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-ODS: Indoor Overhead People Detection and People Counting - Resolving multiple people when close together

Part Number: IWR6843ISK-ODS
Other Parts Discussed in Thread: IWR6843ISK

Hello everyone, I'm facing a problem with my overhead people counting algorithm, it is placed above a hallway in front of an entrance door, the use case is a bi-directional counter for understanding the number of entries and exits.

The algorithm is very lightweight as it considers and processes just the X coordinate of the target it's seeing, this is due to the nature of the use case; the hallway is along the X direction therefore just one coordinate is needed.

The problem is that it struggles in separating people when passing close together, it considers those as a single target (fig1 "current behaviour") while being multiple target (fig2 "Expected behaviour").    

In my case the following parameters are valid: h=2.7m L=6m Θ=64°  for such height i discovered i can reach a max detect distance of ~2.15m (using the formula found here) which is actually enough as my major problem is the detection sensitivity inside this area.

The hallway is about 6m long and large 2m, the sensor is installed on the ceiling (height about 2.7m), you can see the following fig3 for a brief summary map of the installation top view.

The cfg I am using is the following:

% SDK Parameters
% See the SDK user's guide for more information
% "C:\ti\mmwave_sdk_[VER]\docs\mmwave_sdk_user_guide.pdf"
sensorStop
flushCfg
dfeDataOutputMode 1
channelCfg 15 7 0
adcCfg 2 1
adcbufCfg -1 0 1 1 1
lowPower 0 0
% Detection Layer Parameters
% See the Detection Layer Tuning Guide for more information
% "C:\ti\mmwave_industrial_toolbox_[VER]\labs\people_counting\docs\3D_people_counting_detection_layer_tuning_guide.pdf"
profileCfg 0 61.2 60.00 17.00 50 657930 0 55.27 1 64 2000.00 2 1 36
chirpCfg 0 0 0 0 0 0 0 1
chirpCfg 1 1 0 0 0 0 0 2
chirpCfg 2 2 0 0 0 0 0 4
frameCfg 0 2 224 0 120.00 1 0
dynamicRACfarCfg -1 10 1 1 1 8 8 6 4 4.00 6.00 0.50 1 1
staticRACfarCfg -1 4 4 2 2 8 16 4 6 6.00 13.00 0.50 0 0
dynamicRangeAngleCfg -1 10.000 0.0010 2 0
dynamic2DAngleCfg -1 5 1 1 1.00 15.00 2
staticRangeAngleCfg -1 0 1 1
antGeometry0 0 0 -1 -1 -2 -2 -3 -3 -2 -2 -3 -3
antGeometry1 0 -1 -1 0 0 -1 -1 0 -2 -3 -3 -2
antPhaseRot 1 -1 -1 1 1 -1 -1 1 1 -1 -1 1
fovCfg -1 64.0 64.0
compRangeBiasAndRxChanPhase 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0
% Tracker Layer Parameters
% See the Tracking Layer Tuning Guide for more information
% "C:\ti\mmwave_industrial_toolbox_[VER]\labs\people_counting\docs\3D_people_counting_tracker_layer_tuning_guide.pdf"
staticBoundaryBox -4 1.3 -0.9 2 -0.5 2.7
boundaryBox -5 1.4 -1 3 -0.5 2.7
sensorPosition 2.7 0 90
gatingParam 2.5 1.5 1.5 2 3
stateParam 3 3 6 15 3 1000
allocationParam 10 10 0.02 7 1 20
maxAcceleration 1 0.1 1
trackingCfg 1 4 800 20 37 33 120 1
presenceBoundaryBox -4 4 -4 4 0.5 2.5
sensorStart

The boundary box is asymmetrical as the position of the sensor is not centered for physical limitations. As you can see I already changed the parameters as indicated in the TI document regarding trouble-shooting with the board configuration (at this link). In particular I already tried to modify the gatingParam and the allocationParam but it looks like it won't help in increasing the distinction between the targets.
Moreover the radar antennas are covered by a thin sheet of plexiglas (aka PMMA) which is 1mm thick, could you confirm it is not creating interference?
In conclusion I would like to receive some quality advice for coping with the problem of "Resolving multiple people when close together".
Thank you,
Filippo
  • Hi Filippo, 

    How far apart are the people when they are tracked as one? Typically, a minimum of at least 0.5m separation is required. Due to the fundamental limitations of the antenna you will likely have trouble resolving people that are not at least 0.5m apart. 

    Moreover the radar antennas are covered by a thin sheet of plexiglas (aka PMMA) which is 1mm thick, could you confirm it is not creating interference?

    The plexiglass cover should not really be giving you any issues especially since you are dealing with relatively short ranges. If you are interested on the potential effects please read our Radome Design Guide.

    Best Regards,

    Josh

  • Thank you for your fast response. 

    Indeed, my problem is that people often pass by the sensor closer than 0.5m to each other. Are there any method to increase the sensitivity of the sensor to less than 0.5m maybe changing some parameters from the algorithm or from the PCL computing section inside the sensor firmware.

    What other TI radar alternatives exist for people detection at a higher resolution (<0.5m sensitivity)?

    Thanks,

    Filippo

  • Hi Filippo, 

    Let me look into this further and get back to you in the next 24 hours. I appreciate your patience. 

    Best Regards,

    Josh

  • Hi Filippo, 

    I wanted to direct you to our tracker tuning guide which may help as you try to tweak the tracker parameters to improve the performance. 

    What other TI radar alternatives exist for people detection at a higher resolution (<0.5m sensitivity)?

    Is it possible for you to switch to a wall mount configuration and using an antenna pattern like that of the IWR6843ISK EVM? This may help slightly as the better angular resolution in the azimuth direction should give the gTrack algorithm more information to work with. Though I believe this will not entirely resolve the issue if 2 people are very close to each other but you should see some improvement.

    One might also explore using a ML approach to solve this problem. There is an area of ML that looks at these types of problems called Gaussian mixture modelling. This might be something you could explore if you have the resources.

    There is also this paper which explores alternative approaches to solving this problem. However, currently our SW offering only contains the group tracker method for people counting/tracking. 

    Best Regards,

    Josh