Nuitrack Support for Raspberry Pi 3 and Tinkerboard Added

• Gbenel

  • February 09, 2020 13:18

  Yes, but as I learned the hard way, nuitrack does not work with realsense on the raspberry pi (nuitrack does not support)

  0

  Comment actions Permalink
• MartyG

  • February 09, 2020 13:27

  If you are using a Pi that is not Pi 4 and so have a USB 2.0 connection, Nuitrack has advice about what changes to do in order to use Nuitrack with RealSense and USB 2.0.  I have pasted it below.

  **************

  You can connect RealSense via USB 2.0, but in this case you have to set lower resolution for a depth map.  To do this, open nuitrack.config (%NUITRACK_HOME%/data), find the section "Realsense2Module" and add the following lines to "Realsense2Module.Depth":

  "RawWidth": 640,
  "RawHeight": 480,

  0

  Comment actions Permalink
• Gbenel

  • February 09, 2020 13:35

  Ok, I'm using a pi4. It does have a couple USB 2 ports, may be worth trying. I imagine will be very slow though.
  Interesting, when I contacted nuitrack they told me essentially no way to make realsense work with any pi.

  0

  Comment actions Permalink
• Gbenel

  • February 09, 2020 13:49

  I have posenet running very well on my rpi4 with a coral edge device, but without using depth data to find skeletons, it struggles for me when two people are near one another. Was hoping nuitrack may perform better as it uses depth data in establishing identifications.

  0

  Comment actions Permalink
• MartyG

  • February 09, 2020 14:02
  • Edited

  Another skeletal tracking solution for the RealSense D415, RealSense D435 and FRAMOS D435e cameras was highlighted by Intel during January's CES electronics show.  It's called Cubemos, and will also work with the forthcoming new RealSense L515 lidar-based depth camera.

  https://www.cubemos.com/skeleton-tracking-sdk/ 

  Click on the 'Download SDK' button on the above link to go to a download page that has a link to the PDF user manual for Cubemos.

  Here is a video of it running with the L515:

  https://www.youtube.com/watch?v=PTMo4Dvya1Y&feature=emb_logo 

  0

  Comment actions Permalink
• Gbenel

  • February 09, 2020 14:03

  Yes, I have talked with them. Unfortunately they do not support the Jetson platform or raspberry pi platform.

  0

  Comment actions Permalink
• MartyG

  • February 09, 2020 14:25
  • Edited

  I did further research on the subject of finding skeletal tracking solutions for RealSense cameras that might work with a Pi, but the possibilities seem to be slim.  Intel did a seminar a while ago on advanced body tracking that is available on YouTube, though it typically requires multiple cameras to generate a data-set for the tracking before it can be used with a single camera.

  https://www.youtube.com/watch?v=VSHDyUXSNqY

  0

  Comment actions Permalink
• Gbenel

  • February 09, 2020 14:35

  Posenet works great (with an edge device) but when you have multiple people in the frame relatively close together sometimes things run amok and depth data (in finding skeletons) would be useful (it is easy to get depth data after the fact)

  0

  Comment actions Permalink
• MartyG

  • February 09, 2020 15:00

  You can get a basic distance measurement on a person using object recognition with a Deep Neural Network (DNN).  It ought to be able to discern between two different people.  I have seen DNN example programs that can detect more than one person separately.  Here's an example of a program detecting the difference between an object and a human ("Person"):

  https://github.com/twMr7/rscvdnn

  Here is a multi-person example:

  https://github.com/apoorvavinod/Real_time_Object_detection_and_tracking

  0

  Comment actions Permalink
• Gbenel

  • February 09, 2020 22:26

  You can see even in their examples on the readme on github that the same person has their id switched while tracking

  0

  Comment actions Permalink
• MartyG

  • February 09, 2020 23:23

  Another option for skeletal tracking you could investigate is OpenPose

  https://github.com/stevenjj/openpose_ros 

  

  0

  Comment actions Permalink
• Gbenel

  • February 09, 2020 23:41

  Yes, but again, I don’t think openpose uses depth info to find skeletons thus I don’t think it would be any better than posenet (which is optimized for edge devices and works great in 2d)

  0

  Comment actions Permalink
• MartyG

  • February 10, 2020 00:16

  I did further investigation but could not find a solution that would clearly work with Raspberry Pi.  The last resort may be to find out what board will work with existing skeleton tracking solutions and change to that.  I wish you the best of luck.   Sorry I couldn't be of more help.

  0

  Comment actions Permalink
• Gbenel

  • February 10, 2020 01:33

  A Nuc would do it but it has a large power requirement - I am working on a wearable device and it would be hard to carry around such a heavy battery

  0

  Comment actions Permalink
• MartyG

  • February 10, 2020 07:06

  A more compact alternative to a NUC is the Intel Compute Stick pocket-sized PC.

  https://www.intel.co.uk/content/www/uk/en/products/boards-kits/compute-stick.html 

  0

  Comment actions Permalink

Please sign in to leave a comment.