View my account

Objects/people close together

Follow
Mspecter91

Hello all!

I would like to get some help or any ideas I can do to salve this problem.

So I am using d435 to count people, going in and out of entrance.

When two people are close together or touching each other, it recognize as one person rather than two. Such as their shoulders touching, or an adult holding a child's hand.

I tried to use watershed, but it seems like its not working.

I am pretty new to this field so if anyone who knows how to solve this problem any suggestions would be very much appreciated.

Also any suggestions on what would be the best ways for people counting using D435, such as using object detection (Tenserflow, openCV ) etc would also be great!

 

Thank you and looking forward for any suggestions.

 

0

Comments

3 comments

Date Votes
  • MartyG

    Hi Mspecter91  If you are able to make use of pre-made commercial software, the RealSense-compatible Cubemos Skeleton Tracking SDK has a function for outputting the number of detected skeletons.

    https://www.intelrealsense.com/skeleton-tracking/

     

    If you prefer to develop your own application, a way to isolate individual bodies is to position the camera overhead and remove depth detail from the shoulders down so that only the head is rendered on the image.  There is a detailed discussion of the subject in the link below.

    https://github.com/IntelRealSense/librealsense/issues/8044

     

    The RealSense SDK also has a TensorFlow compatibility wrapper that has an example program for detecting persons and determining their height.

    https://github.com/IntelRealSense/librealsense/tree/master/wrappers/tensorflow#part-2---augmenting-output-using-depth-data

     

    0
    Comment actions Permalink
  • Mspecter91

    Thank you for your rely MartyG

     

    I will look into the skeleton detection, looks interesting!

     

    --remove depth detail from the shoulders down so that only the head is rendered on the image.

    --> this one I have tried and in this case, children are not being detected due to their heights.

     

    For an additional note, I am currently trying to make it using  Rasberry Pi 4 model B, if any additional advise what I can do to count people with Raspberry Pi would be extremely helpful.

    Does Cubemos Skeleton Tracking SDK  and TensorFlow work with Raspberry Pi 4? (Or would you recommend using stronger board?)

     

    Thank you, looking forward to hear from you

    0
    Comment actions Permalink
  • MartyX Grover
    • Edited

    I would refer you to the Cubemos website's customer support for queries about the Skeleton Tracking SDK, as Cubemos provide technical support for it.

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

     

    In regard to TensorFlow, I have not seen the RealSense SDK's TensorFlow wrapper used with an Arm device such as Raspberry Pi, so cannot confirm if it is possible.  The link below does though feature an object detection project running on a Pi 3 with TensorFlow and a RealSense D435 camera.

    https://github.com/PINTO0309/MobileNet-SSDLite-RealSense-TF

     

    Regarding alternative computing boards that are suited to AI / vision processing applications and are in a similar price range to Pi 4, you could investigate the Nvidia Jetson Nano.

    0
    Comment actions Permalink

Please sign in to leave a comment.

Didn't find what you were looking for?

New post