• State
  • Locked Locked
  • Replies 2 replies
  • Subscribers 30 subscribers
  • Views 952 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

IWRL6432AOP: Classification and Multi-people detection and tracking

Jack
Mastermind 9145 points
Part Number: IWRL6432AOP
Other Parts Discussed in Thread: IWRL6432

Tool/software:

Hi IWRL6432AOP Champ !

As my customer watched the Smart TV demo video of L6432 application at Radar Toobox and SDK.

My customer would like to start L6432AOP porject with Human and Non-Human Classification and Multi-People Tracking features. 

Could you please clarify the followings?

A customer was impressed with the L6432 promotional video for a Smart TV solution in the Radar Toolbox's Application Guide, and started looking into implementing radar in their set-top box.

The main features they wanted to use were multi-person tracking and Human vs Non-Human Classification for 5 people or less.

  1. According to the following Classification Overview document, it seems to support 3D people counting and tracking, and provided a demo example that supports Human vs Non-Human classification, but I don't have EVK, so it's hard to check the function directly.
    Please let me know the maximum number of people, distance, etc. that can be detected and tracked in the demo.
  2. Also, for the Human vs Non-Human classification feature, I would like to know more about the performance and introduction of the algorithm that determines adult vs child vs pet vs moving machine, etc.
    If there is any related documentation, please share it as well.

  3. If they want to develop an app that includes multi-person tracking for less than 5 people, human vs non-human categorization, and a usecase for measuring vital signs in the same binary, are there any limitations such as memory size, frame rate, maximum distance ?
    If they want to develop firmware that includes all of the following features, please let us know if there are any design restrictions or considerations.
    - Multi-people detection and tracking
    - Human vs Non-Human Callsification
    - Vital Sign detection at least single person.

Thanks and Regards, 

Jack

  • 0
    TI__Mastermind 20991 points

    Hi Jack, 

    I'm looking into your questions. I will reply shortly with more information. 

    Thank you, 

    Josh

  • 0 in reply to Josh Dye
    TI__Mastermind 20991 points

    Hi Jack, 

    Thank you for your patience. 

    Jack said:
    Please let me know the maximum number of people, distance, etc. that can be detected and tracked in the demo.

    The default configuration files for this demo enable tracking of up to 5 targets (humans or non-human moving objects). It is configured for a maximum range of ~10m. It may be possible to extend the range with the demo but we have not tested this. Also, the classifier model is only trained on data collected from targets up to a max distance of ~10m at different angles from -50deg to 50deg.

    It's also worth noting that while the tracker implementation is able to handle tracking of up to 5 people, it typically requires that people are separated by at least 0.5m.

    Jack said:
    for the Human vs Non-Human classification feature, I would like to know more about the performance and introduction of the algorithm that determines adult vs child vs pet vs moving machine, etc.

    First I want to direct you to the following papers for additional information regarding the classifier.

    https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10149609

    https://www.ti.com/lit/wp/swra774/swra774.pdf (case study 1)

    I also want to mention that the existing classifier demo is only a reference design to provide a proof-of-concept for an end-to-end classifier implementation on IWRL6432. Since this is a data driven approach, to leverage it to a product level solution, we always suggest customers to retrain the network with data in their scenarios.

    I'm still working on getting you an answer for #3. I'll get back to you by Monday of next week on this. 

    Best Regards,

    Josh