint = reg[i];int xx_min = std::max(rpoint.x - 1, 0), xx_max = std::min(rpoint.x + 1, img_width - 1);int yy_min = std::max(rpoint.y - 1, 0), yy_max = std::min(rpoint.y + 1, img_height - 1);for(int yy = yy_min; yy <= yy_max; ++yy){uchar* used_row = used.ptr<uchar>(yy);const double* angles_row = angles.ptr<double>(yy);const double* modgrad_row = modgrad.ptr<double>(yy);for(int xx = xx_min; xx <= xx_max; ++xx){uchar& is_used = used_row[xx];if(is_used != USED &&(isAligned(xx, yy, reg_angle, prec))){const double& angle = angles_row[xx];// Add pointis_used = USED;RegionPoint region_point;region_point.x = xx;region_point.y = yy;region_point.used = &is_used;region_point.modgrad = modgrad_row[xx];region_point.angle = angle;reg.push_back(region_point);// Update region's anglesumdx += cos(float(angle));sumdy += sin(float(angle));// reg_angle is used in the isAligned, so it needs to be updates?reg_angle = fastAtan2(sumdy, sumdx) * DEG_TO_RADS;}}}}}void myLineSegmentDetectorImpl::region2rect(const std::vector<RegionPoint>& reg,const double reg_angle, const double prec, const double p, rect& rec) const{double x = 0, y = 0, sum = 0;for(size_t i = 0; i < reg.size(); ++i){const RegionPoint& pnt = reg[i];const double& weight = pnt.modgrad;x += double(pnt.x) * weight;y += double(pnt.y) * weight;sum += weight;}// Weighted sum must differ from 0CV_Assert(sum > 0);x /= sum;y /= sum;double theta = get_theta(reg, x, y, reg_angle, prec);// Find length and widthdouble dx = cos(theta);double dy = sin(theta);double l_min = 0, l_max = 0, w_min = 0, w_max = 0;for(size_t i = 0; i < reg.size(); ++i){double regdx = double(reg[i].x) - x;double regdy = double(reg[i].y) - y;double l = regdx * dx + regdy * dy;double w = -regdx * dy + regdy * dx;if(l > l_max) l_max = l;else if(l < l_min) l_min = l;if(w > w_max) w_max = w;else if(w < w_min) w_min = w;}// Store valuesrec.x1 = x + l_min * dx;rec.y1 = y + l_min * dy;rec.x2 = x + l_max * dx;rec.y2 = y + l_max * dy;rec.width = w_max - w_min;rec.x = x;rec.y = y;rec.theta = theta;rec.dx = dx;rec.dy = dy;rec.prec = prec;rec.p = p;// Min width of 1 pixelif(rec.width < 1.0) rec.width = 1.0;}double myLineSegmentDetectorImpl::get_theta(const std::vector<RegionPoint>& reg, const double& x,const double& y, const double& reg_angle, const double& prec) const{double Ixx = 0.0;double Iyy = 0.0;double Ixy = 0.0;// Compute inertia matrixfor(size_t i = 0; i < reg.size(); ++i){const double& regx = reg[i].x;const double& regy = reg[i].y;const double& weight = reg[i].modgrad;double dx = regx - x;double dy = regy - y;Ixx += dy * dy * weight;Iyy += dx * dx * weight;Ixy -= dx * dy * weight;}// Check if inertia matrix is nullCV_Assert(!