Project · computer-vision parking monitor · in development

Argus

Cameras count free parking spots, residents get told – running on one small server.

Commissioned work, live in production. The running system stays private, so this page walks through how it's built rather than showing the deployment – the link opens a public overview.

The request was concrete: people needed a system that watches the parking area and tells them when something frees up. The chain is simple – city IP cameras, one small VPS, Telegram. It ran for several months against live cameras around the high-rise blocks, and every snapshot it took became training material.

Detection is not just stock YOLO out of the box: the archive was labelled and used to fine-tune the model on my own machine, then only the weights went to the server, and I compared the results against other detectors before settling on what actually performed. Inference runs on CPU alone – the whole thing costs a few euros a month. The link shows the approach on a public camera; the Telegram bot goes live for visitors shortly.

What it does

• Reads existing city IP cameras – no new hardware
• Models trained locally, only the weights ship to the server
• Inference on one minimal CPU server, no GPU
• Snapshot archive feeds the next training round
• Telegram bot answers with photo, map and forecast

Built with