Quick Start with Raspberry Pi, Pi Camera, and Libcamera

Feedback 💡
1

Today many students and others want to get started quickly using a low cost and widely distributed Raspberry Pi with V1, V2, or HQ (V3) Raspberry Pi Cameras. During the past few months there was a transition with Bullseye from what is now termed “legacy camera drivers” to Libcamera and subsequent development of Picamera2, the libcamera-based replacement for Picamera which was a Python interface to the Raspberry Pi’s legacy camera stack.

I am sure Raspberry Pi users around the world would appreciate an update to your Quick Start - YOLOv5 Documentation
to enable a Quick Start with Raspberry Pi with V1 or V2 or HQ (V3) Pi Cameras operating on Bullseye with Libcamera and Picamera2.

2

Thanks for the feedback. We’ll definitely improve our docs and add more content related to edge in the future.

3

I am providing a Quick Start with Raspberry Pi, Pi Camera (or webcam), and Libcamera:
With the release of Raspberry Pi OS Bullseye, the default camera stack is now libcamera.

This guide explains how to deploy a trained YOLOv5 model on a Raspberry Pi 3 or Pi 4 running the 64 bit version of the operating system with the libcamera camera stack.

Install the Raspberry Pi Operating System

Prepare a micro SD card with the latest version of the Raspberry Pi Operating System (64-bit). You can use the Raspberry Pi Imager tool to get everything set up properly:

Install Necessary Packages

  1. Make sure the Raspberry Pi is up-to-date

sudo apt-get update

sudo apt-get upgrade -y

sudo apt-get autoremove -y

  1. Clone the YOLOv5 repository

cd ~

git clone GitHub - ultralytics/yolov5: YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite

  1. Install the required dependencies

cd ~/yolov5

sudo pip3 install -r requirements.txt

  1. Install the correct version of PyTorch and Torchvision (at the time of writing, PyTorch 1.13.0 and Torchvision 0.14.0 were not supported)

sudo pip3 install torch==1.11.0

sudo pip3 install torchvision==0.12.0

Adjust detect.py

By default, detect.py doesn’t allow TCP streams to be used as a source and fails when used via SSH or the Raspberry Pi Command Line Interface. To fix this, make two minor modifications to detect.py:

  1. Open detect.py and find the ‘is_url’ line

cd ~/yolov5

sudo nano detect.py

CTRL + W → is_url → ENTER

  1. Add TCP streams as an accepted url-format

is_url = source.lower().startswith((‘rtsp://’, ‘rtmp://’, ‘tcp://’, ‘http://’, ‘https://’))

  1. Find the line that says ‘view_img = check_imshow(warn=True)’

CTRL + W → view_img = check_imshow(warn=True) → ENTER

  1. Comment out this line

*#*view_img = check_imshow(warn=True)

  1. Save the modifications and close detect.py

CTRL + O → ENTER → CTRL + X

Initiate a TCP stream with the Libcamera driver

libcamera-vid -n -t 0 --width 1280 --height 960 --framerate 1 --inline --listen -o tcp://127.0.0.1:8888

Perform YOLOv5 inference on the TCP stream

  1. Navigate into the yolov5 directory

cd ~/yolov5

  1. Run detect.py with the TCP stream as its source

python3 detect.py --source tcp://127.0.0.1:8888

  1. Note: by commenting out ‘view_img = check_imshow(warn=True)’ we made sure detect.py no longer display its source + predicted bounding boxes by default to enable inference via SSH or the Command Line Interface. To re-enable this feature, run detect.py with the --view-img argument:

python3 detect.py --source tcp://127.0.0.1:8888 --view-img