Skeleton Tracking with the Kinect via ROS

[Nathaniel Lewis]

I got the skeleton tracking package inside ROS to work with the Kinect.
Had to tweak a couple of things to get all of the packages to compile on
Ubuntu 10.10. But here is a video of it working:
http://www.youtube.com/watch?v=juDYwSc-jug

The actual skeleton tracker is NITE by PrimeSense, which is the software
the Xbox 360 is actually running. So enjoy! One thing about running
the demo - you need an Intel CPU. The NITE package being compiled with
the SSSE3 Instuction set (yes 3 S's on the front) which apparently no
AMD cpus implement currently. My Athlon X2 desktop kept throwing
illegal instruction errors so Athlon 64s can't run it. It worked on my
Pentium D laptop though.
Nathaniel

[Pi Robot]

Hi Nathaniel,

This is terrific! Inspired by your results, I picked up a Kinect and
tried it under ROS. While I can get the camera images and point
clouds in RViz, performance is really really slow, about 1 frame per
second for the point cloud update and my CPU load skyrockets to 3.8.
I'm running the latest C Turtle release of ROS( (debian packages) on
an Ubuntu 10.04 machine which is a fairly powerful Fujitsu T730 tablet
2.5 GHz Intel Core i5 with 4 Gb RAM. For comparison, I tried the
connect on my Windows XP machine with RoboRealm and frame updates were
very fast (though there is no point cloud in RoboRealm). Do you think
the sluggishness might be caused by a poor graphics driver for my
Fujitsu? Any other thoughts?

Also, is it possible to get the false color distance camera image in
RViz running just the kinect_node? All I can get is a black-and-white
depth map--just two-toned with no shades of color to represent depth.

Thanks!
patrick



On Dec 12, 11:28 am, Nathaniel Lewis <linux.robotd...@gmail.com>
wrote:

[Pi Robot]

Just an update on my own post. Seems I get much better frame rates if
I set the Point Cloud display type to Points rather than Billboard
Spheres, though I would estimate I'm still getting only 5-10 fps.
This is true whether I use the ROS/Kinect or OpenNI drivers. Also,
when using the OpenNI drivers, I do not get any camera images in
RViz. The status error messages I get are one or all of the below:

No CameraInfo received on [/camera/rgb/camera_info]. Topic may not
exist.
Unsupported image encoding [32FC1]
For frame []: Frame [] does not exist

On the other hand, I was able to run the NITE demo just fine--fast
refresh and very nice tracking of multiple users.

--patrick

[Alex Bravo]

Patrick,

On Dec 15, 6:39 am, Pi Robot <pirobotproducti...@gmail.com> wrote:
> Just an update on my own post.  Seems I get much better frame rates if
> I set the Point Cloud display type to Points rather than Billboard
> Spheres, though I would estimate I'm still getting only 5-10 fps.

I believe they fixed Billboard issue in rviz, but it's only available
in unstable.

>I'm running the latest C Turtle release of ROS( (debian packages) on
>an Ubuntu 10.04 machine which is a fairly powerful Fujitsu T730 tablet
>2.5 GHz Intel Core i5 with 4 Gb RAM.

Did you see Kinect used with quadrotor? They have 1.6GHz Atom board
there.

One of the problems with Kinect is there's no full support of OpenNI's
QVGA 320x240 mode.
PrimeSense's PSDK5.0 sensor does support it though.

So alternatively they can downsample 640x480 to 320x240 in PCL, but
that's a lot of computation.
At 640x480x30Hz it's a lot more data that ROS ever saw (according to
ROS people themselves :) ).
So Radu and team are working hard to profile/optimize PCL.

Here are results of latest profiling:
http://kinect-with-ros.976505.n3.nabble.com/Re-Ros-kinect-OpenNI-dev-Performance-td2066691.html#a2090724

Alex Bravo, RobotNV

[Nathaniel Lewis]

I get the same problem with RViz as well. I originally used points and
I get 10-15 fps. I can think of a few things, maybe it is because it is
pushing the image + depth data over the network. The base libfreenect
driver is fully capable of doing 30 fps (I tested that). RViz gives me
low performance, but alas, I have a GPU not well supported by Ogre3D
(The 3D library RViz uses) in my laptop (Intel GMA 4000), which could be
just as true. If anyone has a decent GPU (possible a medium to high
range Geforce or Quadro NVidia card) could they possible test it for us?
Linux really doesn't like the ATI Radeon HD 5770 1 GB in my desktop,
even with the proprietary drivers.
Nathaniel

[anfederman@comcast]

Did I hear that right? I don't see it on the calendar, but I thought I
heard it mentioned at the November meeting.

Go Twins! (Too many Minnesotans in the club.)

Alan N. Federman, PhD
http://federman.best.vwh.net
http://www.linkedin.com/in/alan1federman
(408)227-6106 Home
(408)826-1795 Mobile

[Dave Curtis]

Yes, that's the plan.
-dave

[Pi Robot]

Thanks Alex and Nathaniel. Glad to hear it isn't just my machine. I
have since gotten some more info from the ros-users and ros-kinect
mailing lists and I think I'll try a voxel grid downsampling on the
Kinect point cloud once I figure out how!

--patrick

[Alan Downing]

FYI, there's now faster lower resolution modes and 32bit NiTE skeletal
tracking support.
http://www.ros.org/news/2010/12/openni-updates.html

Alan

> --
> You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
> To post to this group, send email to hbrob...@googlegroups.com.
> To unsubscribe from this group, send email to hbrobotics+...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/hbrobotics?hl=en.
>
>

[Pi Robot]

Thanks for the heads-up Alan! Looks like some nice improvements.

--patrick

[Pi Robot]

I am trying out the latest OpenNI package with dynamic reconfigure and
setting all image topics to QQVGA makes a big difference to the fps of
the point cloud image in RViz. I'm still using Points as the Style as
even with the unstable version of RViz, Billboards or Billboard
Spheres slows things to a crawl. Nevertheless, the points image is
quite nice with some cool shadowing effects if you stand in front of a
wall.

--patrick

[Camp Peavy]

The "Holiday Junk Exchange" (Dec 22) is mentioned on the homepage and in the calender: http://hbrobotics.org/.   Ho-Ho-Hope to see you there, o<|:^) Camp --- On Wed, 12/15/10, anfederman@comcast <anfed...@comcast.net> wrote:

-- You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group. To post to this group, send email to hbrob...@googlegroups.com.

To unsubscribe from this group, send email to hbrobotics+unsub...@googlegroups.com.