I’ve been working on a locally deployed, edge-AI tool called Lot Vulture that turns existing IP / RTSP security cameras into real-time parking space occupancy monitors (using an interactive polygon editor and on-device computer vision).
I’m currently thinking about integrating it into Home Assistant to do things like automate driveway lights, trigger gate relays, and show live spot status on dashboard tablets. It currently supports publishing events to an MQTT broker, but before I lock in the architecture, I wanted to get some feedback from Home Assistant experts on the best integration pattern.
How it currently works:
1. You point Lot Vulture to your RTSP stream (or Blue Iris / NVR substream).
2. You trace 4-point polygons over your driveway or parking spaces.
3. An on-prem C++ inference engine runs against the feed with dual-threshold hysteresis (to prevent headlight/shadow flickering).
4. When a spot state changes, it publishes an alert envelope over MQTT:
{
"event_id": 123,
"rule_id": 7,
"rule_name": "Driveway Spot 1",
"trigger_type": "space_occupied",
"fired_at": "2026-09-05T12:00:00.000000+00:00",
"data": {
"trigger_type": "space_occupied",
"edge": "became_occupied",
"space_state": "occupied",
"space_id": 1,
"space_name": "Driveway Spot 1",
"camera_id": 2,
"camera_name": "Driveway Camera"
}
}
Needing Advice:
1. MQTT Discovery vs Generic Topics: Would you prefer having Home Assistant MQTT Auto-Discovery (where Lot Vulture automatically registers each parking space as a binary_sensor entity in HA with zero YAML), or do you prefer simple manual topic routing?
2. Topic Hierarchy: Right now it publishes to a central lotvulture/alerts topic. Would a per-entity state topic like lotvulture/<camera_id>/<space_id>/state (ON/OFF) with the retain flag set be much easier for your automations?
3. Camera Feed Overlays: Would it be useful to expose an MQTT image entity or camera stream with the live green/red bounding polygons drawn so it can be embedded on dashboard cards?
4. Is it possible/sensible to pull video from Home Assistant, or is it easier/better for the tool to pull its own video?
Code:
The core software is free and source-available (runs on Docker or native Windows .exe):
GitHub:
There’s a commercial option that provides a higher accuracy AI engine.
I appreciate any guidance on how to best structure it for HA.
