RasPiCam Vision Lite is a minimalistic and lightweight Flask web app that serves on Raspberry Pi and streams live video from its camera module at high frame rates while doing on-device image classification asynchronously with TensorFlow Lite models.

• RasPiCam Vision Lite allows you to efficiently serve and switch over multiple TensorFlow Lite models for image classification with just a few clicks.

• It takes advantage of concurrency and multiprocessing and shifts the computational heavy lifting of inference into dedicated subprocesses and threads, independent of video streaming feed, which to a large extent spares you the annoyance of choppy/stuttering frames when they're processed synchronously.

• You get more things done with a simple, beautiful and flat design. Because why not? (Click on the preview below for a demo video)

1. Open a terminal, SSH into your Raspberry Pi and clone the repository.

2. (Optional) To use models other than the defaults that come with this repository, you can still download and use some other offically released hosted models as this project does. Alternatively, train your own custom TensorFlow Lite (optimally quantized) models either via transfer learning or from scratch, and then place them along with labels inside models/ simply as the rest of them:

   .
   └─ models
      └── {your-model-version}
          ├── {your-model-version}.tflite
          └── labels.txt
   

3. Start a Flask web server by running python3 main.py under the repository root.

4. Open a browser and access the IP address distributed to Raspberry Pi in your local network with default port 5000 (e.g. 192.168.0.104:5000) . You'll be greeted by the web interface as shown above.

5. Go select one of the models from the dropdown list, and press the SERVE button to watch video live streaming as the serving model does inferencing in the background. Try selecting a different model and serve it again.

6. Shut down the server safely when you're done by clicking SHUTDOWN.

Following setups are fully tested. Some variations could also work, but no guarantee.

Hardwares

• Raspberry Pi 3 Model B+, 4 Model B
• Pi Camera Module V2

Softwares

• Raspbian 9+
• Python 3.5+
• TensorFlow 2.0
• OpenCV 4.1
• Flask 1.1.0+
• Picamera 1.13
• Chrome, Safari

MIT