Fusion body tracking can't achieve 30FPS running with 4 zed and RTX4090 while tracking multiple people?

Software C++
1

I used 4 zed2i cameras for multi people body tracking with one RTX4090 on windows 11 pc, now the configuration is below:
ZED SDK: V4.1.3
GPU Driver version: 560.70
CUDA nvcc version: 12.1
zed-unity-livelink-fusion program

  • InitParameters:
    • depth_mode: NEURAL
    • camera_resolution: HD720
    • camera_fps: 30
  • BodyTrackingParameters
    • detection_model: HUMAN_BODY_ACCURATE
    • body_format: BODY_18
    • enable_body_fitting: false (to avoid bad quality)
    • enable_tracking: false (to avoid bad quality)
  • BodyTrackingFusionParameters
    • enable_tracking: true (no data will be generated if false)
    • elable_body_fitting: true (to force output BODY_34 for unity demo)

But the release version of the program only get 20~23FPS when 4 people is being tracked, and in the situation with only one people will get almost 30FPS, How could I get higher FPS and lower latency when tracking 4 people?
Is there possible to use 2 RTX4090 in one pc to achieve better FPS? and how to bind each camera to different GPU core?

Thanks.

2

Hi David,

Thanks for reaching out.

I don’t think the bottleneck is the GPU here, adding a second one will not help a lot.

The metrics you are sharing are quite low, I think there is a problem in the fusion module slowing down the process.
I’ve shared your results internally and we’ll be carrying out further tests in the coming days to find the source of this problem.

With the API, you can select on which GPU the instance of the SDK will run, using the Initparameters::sdk_gpu_id (InitParameters Class Reference | API Reference | Stereolabs).

1 Like
3

@BenjaminV Thanks for the test and the guide of api.

Maybe my previous description about FusionParametrer::enable_tracking is inappropriately. The fusion program will show the skeleton of people but didn’t send out the data to network when I set FusionParametrer::enable_tracking to false.

I used the GPU-Z 2.59.0 to check the status of GPU, and found that the GPU Load won’t exceed 60% when 4 people being tracked, I don’t know if that is normal.

4

Hi,

You have to enable the tracking to be able to use the fitting option.
But you are completely right, there is no log warning the user or anything right now.

We’ll fix this.