(double_equal(Ixx, 0) && double_equal(Iyy, 0) && double_equal(Ixy, 0)));// Compute smallest eigenvaluedouble lambda = 0.5 * (Ixx + Iyy - sqrt((Ixx - Iyy) * (Ixx - Iyy) + 4.0 * Ixy * Ixy));// Compute angledouble theta = (fabs(Ixx)>fabs(Iyy))?double(fastAtan2(float(lambda - Ixx), float(Ixy))):double(fastAtan2(float(Ixy), float(lambda - Iyy))); // in degstheta *= DEG_TO_RADS;// Correct angle by 180 deg if necessaryif(angle_diff(theta, reg_angle) > prec) { theta += CV_PI; }return theta;}bool myLineSegmentDetectorImpl::refine(std::vector<RegionPoint>& reg, double reg_angle,const double prec, double p, rect& rec, const double& density_th){double density = double(reg.size()) / (dist(rec.x1, rec.y1, rec.x2, rec.y2) * rec.width);if (density >= density_th) { return true; }// Try to reduce angle tolerancedouble xc = double(reg[0].x);double yc = double(reg[0].y);const double& ang_c = reg[0].angle;double sum = 0, s_sum = 0;int n = 0;for (size_t i = 0; i < reg.size(); ++i){*(reg[i].used) = NOTUSED;if (dist(xc, yc, reg[i].x, reg[i].y) < rec.width){const double& angle = reg[i].angle;double ang_d = angle_diff_signed(angle, ang_c);sum += ang_d;s_sum += ang_d * ang_d;++n;}}CV_Assert(n > 0);double mean_angle = sum / double(n);// 2 * standard deviationdouble tau = 2.0 * sqrt((s_sum - 2.0 * mean_angle * sum) / double(n) + mean_angle * mean_angle);// Try new regionregion_grow(Point(reg[0].x, reg[0].y), reg, reg_angle, tau);if (reg.size() < 2) { return false; }region2rect(reg, reg_angle, prec, p, rec);density = double(reg.size()) / (dist(rec.x1, rec.y1, rec.x2, rec.y2) * rec.width);if (density < density_th){return reduce_region_radius(reg, reg_angle, prec, p, rec, density, density_th);}else{return true;}}bool myLineSegmentDetectorImpl::reduce_region_radius(std::vector<RegionPoint>& reg, double reg_angle,const double prec, double p, rect& rec, double density, const double& density_th){// Compute region's radiusdouble xc = double(reg[0].x);double yc = double(reg[0].y);double radSq1 = distSq(xc, yc, rec.x1, rec.y1);double radSq2 = distSq(xc, yc, rec.x2, rec.y2);double radSq = radSq1 > radSq2 ? radSq1 : radSq2;while(density < density_th){radSq *= 0.75*0.75; // Reduce region's radius to 75% of its value// Remove points from the region and update 'used' mapfor (size_t i = 0; i < reg.size(); ++i){if(distSq(xc, yc, double(reg[i].x), double(reg[i].y)) > radSq){// Remove point from the region*(reg[i].used) = NOTUSED;std::swap(reg[i], reg[reg.size() - 1]);reg.pop_back();--i; // To avoid skipping one point}}if(reg.size() < 2) { return false; }// Re-compute rectangleregion2rect(reg ,reg_angle, prec, p, rec);// Re-compute region points densitydensity = double(reg.size()) /(dist(rec.x1, rec.y1, rec.x2, rec.y2) * rec.width);}return true;}double myLineSegmentDetectorImpl::rect_improve(rect& rec) const{double delta = 0.5;double delta_2 = delta / 2.0;double log_nfa = rect_nfa(rec);if(log_nfa > LOG_EPS) return log_nfa; // Good rectangle// Try to improve// Finer precisionrect r = rect(rec); // Copyfor(int n = 0; n < 5; ++n){r.p /= 2;r.prec = r.p * CV_PI;double log_nfa_new = rect_nfa(r);if(log_nfa_new > log_nfa){log_nfa = log_nfa_new;rec = rect(r);}}if(log_nfa > LOG_EPS) return log_nfa;// Try to reduce widthr = rect(rec);for(unsigned int n = 0; n < 5; ++n){if((r.width - delta) >= 0.5){r.width -= delta;double log_nfa_new = rect_nfa(r);if(log_nfa_new > log_nfa){rec = rect(r);log_nfa = log_nfa_new;}}}if(log_nfa > LOG_EPS) return log_nfa;// Try to reduce one side of rectangler = rect(rec);for(unsigned int n = 0; n < 5; ++n){if((r.width - delta) >= 0.5){r.x1 += -r.dy * delta_2;r.y1 += r.dx * delta_2;r.x2 += -r.dy * delta_2;r.y2 += r.dx * delta_2;r.width -= delta;double log_nfa_new = rect_nfa(r);if(log_nfa_new > log_nfa){rec = rect(r);log_nfa = log_nfa_new;}}}if(log_nfa > LOG_EPS) return log_nfa;// Try to reduce other side of rectangler = rect(rec);for(unsigned int n = 0; n < 5; ++n){if((r.width - delta) >= 0.5){r.x1 -= -r.dy * delta_2;r.y1 -= r.dx * delta_2;r.x2 -= -r.dy * delta_2;r.y2 -= r.dx * delta_2;r.width -= delta;double log_nfa_new = rect_nfa(r);if(log_nfa_new > log_nfa){rec = rect(r);log_nfa = log_nfa_new;}}}if(log_nfa > LOG_EPS) return log_nfa;// Try finer precisionr = rect(rec);for(unsigned int n = 0; n < 5; ++n){if((r.width - delta) >= 0.5){r.p /= 2;r.prec = r.p * CV_PI;double log_nfa_new = rect_nfa(r);if(log_nfa_new > log_nfa){rec = rect(r);log_nfa = log_nfa_new;}}}return log_nfa;}double myLineSegmentDetectorImpl::rect_nfa(const rect& rec) const{int total_pts = 0, alg_pts = 0;double half_width = rec.width / 2.0;double dyhw = rec.dy * half_width;double dxhw = rec.dx * half_width;edge ordered_x[4];edge* min_y = &ordered_x[0];edge* max_y = &ordered_x[0]; // Will be used for loop rangeordered_x[0].p.x = int(rec.x1 - dyhw); ordered_x[0].p.y = int(rec.y1 + dxhw); ordered_x[0].taken = false;ordered_x[1].p.x = int(rec.x2 - dyhw); ordered_x[1].p.y = int(rec.y2 + dxhw); ordered_x[1].taken = false;ordered_x[2].p.x = int(rec.x2 + dyhw); ordered_x[2].p.y = int(rec.y2 - dxhw); ordered_x[2].taken = false;ordered_x[3].p.x = int(rec.x1 + dyhw); ordered_x[3].p.y = int(rec.y1 - dxhw); ordered_x[3].taken = false;std::sort(ordered_x, ordered_x + 4, AsmallerB_XoverY);// Find min y. And mark as taken. find max y.for(unsigned int i = 1; i < 4; ++i){if(min_y->p.y > ordered_x[i].p.y) {min_y = &ordered_x[i]; }if(max_y->p.y < ordered_x[i].p.y) {max_y = &ordered_x[i]; }}min_y->taken = true;// Find leftmost untaken point;edge* leftmost = 0;for(unsigned int i = 0; i < 4; ++i){if(!ordered_x[i].taken){if(!leftmost) // if uninitialized{leftmost = &ordered_x[i];}else if (leftmost->p.x > ordered_x[i].p.x){leftmost = &ordered_x[i];}}}CV_Assert(leftmost != NULL);leftmost->taken = true;// Find rightmost untaken point;edge* rightmost = 0;for(unsigned int i = 0; i < 4; ++i){if(!ordered_x[i].taken){if(!rightmost) // if uninitialized{rightmost = &ordered_x[i];}else if (rightmost->p.x < ordered_x[i].p.x){rightmost = &ordered_x[i];}}}CV_Assert(rightmost != NULL);rightmost->taken = true;// Find last untaken point;edge* tailp = 0;for(unsigned int i = 0; i < 4; ++i){if(!ordered_x[i].taken){if(!tailp) // if uninitialized{tailp = &ordered_x[i];}else if (tailp->p.x > ordered_x[i].p.x){tailp = &ordered_x[i];}}}CV_Assert(tailp != NULL);tailp->taken = true;double flstep = (min_y->p.y != leftmost->p.y) ?