An open-source software solution for situational awareness from a network of video and audio sources. Utilizing Home Assistant, addons, the LINUX Foundation Open Horizon edge fabric, and edge AI services, the system enables personal AI on low-cost devices; integrating object detection and classification into a dashboard of daily activity.
Start-to-finish takes about thirty (30) minutes with a broadband connection. There are options to consider; a non-executable example script may be utilized to specify commonly used options. Please edit the example script for your environment.
The following two (2) sets of commands will install
motion-ai on the following types of hardware:
arm); 2GB recommended
amd64); 2GB, 2vCPU recommended
arm64); 4GB required
The initial configuration presumes a locally attached camera on
/dev/video0. Reboot the system after completion; for example:
sudo apt update -qq -y sudo apt install -qq -y make git curl jq git clone http://github.com/dcmartin/motion-ai cd ~/motion-ai sudo ./sh/get.motion-ai.sh sudo reboot
When the system reboots install the official MQTT broker (aka
core-mosquitto) and Motion Classic (aka
motion-video0) add-ons using the Home Assistant Add-on Store (n.b. Motion Classic add-on may be accessed by adding the repository http://github.com/dcmartin/hassio-addons to the Add-On Store.
Select, install, configure and start each add-on (see below). When both add-ons are running, return to the command-line and start the AI's. After the MQTT and Motion Classic addons have started, run the
make restart command to synchroize the Home Assistant configuration with the Motion Classic add-on, for example:
cd ~/motion-ai make restart
Once the system has rebooted it will display a default view; note the image below is of a configured system:
A more detailed interface is provided to administrators only, and includes both summary and detailed views for the system, including access to NetData and the motion add-on Web interface.
The Motion Classic configuration includes many options, most which typically do not need to be changed. The
group is provided to segment a network of devices (e.g. indoor vs. outdoor); the
device determines the MQTT identifier for publishing; the
client determines the MQTT identifier for subscribing;
timezone should be local to installation.
Note: No capital letters [A-Z], spaces, hyphens (-), or other special characters may be utilized for any of the following identifiers:
group- The collection of devices
device- The identifier for the hardware device
name- The name of the camera
cameras section is a listing (n.b. hence the
-) and provide information for both the motion detection as well as the front-end Web interface. The
w3w attributes are required. The
icon attributes are optional and are used to locate the camera on the overview image. The
height, and other attributes are optional and are used for motion detection.
... group: motion device: raspberrypi client: raspberrypi timezone: America/Los_Angeles cameras: - name: local type: local w3w:  top: 50 left: 50 icon: webcam width: 640 height: 480 framerate: 10 minimum_motion_frames: 30 event_gap: 60 threshold: 1000 - name: network type: netcam w3w: - what - three - words icon: door netcam_url: 'rtsp://192.168.1.224/live' netcam_userpass: 'username:password' width: 640 height: 360 framerate: 5 event_gap: 30 threshold_percent: 2
Return to the command-line, change to the installation directory, and run the following commands to start the AI's; for example:
cd ~/motion-ai ./sh/yolo4motion.sh ./sh/face4motion.sh ./sh/alpr4motion.sh
These commands only need to be run once; the AI's will automatically restart whenever the system is rebooted.
The overview image is used to display the location of camera icons specified in the add-on (n.b.
left percentages). The mode may be
local, indicating that a local image file should be utilized; the default is
overview.jpg in the
www/images/ directory. The other modes utilize the Google Maps API; they are:
The zoom value scales the images generated by Google Maps API; it does not apply to
motion-ai solution is composed of two primary components:
Home Assistant add-ons:
motion- add-on for Home Assistant - captures images and video of motion (n.b. motion-project.github.io)
MQTT- messaging broker
FTP- optional, only required for
Open Horizon AI services:
yolo4motion- object detection and classification
face4motion- face detection
alpr4motion- license plate detection and classification
pose4motion- human pose estimation
Data may be saved locally and processed to produce historical graphs as well as exported for analysis using other tools (e.g. time-series database InfluxDB and analysis front-end Grafana). Data may also be processed using Jupyter notebooks.
amd64- Intel/AMD 64-bit virtual machines and devices
aarch64- ARMv8 64-bit devices
armv7- ARMv7 32-bit devices (e.g. RaspberryPi 3/4)
aarch64- with nVidia GPU
amd64- with nVida GPU
armv7- with Google Coral Tensor Processing Unit
armv7- with Intel/Movidius Neural Compute Stick v2
Installation is performed in five (5) steps; see detailed instructions. The software has been tested on the following devices:
This configuration includes dual OLED displays to provide display of annotations text and image, as well as a USB-attached camera (n.b. Playstation3 PS/Eye camera). The Intel/NCS2 implemtation is still in alpha mode and not in the
Low-cost computing (e.g. RaspberryPi, nVidia Jetson Nano, Intel NUC) as well as hardware accelerators (e.g. Google Coral TPU, Intel Movidius Neural Compute Stick v2) provide the opportunity to utilize artificial intelligence in the privacy and safety of a home or business.
To provide for multiple operational scenarios and use-cases (e.g. the elder's activities of daily living (ADL)), the platform is relatively agnostic toward AI models or hardware and more dependent on system availability for development and testing.
An AI's prediction quality is dependent on the variety, volume, and veracity of the training data (n.b. see Understanding AI, as the underlying deep, convolutional, neural-networks -- and other algorithms -- must be trained using information that represents the scenario, use-case, and environment; better predictions come from better information.
The Motion Ã👁 system provides a personal AI incorporating both a wide variety artificial intelligence, machine learning, and statistical models as well as a closed-loop learning cycle (n.b. see Building a Better Bot); increasing the volume, variety, and veracity of the corpus of knowledge.
This system may be used to build solutions for various operational scenarios (e.g. monitoring the elderly to determine patterns of daily activity and alert care-givers and loved ones when aberrations occur); see the Age-At-Home project for more information; example below:
Releases are based on Semantic Versioning, and use the format
MAJOR.MINOR.PATCH. In a nutshell, the version will be incremented
based on the following:
MAJOR: Incompatible or major changes.
MINOR: Backwards-compatible new features and enhancements.
PATCH: Backwards-compatible bugfixes and package updates.
David C Martin ([email protected])
localcamera and let me know
motion-ai as upstream to your repository:
git remote add upstream http://github.com/dcmartin/motion-ai.git
Please make sure you keep your fork up to date by regularly pulling from upstream.
git fetch upstream git merge upstream/master