imagine
Imagine, an AI video monitoring framework
Imagine is a crystal lang web platform for AI processing and monitoring video streams.
- takes UDP H.264 video streams as input (needs to be web compatible format)
- runs frames through TensorFlow Lite models
- outputs detection data to a websocket and optionally a Redis channel or webhook
- detection coordinates only
- optionally the PNG image (websocket and webhook)
- optionally the PNG image with markup applied (websocket and webhook)
- grabs segments from the video stream and makes it available as MPEG-DASH for browser viewing
- this will be pushed to a local volume for serving
- and optionally can be pushed to a S3 for serving video at scale
This is intended to run independently of any applications that make use of the output. KISS
Documentation
Use the provided docker-compose to launch the application.
It's not intended to be public facing as it will not scale (without wasting a lot processing power) but is designed so that it can be used in scalable applications. Obviously works fine as a frontend for your personal Raspberry Pi projects.
Pre-prepared TF Lite models
These are a bunch of models available that will work with Coral Edge TPUs and the example object detection model
Converting TF Models for use
- List of pre-built detection models
- guide for converting these models to run on TF Lite
- Download the tooling:
pip3 install tf-models-official
pip install tensorflow-object-detection-api
- Download the tooling:
Repository
imagine
Owner
Statistic
- 4
- 0
- 0
- 1
- 3
- about 1 year ago
- December 18, 2022
License
MIT License
Links
Synced at
Wed, 06 Nov 2024 21:36:41 GMT
Languages