I have been trying to fit a cylinder model to a generated point cloud using the pcl tools. I have been adapting the example code that is provided in the documentation here.
From my understanding, the pcl::SampleConsensusModelCylinder requires normals data so I have added a new check in the code that checks for a -cf argument. The code then calculates the normals of each point in a pcl::Normals-type point cloud called cloud_normals and then concatenates this with the original pcl::PointXYZ-type point cloud, cloud . I save this to a new pcl::PointNormal-type point cloud called cloud_normalpoints and with this I attempt to fit the the pcl::SampleConsensusModelCylinder using RAndom SAmple Consensus (RANSAC).
I have included the code below:
else if (pcl::console::find_argument (argc, argv, "-cf") >= 0 )
{
//TODO: find fastest way to fit cylinder model to point cloud
// the cloud_normals point cloud will be used to store the point cloud or normals
pcl::PointCloud<pcl::Normal>::Ptr cloud_normals (new pcl::PointCloud<pcl::Normal>);
cloud_normals->width = 500;
cloud_normals->height = 1;
// is_dense is True if no points have NaN or Inf in any of their floating points field
cloud_normals->is_dense = false;
cloud_normals->points.resize (cloud_normals->width * cloud_normals->height);
// the NormalEstimation object ne is created and will estimate the normals and curvature at each point
pcl::NormalEstimation<pcl::PointXYZ, pcl::Normal> ne;
// the search:KdTree object pointer points to a search method for finding points in 3D space (3D point clouds)
pcl::search::KdTree<pcl::PointXYZ>::Ptr tree (new pcl::search::KdTree<pcl::PointXYZ> ());
// use the filtered cloud as an input to the normal estimator
ne.setInputCloud (cloud);
// set number of k-nearest neighbours to use for feature estimation to 50
ne.setKSearch (50);
// compute normals and save these to the clouds_normals point cloud
ne.compute (*cloud_normals);
pcl::PointCloud<pcl::PointNormal>::Ptr cloud_normalpoints (new pcl::PointCloud<pcl::PointNormal>);
cloud_normalpoints->width = 500;
cloud_normalpoints->height = 1;
// is_dense is True if no points have NaN or Inf in any of their floating points field
cloud_normalpoints->is_dense = false;
cloud_normalpoints->points.resize (cloud_normalpoints->width * cloud_normalpoints->height);
pcl::concatenateFields(*cloud,*cloud_normals,*cloud_normalpoints);
//TODO: Solve normals not given error
pcl::SampleConsensusModelCylinder<pcl::PointNormal, pcl::Normal>::Ptr
model_c (new pcl::SampleConsensusModelCylinder<pcl::PointNormal, pcl::Normal> (cloud_normalpoints));
// Declares ransac as a ransac implementation searching for a cylinder (according to model_c -> in cloud)
pcl::RandomSampleConsensus<pcl::PointNormal> ransac (model_c);
// Set distance threshold of .01 -> believe this is for inliers
ransac.setDistanceThreshold (.01);
// Compute model coefficients and find inliers
ransac.computeModel();
// Return indices of best set of inliers so far for this model
ransac.getInliers(inliers);
}
I also added some more code to generate an original point cloud containing a cylinder, but this works so I will not go into it here.
When I run my code it gets to the compute model stage and then throws the following error:
[pcl::SampleConsensusModelCylinder::computeModelCoefficients] No input dataset containing normals was given!
Does anyone know why this is? The cloud_normalpoints cloud includes the normal data that was found for each point. Should I be setting up the RANSAC estimator differently? Should I use a different point type? I am relatively new to pcl so any help would be greatly appreciated!
You have to call the function setInputNormals of your model_c, where you pass the cloud_normals.
The cloud you pass in the constructor of SampleConsensusModelCylinder only sets the XYZ information, but is not used for normals.
This tutorial could also be interesting for you: https://pcl.readthedocs.io/projects/tutorials/en/latest/cylinder_segmentation.html