Concerning depth cameras, here's my current survey:
There appears to be a new Kinect in the pipeline, "Project Kinect
for Azure", aka Kinect v4, and its performance
looks amazing. There have
been a couple of surveys on desired features, and body segmentation
has placed highly. You can download a
recent
whitepaper on it, and there is a
notice
mailing list. Have not heard any timeline though, I expect it
is at least a year away.
The Orbbec is a nice camera, basically a better Primesense/XTion1.
It has on OpenNI2 API, so it dropped right into my software. The
Persee has its own processor, so it can do pre-processing and send
data over a network connection. There is an article about doing
body
tracking with it and NUITrack, but I have not tried it.
Another interesting device is the
VicoVR,
which does skeleton tracking and sends it over WiFi, a very compact
solution. I got one off Kickstarter but have never had time to
check it out.
I am using a lot of
StereoLabs
ZEDs in a large interactive art installation in NYC, since
they are stereo vision and work in sunlight. They require a pretty
good nVidia CUDA capable graphics card though, which adds to their
cost and system complexity. But in the right conditions they can
approach a Kinect v2 in depth quality.
If you have not checked the Intel RealSense cams in a while, you
should. The
DS435 is
an excellent, small, USB-powered camera that does the depth
processing internally so it has low system demand. It uses stereo
vision, but in the infrared, and it has its own built-in speckle
illuminator so it can work in darkness to direct sunlight. The
speckle also really improves the depth quality indoors.
But I have to admit, when the Kinect v2 went out of production, I
went and bought a pantload of them at the local GameStop; lots and
lots of gamers that got one with their XBox turned them in, so used
ones are
only
$40 here in the US. With the simple power mod, they do not
need the interface box and can directly plug into USB3. I still use
them a lot - the quality of their pointcloud allows me to combine
multiple cameras with high accuracy. And with LibFreenect2, you can
use them with OpenNI2 applications.
So, did I miss anything new? Sorry can't help with the skeleton
tracking more, but I rolled my own simpler version in order to get
device and SDK independence.
Good luck, share what you find out!
- Lorne
Lorne Covington
NOIRFLUX - Art in Interaction
http://noirflux.com