I had a similar problem. It turned out in my case that the two point clouds
were further apart than the max correspondence distance.  This resulted in
no correspondences being found, which resulted in a segfault when the ransac
algorithm tried to choose samples to do the model alignment.

I traced the segfault to the following line in computeModelCoefficients:

    indices_tgt[i] = (*indices_tgt_)[samples[i]];

Adding a check to make sure that returns false when indices_tgt_ has less
than samples[i]+1 elements fixed this for me, but there is probably a better
place to put this check.

-Jon B


On Fri, Dec 10, 2010 at 3:57 PM, paulhebert <paulhebert@gmail.com> wrote:

>
> Hi All,
>
> I can't for the life of me, figure what is going wrong. Perhaps, I do not
> know the structure of PointCloud very well yet but I keep getting a
> segfault
> at the last line in "reg.align". I would like to fill in the point cloud
> with a buffer of floats/doubles and then run ICP on the two cloud points.
> Is
> this the right way to do it? It seems to work if I use PCD files but not
> when I manually input data like below. Hope someone can help.
>
>
>    // Fill in the model cloud data
>    model.width    = modelN;
>    model.height   = 1;
>    model.is_dense = false;
>    model.points.resize (model.width * model.height);
>
>    for (size_t i = 0; i < model.points.size (); i++)
>    {
>        model.points[i].x = (float)modelPointsT[3*i + 0];
>        model.points[i].y = (float)modelPointsT[3*i + 1];
>        model.points[i].z = (float)modelPointsT[3*i + 2];
>    }
>   // Fill in the model cloud data
>    data.width    = dataN;
>    data.height   = 1;
>    data.is_dense = false;
>    data.points.resize (data.width * data.height);
>
>    for (size_t i = 0; i < data.points.size (); i++)
>    {
>        data.points[i].x = (float)dataPointsT[3*i + 0] + 5;
>        data.points[i].y = (float)dataPointsT[3*i + 1];
>        data.points[i].z = (float)dataPointsT[3*i + 2];
>    }
>
>    pcl::PointCloud<pcl::PointXYZ> output;
>
>    pcl::IterativeClosestPointNonLinear<pcl::PointXYZ, pcl::PointXYZ> reg;
>
>    reg.setMaxCorrespondenceDistance (0.1);
>    reg.setTransformationEpsilon (1e-6);
>    reg.setMaximumIterations (50);
>
>    reg.setInputCloud(boost::make_shared<const
> pcl::PointCloud<pcl::PointXYZ> > (data));
>    reg.setInputTarget(boost::make_shared<const
> pcl::PointCloud<pcl::PointXYZ> > (model));
>    Eigen3::Matrix4f transform;
>    //reg.computeTransformation(output);
>    //reg.estimateRigidTransformationLM(data,model,transform);
>    printf("Done\n");
>    reg.align(output);
>
>
> --
> View this message in context:
> http://ros-users.122217.n3.nabble.com/ICP-PCL-with-foreign-data-tp2066643p2066643.html
> Sent from the ROS-Users mailing list archive at Nabble.com.
> _______________________________________________
> ros-users mailing list
> ros-users@code.ros.org
> https://code.ros.org/mailman/listinfo/ros-users
>