LatropPoint a trail camera at the yard; get it to tell deer from raccoons from "that is just a branch moving again," and ping you only for the stuff you care about.
Next to-dos
- Improve night / IR-footage accuracy (the hard part)
- On-device vs cloud detection — decide on cost/privacy/latency
- Build the timeline review UI with one-tap label correction
- Wire push notifications (ntfy or Pushover) with per-category rules
- Decide clip retention policy (keep all vs flagged-only)
Recent activity
- Created project · 4 hours ago
- To-do added — Ingest clips from the camera upload folder · 4 hours ago
- To-do added — Daytime detection working end to end · 4 hours ago
- To-do added — Improve night / IR-footage accuracy (the hard part) · 4 hours ago
- To-do added — On-device vs cloud detection — decide on cost/privacy/latency · 4 hours ago
- To-do added — Build the timeline review UI with one-tap label correction · 4 hours ago
- To-do added — Wire push notifications (ntfy or Pushover) with per-category rules · 4 hours ago
- To-do added — Decide clip retention policy (keep all vs flagged-only) · 4 hours ago
Design doc
Trailcam Classifier — design doc
What it is: Software that watches a trail/security camera, classifies what triggered each clip (deer, raccoon, person, vehicle, or "wind"), and only notifies you for the categories you've opted into.
The problem it solves
Motion-triggered cameras cry wolf constantly — a swaying branch fires 40 clips a night. The signal is buried. This filters the noise so a "there's a person in the driveway at 2 a.m." alert actually means something.
Approach
- Ingest clips from the camera (RTSP or the camera's upload folder).
- Run a lightweight detector per clip → labels + confidence.
- Rules engine: notify on {person, vehicle}; log-only on {deer, raccoon}; drop {empty, wind}.
- A simple timeline UI to review and correct labels (corrections feed back into the rules).
Stack
A small detection model (YOLO-class or a hosted vision API) · Python worker · SQLite event log · push notifications via ntfy/Pushover.
Open questions
- On-device detection vs a cloud API — cost vs. privacy vs. latency?
- How much labeled local data do we need before accuracy is trustworthy at night (IR footage)?
- Clip storage policy — keep everything, or only flagged events?
Status
Detection works on daytime clips; night/IR is the hard part. Actively poking at it between other projects.