Using a custom YOLOv8-based image classification model to detect if my driveway gate is open or closed and set state in Home Assistant.
I like home automation and I've already automated the garage door, so I wanted to automate the driveway gate too. It is a swing gate controlled by a FAAC EO24S controller which I've already set up a WiFi relay to open and close, but I couldn't see a way to get the open or closed state of the gate.
I could have used a magnetic reed switch like most smart garage controllers do, but I didn't like the idea of a wire moving around with the swing arms.
I already had a CCTV camera with unobstructed view of the gate, which gave me an idea to give neural networks a go which I haven't done before.
I came across YOLOv8, an impressive free and open-source image object detection, segmentation and classiciation model.
For what I need, image classification was an ideal fit because I don't need to identify where the gate was in the image or its shape outline, just if it was open or closed, effective a binary state.
I used Roboflow to quickly label a dataset from my CCTV camera footage.
Since I was planning to be only use a crop of the gate for the actual model inference, I also pre-cropped all my training images before uploading them to Roboflow.
I tried to find examples of open and closed gates, in different weather and lighting conditions, as well as with vehicles and people across it. I ended up with 87 images roughly equally split between open and closed.
I trained 300 epochs using the yolov8s-cls.pt pretrained model using my RTX 4070 Ti GPU which only took a few minutes.
I planned to run this on my Windows home server which only has an integrated GPU, so it was practically going to run on CPU. I found the ONNX format which could be ran in Node.js using Microsoft's ONNX Runtime so I exported the model to ONNX which YOLO CLI had built-in support for.
I then wrote a simple Node.js script to basically loop through downloading the latest snapshot from my CCTV camera, crop it, run inference with the custom model, filter and find the most confident prediction.
Because I used the "small" pretrained model, inference only took approximately 80ms on a Intel i5-9600K CPU. In fact downloading the image was significantly slower at approximately 900ms. So the whole update loop takes around 1 second which is good enough for my needs.
I set up my Node script to push the state into my Home Assistant as a binary_sensor by simplying POSTing the state to Home Assistant.
{
state: classification === "open" ? "on" : "off",
attributes: {
device_class: "garage_door",
friendly_name: "Driveway gate",
},
},I wanted to tie together this open/closed state with my existing Wifi relay to more accurately represent the gate as a cover entity in Home Assistant so I used a template to merge the two together
cover:
- platform: template
covers:
driveway_gate:
device_class: gate
friendly_name: "Driveway gate"
value_template: "{{ is_state('binary_sensor.driveway_gate', 'on') }}"
open_cover:
service: switch.turn_on
target:
entity_id: switch.sonoff_gate
close_cover:
service: switch.turn_on
target:
entity_id: switch.sonoff_gate
stop_cover:
service: switch.turn_on
target:
entity_id: switch.sonoff_gateFinally, since I do most of my actual home management in Apple Home, so I re-exported the merged gate entity to HomeKit.