Motion Ã👁

My primary hacking effort is focused (currently) on the underlying platform for AI at the edge; to wit: http://github.com/motion-ai/motion-ai

The capabilities of AI are very compelling, at least in the laboratory and corresponding research paper. However, deploying AI into the field requires both an understanding of the signal produced, but also the mechanisms for attenuation of that signal and the refinement of the AI itself.

The first objective is to generate a signal and understand it sufficiently to answer a question; e.g. did anyone get up today and if so, when; if not, when did you last see a person and where?

The Motion Ã👁 platform comprises a set of Docker containers running on a distributed network of computers from the very small, e.g. RaspberryPi model 0w, to the more robust, e.g. AMD64 server with nVidia GPU accelerator.

The containers provide different services:

1. The motion addon for Home Assistant (aka the motion2mqtt service for Open Horizon) provides the starting point for collecting video information using the Motion Project (motion-project.github.io) software to interface to RTSP, HTTP, and V4L2 connected devices.

2. A MQTT broker to provide communications between and amongst the services as well as to integrate with Home Assistant using sensors and cameras.

3. A suite of analysis containers, notably those for AI detection and classification algorithms:

   1. YOLO - You Only Look Once; a high-performance, light-weight entity detector and classifier. Versions tiny-v2, tiny-v3, v2, and v3. Supports both nVidia GPU accelerated and CPU-only inferencing on AMD64, ARM64 and ARM7 architectures.

   2. FACE - An open source face detector -- not recognizer -- to be used in conjunction with detection using YOLO with an objective of improving the signal quality; for example, building a database of faces for eventual face recognition.

   3. ALPR - An open-source license plate detector and recognizer for EU and USA, including options for country and state variations.

Together these services provide information to Home Assistant to capture and process through its complex-event processor, i.e. automations, templates, and other mechanisms to provide a simple interface.

Analysis of these events can be performed using more powerful time-series database, query, and visualization tools. Home Assistant addons for InfluxDB and Grafana provide a starting point and full-featured data analysis may be performed using the Jupyter notebook addon.