Experimental pipeline processing for Python. Intended for developing basic AI or IoT pipelines for testing ideas and developing PoC. The idea is to have something that reminds of Google MediaPipe for edge computer vision and audio pipelines but implemented in plain Python rather than in C++. The pipeline configuration format is heavily inspired by Google MediaPipe.
You need Python 3, Package Installer for Python (PIP), PyAudio and OpenCV.
> sudo apt update
> sudo apt install python3 python3-pip
> sudo apt install python3-pyaudio
> # Some system have OpenCV pre-installed or might require manual installation
> sudo apt install python3-opencv
> # If xcode not installed before (needed for portaudio)
> xcode-select --install
> # If Python 3 not already installed
> brew install python@3.8
> brew install portaudio
> pip3 install pyaudio
> brew install opencv
> pip3 install opencv-python
Please search online for instructions for your operating system.
It is highly recommended to use Python virtual environments to enable separated environments for different Python projects.
> pip3 install virtualenv
> pip3 install virtualenvwrapper
Some systems might require extra installation steps. Please search online for instructions for your operating system.
Create virtual environment for Edgepipes.
> mkvirtualenv edgepipes -p python3
> workon edgepipes
> cd edgepipes
> pip3 install -r requirements.txt
Some examples have additional requirements. To run the face recognition example:
> pip3 install -r face/requirements.txt
Edgepipes are run with computer vision example by default (for now)
> ./edgepipes.py graphs/edge_detection.pbtxt
This will run the same graph as is available in Google MediaPipe as an example of the similarities. Press 'q' in the popup window to quit the example.
To specify the input video source:
> # First video device
> ./edgepipes.py graphs/edge_detection.pbtxt --input 0
> # Use IP camera via RTSP
> ./edgepipes.py graphs/edge_detection.pbtxt --input rtsp://192.168.1.237:7447/5c8d2bf990085177ff91c7a2_2
> # Screen capture (first monitor)
> ./edgepipes.py graphs/edge_detection.pbtxt --input screen
The input argument (except for screen) will be passed directly to OpenCV VideoCapture function. Please see OpenCV documentation for more options.
All examples use the default camera unless told otherwise using the --input argument.
Press 'q' in the popup window to quit the example.
Edge detection
> ./edgepipes.py graphs/edge_detection.pbtxt
Yolo 3 detector
> ./edgepipes.py graphs/yolov3detector.pbtxt
Hand tracker
> ./edgepipes.py graphs/handtracker.pbtxt
Face recognition - more images of "known" people can be added in the directory edgepipes/face/known
> ./edgepipes.py graphs/face_recognition.pptxt
Yolo 3 detector publishing detections via MQTT. This example requires a running MQTT broker in local computer.
> ./edgepipes.py graphs/yolov3detector_mqtt.pbtxt
Another option to start Edgepipes is to run it from an interactive CLI.
> ./pipecli.py
This will allow you to play with loading pipeline, starting stopping and printing the pipeline. Future features will be to add and remove parts of the pipeline at runtime, run at different speeds, debug, etc.
Download a model from https://alphacephei.com/vosk/models and unpack as 'model' in the current folder where Edgepipes are run.
A recommended model to start with is vosk-model-small-en-us-0.4.
- Add a way to distribute the pipeline processing over multiple threads and machines.
- Add a way to send messages over MQTT instead of passing results internally in Python.
- Serialize messages in Protobuf between processes (over network).
- Allow same config to run in different modes (e.g. local in single process or over MQTT with Protobuf) without massive configuration change.