Help with Pose Estimation on Raspberry Pi AI Hat (Hailo-8) – Project by Student Developer

1

Help with Pose Estimation on Raspberry Pi AI Hat (Hailo-8) – Project by Student Developer

2

Hi @Mridul303453 and welcome to the forums. Glad to have you here.

Forum members would be happy to help out with your project, but we will need a little more information about the specific problems you are encountering so that we can give adequate guidance.

1 Like
3

face-recognition integration on Raspberry Pi 5

4

To: jane@coreelectronics.com.au
Subject: Request for help with Hailo-8 installation, performance, and face-recognition integration on Raspberry Pi 5

Dear Jane,

I hope you’re doing well. Thank you very much for your continued support — your help has made it much easier for me to work with the Raspberry Pi AI Kit (Hailo-8 + Sony IMX500).

I wanted to share a few issues I’ve encountered and a small request regarding future tutorials.

1. Installation issue on Debian 13 (Trixie)

When I run:

sudo apt install hailo-all

I get:

Unable to locate package hailo-all

However, when I switch to Raspberry Pi OS Bookworm (64-bit), everything installs perfectly.
Could you please confirm if Debian 13 (Trixie) is currently unsupported for Hailo, or if there’s a manual method to install it (like a .deb file or alternate repository)?

2. Performance and QoS warnings

On Bookworm, the pose estimation and object detection demos work but run slowly. The terminal repeatedly shows:

Lots of QoS messages received: 100. Consider optimizing the pipeline or reducing the pipeline frame rate (see ‘–frame-rate’ flag).

Could you please suggest:

  • The best frame rate and resolution settings for Raspberry Pi 5.

  • Any pipeline optimization steps (queue, buffer, or frame-drop options).

  • How to check that the PCIe interface for the AI Hat is configured properly (Gen 3 mode, cooling, and power requirements).

3. Face-recognition integration

I’m also trying to combine Hailo-based person detection with face recognition (using face_recognition and dlib).

Could you please tell me:

  • Whether the Hailo accelerator can help speed up face-recognition inference.

  • Or if the recommended approach is to run YOLOv5s person detection on Hailo and use the Pi’s CPU for face matching.

  • If there are any example pipelines or SDK demos showing such a hybrid setup.

4. Tutorial request

Could you kindly forward a request to your technical team to create a dedicated tutorial for the Raspberry Pi AI Hat Kit, including:

  • Full installation on Raspberry Pi 5 (64-bit).

  • Hailo environment setup.

  • Object, pose, and face-recognition examples.

  • Tips for achieving smooth real-time performance (20 FPS +).

Such a guide would be extremely useful for makers like me working on AI-based robotics projects.

Thank you so much for your time, patience, and helpful guidance. Your team’s tutorials have been incredibly valuable for my learning.

Warm regards,
Mridul Pratihar
Student Maker / Raspberry Pi AI Enthusiast
Etah, Uttar Pradesh, India
:envelope: mradulpratihar333@gmail.com

5

Hey @Mridul303453,

Thanks for the kind words. It looks like you have done quite a bit with computer vision on the Pi 5! Some very good questions as well, let’s go through them!

  1. Unfortunately, OS Trixie does not support the AI HAT packages yet. Raspberry Pi said they are coming soon though!

Please note that packages for the Raspberry Pi AI HAT+ and AI Kit, for the TV HAT, and for Wolfram Mathematica are not yet available in Trixie. If you are using any of these products, we advise you to continue to use your existing Bookworm image for the time being. Versions of these packages will be made available for Trixie in the near future.

  1. Have you modified the pipelines at all? And are you running a larger size YOLO model? Previously, the pipelines have limited the speeds to 30 FPS? You should be able to run the default model that comes with it at 30 FPS without any issues - you shouldn’t need to change the resolution to achieve this. It also sounds like you might be more of an expert on the pipelines here! I explored them when the AI HAT launched and they have since been changed greatly.

    The AI HAT library should automatically enable Gen3 PCIe speeds by default as well. You can manually override this though by opening up the boot file with the following command in the terminal:

    sudo nano /boot/firmware/config.txt**
    **
    Then insert the following line somewhere near the top:

    ***dtparam=pciex1_gen=3

    There is quite a noticeable difference between gen2 and gen3 performance so if you are using gen2 this might be your issue!

  2. The library we use for face recognition does not run on the AI HAT. It is not very CPU-intensive and will happily run on the Pi itself. If you want to combine it with person detection, that is also entirely possible! Although, I’m not sure how that would help other than to know that there is a person there.

    In terms of pipelines, I don’t think you would need to create a custom pipeline to integrate this all. I think a good approach might be to just do it in your Python Library. Install all the packages into the same virtual environment and use each library as needed.

  3. That is a good idea for a guide! We have a big list of things we are working on for guides, but it is on our list to make some more AI HAT guides. Unsure when we will make them, possibly not for a while, but when we do we will explore what has changed since then and any tips and tricks that we have learn’t since the old videos!

Sounds like you have quite a project on your hands as well! Very awesome to see people actually applying this stuff in the real world - especially in schools!

Cheers,
Jaryd

2 Likes
6

integration on Raspberry Pi 5