Skip to content

Neighborhood Hacks 2026 · Safety & Emergency Response

ARGO

A hundred eyes on Rome’s metro.

ARGO reads live CCTV and flags violence, weapons and snatching within seconds, so the operator acts on the moment instead of reviewing it afterwards.

Six real videos, normal ones last. Everything drawn on them is the real output of ARGO: skeletons, the weapon box, violence and the alarm. Public clips (CC BY, Pexels).

I

The neighbourhood

Termini, the Metro B platforms, the escalators at rush hour: the same scenes repeat every day. A phone pulled from a hand as the doors close, a bag lifted in the crowd, an argument that turns into a fight on the mezzanine. The State has answered with more police in the stations, and since January 2025 the area around Termini is a reinforced-surveillance zone. Crime has not gone away. The weak point is not the number of cameras, it is the attention behind them: one operator cannot watch hundreds of feeds, and most footage is reviewed only after the fact, when the damage is done. ARGO starts from there.

II

Rome, the crimes

Official figures show where the pressure is. Below are the three street crimes ARGO is built for, across Rome’s 15 districts: the totals are official, the split by district is our declared estimate. Thefts from people outnumber everything else.

Hover over a districtto see population and estimate

Boundaries, metro, stations: © OpenStreetMap contributors, ODbL

Sources and method

    III

    Weapons

    Statistics count crimes, not the tools used to commit them. The Ministry of the Interior sorts Rome’s crimes into 34 categories, and none of them separates offences committed with a weapon: nobody can say how many robberies on the metro involved a knife. A camera can see it, if someone is watching. That is why ARGO treats weapons as a separate signal: a dedicated detector looks for pistols, knives and rifles in every frame, and a weapon only counts when it is in someone’s hand.

    The 34 categories the Italian Ministry of the Interior uses to count crimes in Rome (2024). None of them separates crimes committed with a weapon.

      Source: Italian Ministry of the Interior, crimes committed in Rome 2021–2024

      IV

      How it works

      ARGO does not watch pixels the way a person does. It turns every frame into skeletons, reads 1.6 seconds of movement at a time, and asks two networks and a weapon detector the same question: is someone in danger right now?

        V

        Results

        Loading the results

        The test results are read from dati/risultati.json.

        VI

        Who would use it

        ARGO is built for law enforcement and for whoever sits in front of the camera wall: metro and station control rooms, local police, private security. Instead of searching dozens of feeds, the operator receives the event: which camera, what is happening, how confident the system is. The same alert can be pushed by software to officers in the field, so the patrol closest to the platform knows before anyone calls. ARGO does not decide and does not intervene: every alarm is checked by a person.

        Limits and next steps

        Fights are ARGO’s strength: 90% accuracy on 30-second clips it has never seen. Armed violence and snatching are still weak (F1 0.10 and 0.13) because public footage of them is scarce, and the WEAPON WARNING catches only a small share of real weapons. On a laptop the panel runs with a delay. Next: real footage from a transport operator, a zoom on the hands for small weapons, a larger detector on GPU, and alerts delivered straight to officers’ devices.

        Measured during development

        On a frame of Largo Argentina filmed from above, the YOLO26s detector finds 22 people and the YOLO26s-pose model 0: with small, distant people the pipeline can lose them.

        Tools and artificial intelligence

        Trained on Kaggle’s free GPUs, on public datasets only. Built with Claude Code as programming assistant: every design choice was made, measured and logged by the author, and every number on this page comes from a test the code can repeat.