Realtime Speed (FPS) for YOLOv8 and YOLOv9 on Raspberry Pi 5/4: Google Coral Edge TPU | Ultralytics

🚀 Dive deeper into the world of edge computing with our demo on 'Edge TPU Silva,' an exceptional framework tailored for the Google Coral Edge TPU, showcasing its integration with the versatile and powerful Raspberry Pi 4 and 5. Discover how these compact yet mighty devices are revolutionizing machine learning at the edge, making AI more accessible and efficient. Explore the project on GitHub: github.com/DAV....
🔍 Enhanced by Raspberry Pi: This guide shines a spotlight on the synergy between Raspberry Pi 4/5 and Google Coral Edge TPU in executing TensorFlow models efficiently. Whether you're utilizing the proven performance of the Raspberry Pi 4 or tapping into the advanced capabilities of the Raspberry Pi 5, you'll discover how these boards revolutionize machine learning deployment at the edge.
📌 Leveraging Raspberry Pi for Edge AI:
Learn about the integration of Raspberry Pi 4 and 5 with the Google Coral Edge TPU for seamless AI deployments.
Explore practical examples and step-by-step tutorials on setting up and deploying your AI projects on these platforms, available at our GitHub repository.
Discover the enhancements and optimizations specific to Raspberry Pi 5 that empower your machine-learning applications even further.
💡 Ideal For:
Tech enthusiasts exploring the frontier of edge AI with the latest hardware innovations.
Developers and makers seeking scalable and effective solutions for real-world AI applications.
Educators and students who are interested in hands-on AI learning experience through cutting-edge technology.
Join us as we explore the transformative potential of combining Edge TPU Silva with Raspberry Pi 4 and 5, ushering in a new era of accessible, powerful, and efficient AI at the edge. Experience firsthand how these technologies are redefining machine learning deployment and innovation.

[@oliverexcellent]

Thank you, thank you, thank you! I've been struggling to get my YOLOv8 + RPi 4 + Coral TPU up and running. I'm going to follow your tutorial this weekend and see if I can get everything working. Exciting!!

[@davidnyarko7300]

That sounds great. Post here if you facing any issues.

[@muthannaso]

Thank you for this amazing tutorial

[@davidnyarko7300]

Glad it was helpful!

[@anuj1788]

Hi, Can you help me with this error. silvatpu-linux-setup Illegal instruction

[@abelrey1576]

pip uninstall torch pip uninstall torchvision pip install torch==2.0.1 pip install torchvision==0.15.2 github.com/ultralytics/hub/issues/787#issuecomment-2263946340

[@rekomarie8779]

Hello Koby, i have a problem when i run the code like you did at 28:41 my RP4 reboot, i do not know why i have tried many time and it reboot again , i need to finish your tutorial so i could use my usb web cam to for live object detection, thank you in advance Edit: the RP does not reboot it closed remotely

[@anthonyholmes1312]

one issue I can't seem to get around when going back to Python 3.9 is that I cannot get libcamera to work for the pi camera module 3...either coral doesn't work with 3.11 or libcamera doesn't work with 3.9...driving me crazy!

[@andrei1916dtz]

Thank you so much for this amazing tutorial, it helped a lot with my project! Could you please help me with some informations about how I could use this setup for an USB webcam live inference object detection? Instead of being on .mp4 video files. I tried using the ultralytics script, but it didn’t work. Thanks again!

[@davidnyarko7300]

Hey! For a USB webcam, you can specify 0 or 1 or (your USB camera index) as the input and everything works the same with your camera.

[@davidnyarko7300]

Please provide context for the Ultralytics script and error.

[@samoo6931]

@davidyarko7300 please help me

[@sandrocain-l2v]

nice video~

[@ConsultingjoeOnline]

Great video. Thanks!

[@PenguinYan]

Thank for your tutorial maybe it's have possible to deploy YOLOv7 model on TPU with RPI4 I will try it when i'm get some time

[@davidnyarko7300]

You are most welcome.

[@arinjayagrawal9814]

Do you think this would work on the google coral devboard too?

[@AmanChauhan-xu7iw]

Hello Koby. Many thanks for this tutorial. I was earlier using RPI4 (32 bit legacy os) along with google coral accelerator and using edge tpu i was able to recognize object with my pi camera. I think it was using SSD model. I want to extend by using YOLO model and referring to this project. I am having a short query that this model work on edge tpu correct since we are using coral accelerator

[@alimohsen7071]

‏‪35:24‬‏ Thank you. The explanation was very good. But I have one problem. At this time 35:24 , I keep getting the same error. I could not solve it.

[@davidnyarko7300]

Thanks for the comment. The mentioned error occurs when the USB accelerator is connected to your Raspberry Pi when the code is executed. If it was connected then there is a loose contact just like in my case while working with pi5. You can change the port/ make sure you have firm USB contact.

[@alimohsen7071]

Thank you, the problem has been solved❤

[@JoeNassar-og3qe]

Hello Koby..Thanks for this video...I just bought dual edge coral TPU but in pci case for raspberry pi 5. . Can not this be used with this library?

[@shahriarahmadfahim6457]

Thank you for your video. How much FPS did you get for YOLOv8 and YOLOv9 respectively?

[@davidnyarko7300]

We did get several FPS in this video depending on the process type and model size. Some combinations gave the peek 60FPS, and 30FPS for the most part.

[@maksp6208]

Hello Koby. Can you help me pls? I want to run an object detection model on my raspberry pi 4 with a Coral USB accelerator and a model 3 wide camera. But I am having big problems with FPS currently only 3-4fps. Can you recommend me the best approach? I need to detect objects at a long distance but I have only class 1. I have been using mobilenet ssd fpn lite 320 and 640. I need to have minimum 25fps to close my project. I would appreciate any help. Thank you.

[@makratorian]

Same with steam deck 6fps 9 classes I had to resize to 240x240 2 classes now I get 22fps. I need 45fps min

[@samoo6931]

Please help me in that , i have a rpi 4 with 64 bit os , and i want to use camera module ip , with script ultrlytics

[@alimohsen7071]

Can I create my own model using yolov8? How do I convert it to tflite?

[@music_love21]

Hi! We are creating a system that classifies tomato ripeness levels using image processing in CNN architecture with the YOLOv8 model. We are using Raspberry Pi 4 OS with 4GB RAM and we have encountered a problem - the system has 2-3 minute delay/lag in classifying the ripeness level. Would you happen to have any recommendation/suggestion sir on this problem?

[@davidnyarko7300]

My first recommendation to you is this video, you could get the edge tpu USB accelerator to help you speed up the process.

[@anthonyholmes1312]

switched to a usb camera, but still struggling with getting it to run the model with a video feed. Is there an example code somewhere that does this using this model? No matter what I try I still get the "ValueError: Failed to load delegate from libedgetpu.so.1" error.

[@anthonyholmes1312]

ok so the cord that comes with the coral USB is crap. Replaced it and the TPU is actually recognized. ...facepalm...BUT I'm still struggling to find the right code to detect properly.

[@ucnguyen8653]

Can i use this set up for Raspberrypi zero 2w