• State Resolved
  • Locked Locked
  • Replies 6 replies
  • Answers 2 answers
  • Subscribers 30 subscribers
  • Views 1123 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.

IWR6843ISK: Gesture recognition

Atsu
Prodigy 140 points
Part Number: IWR6843ISK
Other Parts Discussed in Thread: IWR6843AOP, ,

Hello TI team,

I'm trying Gesture_with_Machine_Learning demo in radar_toolbox_1_00_00_26.
While reading its source code, I got confused a little bit.

1.This code seems to use only 1Tx x 4Rx configuration in spite of performing 4 direction recongnition.
    cli.c line 83
     "channelCfg 15 1 0",
   Is this correct?   Should I chage it when I use 6843ISK?

2.This demo has two projects, that is, for 6843ISK and 6843AoP.
   But both projects look to use the same ANN parameter.  
   Which one does the parameter embedded in the code suit for?  Or do you intend to use the same ANN for both EVMs?

Anyway the pre-built binary worked fine for PoC.
I understand this demo is provided "as is" but...
Since this demo is very good start point for me, I want to exprore more.
As the second step, I'd like to reproduce the same binary code as you provided from the souce code.
For this purpose, could you please answer above my questions and whether this source code can make the same binary output you provided in the prebuilt folder?

Best regards,
Atsu

  • 0
    TI__Mastermind 21001 points

    Hi Atsu, 

    First I want to clarify about the supported EVMs for this demo. This demo supports IWR6843AOP EVM and IWR6843ISK-ODS. IWR6843ISK-ODS differs from IWR6843ISK in the antenna pattern. So IWR6843ISK is not supported for this demo, several modifications would be required in order to potentially support this EVM. 

    1. This setting is correct for both ODS and AOP EVMs. Because of the antenna design for these EVMs when only TX1 is enabled, we still have a 2 dimensional virtual antenna array with 2 azimuth elements and 2 elevation elements. 

    6843AOP EVM:

    6843ODS EVM:

    I understand your confusion because with TX1 only for ISK EVM you would get a 1 dimensional virtual antenna array with all elements in a single elevation. So yes, If you wished to try to enable ISK EVM for this demo, additional TX antenna should be used. 

    2. The same ANN is used for both the 6843AOP project and the 6843ISK-ODS project. The difference is just in the angle features processing. Due to the difference in the order of the antennas (see images above), the input to the angle FFT must be arranged differently for each EVM. 

    Atsu said:
    Anyway the pre-built binary worked fine for PoC.

    Can you please clarify? Are you using IWR6843ISK or IWR6843ISK-ODS? I would not expect the prebuilt binary to work for IWR6843ISK. Are you saying that is the case?

    Best Regards,

    Josh

  • 0 in reply to Josh Dye
    Prodigy 140 points

    Hi Josh,

    Thank you very much for your prompt reply.  Got cleared.
    My apologies I overlooked "-ODS" written in user manual and might confuse documents with older one.

    Regarding your question, I've tried the prebuilt binary on "IWR6843-ISK" and that resulted in poor output in the case of "up to down" though,
    I had accepted it as a challenge for future work.
    And I checked AoP could get much better result of course.

    Can I ask a question again because I am a beginner in ML.
    6843 ISK-ODS and 6843 AoP are similar indeed but the perfomance seems to be a bit different regarding gain and directivity. But you use same ANN.
    Does such a difference ,depends on how much different though, not impact much to the result?  Only alignment is important due to effect of normalize
    or something?

    Best regards,
    Atsu

  • 0 in reply to Atsu
    TI__Mastermind 21001 points

    Hi Atsu, 

    No problem! 

    Atsu said:
    6843 ISK-ODS and 6843 AoP are similar indeed but the perfomance seems to be a bit different regarding gain and directivity. But you use same ANN.
    Does such a difference ,depends on how much different though, not impact much to the result?  Only alignment is important due to effect of normalize

    Your understanding here is correct. While slight differences do exist, the feature extraction is robust enough such that the neural network is easily able to recognize the patterns. There is no noticeable difference in performance of the ANN between the two EVMs. 

    Best Regards, 

    Josh

  • 0 in reply to Josh Dye
    Prodigy 140 points

    Hi Josh,

    Thanks.
    Sorry, one more...
    One of my goal is to expand the area of gesture detection, if possible,for whole room.  For instance I want to extend the distance up to 3m.
    I wonder if this will be feasible.   
    I guess radar configuration maybe OK, but the results of recongnision will be worse, furthermore training process will be much tougher, in my guess though.
    Can you give me some guidances or advices for it if you have already tested such a case?

    Best regards,
    Atsu

  • 0 in reply to Atsu
    TI__Mastermind 21001 points

    Hi Atsu, 

    We have done some minimal testing at 2m distance using the same chirp design and got somewhat promising results. I think the maximum detectable range allowed by the configuration used for this demo is somewhere close to 3m so you are correct the radar configuration should be okay. You can confirm this by inputting the chirp parameters into the mmWave Sensing Estimator to see the derived high level parameters.

    One thing you will definitely need to change is specifying which range bins the feature extraction is to be performed on. This is defined in gesture.h by RANGE_BIN_START and RANGE_BIN_END. Currently, range bins 1-8 are used which corresponds to about 30-40cm in front of the radar with the configuration used here. 

    I'm also wondering, do you want to make the gesture recognition work anywhere up to 3m or are you trying to have it work at a fixed distance of 3m? The former would require a much larger data collection effort for retraining the neural network as data should be collected at all distances up to 3m. 

    Best Regards,

    Josh 

  • 0 in reply to Josh Dye
    Prodigy 140 points

    Hi Josh,

    My apologies for my late reply due to holiday week in Japan.
    And thank you very much for your answer. That sounds great.  I'll check it soon and want to share the result with my customer. Thanks!

    Our (customer's) goal is to use gesture everywhere up to 3m.  I cannot imagine how much diffilcult to take data for learning and testing them so far...
    I'm gonna try step by step.
    Anyway, I'd like to close this case.  Thank you very much for your support.

    Atsu