jqq wrote:Yes it does work. The problem is certain filters like Euclidean Cluster Extraction freeze on the cloud. If I first pass it through a Voxel Grid though, they work. I'm pretty sure this is because the grid makes a completely new cloud, and the generated cloud has some bad points. As per your grabber, there is a TODO regarding NaN points. Have you tried this method mentioned here:
http://www.pcl-users.org/Conversion-fro ... l#a4023943
i.e. this test: if (depthValue == depthValue) // if depthValue is not NaN
I tried it but still got the filters freezing unless I voxelised the cloud first.
Also, where did you get your constants for the cloud generation:
#define DEPTH_CX 319.5
#define DEPTH_CY 239.5
#define DEPTH_FX 525.0
#define DEPTH_FY 525.0
Are these always valid? Do you know if they are valid for Carmine 1.09? Are they valid for all OpenNI devices? Is there a more robust way? (sorry for all the questions!)
Thanks
You are probably right about the NaN points. Generally I always downsample (even on a desktop) for speed anyways. There is a function in PCL that does this automatically...
pcl::removeNaNFromPointCloud(*cloudIn, *cloudIn, index);
I don't remember why I added the TODO, oh well
The constants CX and CY are the center points of 640 x 480, respectively (639/2 and 479/2 to be exact, because of the 0 number base). FX and FY are the focal length parameters. The carmine is for close range, right? If so, FX and FY may be different.
Those numbers are technically not optimal, but it is what OpenNI uses as the default. You can optimize the parameters using the OpenCV camera calibration or MATLAB calibration toolboxes. CX and CY are where the optical center of the camera/lens assembly passes through the CMOS sensor. In practice, this is never perfectly centered.
An example of this is when you look at edges in a point cloud. Often you will get some colors bleeding across different depths. This is because the calibration parameters are not optimal, and the registration between the color and depth images is somewhat misaligned (or the transformation between the RGB and D cameras is not accurate, which you can solve with stereo camera calibration).
If you want to calibrate the intrinsic parameters of the D camera yourself (CX,CY,FX,FY), you need to get the raw IR image (the IR image itself, not the depth) of a checkerboard with known dimensions from several orientations in an environment flooded with IR light (to override the IR projector, I just use a halogen lamp). You then use a camera calibration toolbox of choice, and it will give you the "more optimal" solution for your particular device. There is a slight deviation between each device due to manufacturing tolerances, but you only need to do it once for your device and it can be performed on any computer.
This link talks about it a little...
http://wiki.ros.org/kinect_calibration/technical