(min_y->p.x - leftmost->p.x) / (min_y->p.y - leftmost->p.y) : 0; //first left stepdouble slstep = (leftmost->p.y != tailp->p.x) ?(leftmost->p.x - tailp->p.x) / (leftmost->p.y - tailp->p.x) : 0; //second left stepdouble frstep = (min_y->p.y != rightmost->p.y) ?(min_y->p.x - rightmost->p.x) / (min_y->p.y - rightmost->p.y) : 0; //first right stepdouble srstep = (rightmost->p.y != tailp->p.x) ?(rightmost->p.x - tailp->p.x) / (rightmost->p.y - tailp->p.x) : 0; //second right stepdouble lstep = flstep, rstep = frstep;double left_x = min_y->p.x, right_x = min_y->p.x;// Loop around all points in the region and count those that are aligned.int min_iter = min_y->p.y;int max_iter = max_y->p.y;for(int y = min_iter; y <= max_iter; ++y){if (y < 0 || y >= img_height) continue;for(int x = int(left_x); x <= int(right_x); ++x){if (x < 0 || x >= img_width) continue;++total_pts;if(isAligned(x, y, rec.theta, rec.prec)){++alg_pts;}}if(y >= leftmost->p.y) { lstep = slstep; }if(y >= rightmost->p.y) { rstep = srstep; }left_x += lstep;right_x += rstep;}return nfa(total_pts, alg_pts, rec.p);}double myLineSegmentDetectorImpl::nfa(const int& n, const int& k, const double& p) const{// Trivial casesif(n == 0 || k == 0) { return -LOG_NT; }if(n == k) { return -LOG_NT - double(n) * log10(p); }double p_term = p / (1 - p);double log1term = (double(n) + 1) - log_gamma(double(k) + 1)- log_gamma(double(n-k) + 1)+ double(k) * log(p) + double(n-k) * log(1.0 - p);double term = exp(log1term);if(double_equal(term, 0)){if(k > n * p) return -log1term / M_LN10 - LOG_NT;else return -LOG_NT;}// Compute more terms if neededdouble bin_tail = term;double tolerance = 0.1; // an error of 10% in the result is acceptedfor(int i = k + 1; i <= n; ++i){double bin_term = double(n - i + 1) / double(i);double mult_term = bin_term * p_term;term *= mult_term;bin_tail += term;if(bin_term < 1){double err = term * ((1 - pow(mult_term, double(n-i+1))) / (1 - mult_term) - 1);if(err < tolerance * fabs(-log10(bin_tail) - LOG_NT) * bin_tail) break;}}return -log10(bin_tail) - LOG_NT;}inline bool myLineSegmentDetectorImpl::isAligned(int x, int y, const double& theta, const double& prec) const{if(x < 0 || y < 0 || x >= angles.cols || y >= angles.rows) { return false; }const double& a = angles.at<double>(y, x);if(a == NOTDEF) { return false; }// It is assumed that 'theta' and 'a' are in the range [-pi,pi]double n_theta = theta - a;if(n_theta < 0) { n_theta = -n_theta; }if(n_theta > M_3_2_PI){n_theta -= M_2__PI;if(n_theta < 0) n_theta = -n_theta;}return n_theta <= prec;}} // namespace cv
在LSDDetector_custom.cpp文件中引用头文件
#include "my_lsd.hpp"再次编译运行
四、运行
source devel/setup.bash
roslaunch plvins_estimator plvins_show_linepoint.launch
rosbag play MH_05_difficult.bag注意：我们此时需要需要将src/PL-VINSvins_estimator/launch/下的plvins-show-linepoint.launch改为plvins_show_linepoint.launch（注意是下划线）文件名
数据集下载地址如下：
kmavvisualinertialdatasets – ASL Datasets
博主github仓库中的代码全部都修改完毕，适用ubuntu20.04 opencv4
