AI Violence Detection for K-12 Schools
A real-time, privacy-first computer vision system that turns the IP cameras you already have into a violence-detection layer — and notifies the right adult before a fight escalates.
Last updated: April 25, 2026
What problem does GuardianAI solve for K-12 schools?
GuardianAI converts existing CCTV into a real-time fight-alerting system. The average school administrator learns about a physical altercation 8–12 minutes after it starts — through a student tip, a teacher walking in, or an injury at the nurse's office. By that point, the fight is over and the school has only post-hoc reports. GuardianAI compresses that 8–12-minute lag to 1–2 seconds.
Every K-12 school in the country already has cameras. Most don't have anyone watching them. Traditional CCTV is a forensic tool: it tells you what happened after it happened. GuardianAI is the live alerting layer on top of the cameras you already own. A pose-based AI model watches every stream 24/7 and pings the school resource officer (SRO), vice principal, or designated on-call adult when it sees fighting, hitting, or shoving — usually within 1–2 seconds of the first contact.
How is GuardianAI deployed in a school?
- Connect existing cameras. Any RTSP stream works — Hikvision, Dahua, Avigilon, Axis, Reolink, even a USB webcam. No new hardware required.
- On-prem AI inference.The detection runs on a single mid-tier GPU box (NVIDIA Jetson Orin or equivalent) installed inside the school's network. Video never leaves the building.
- Skeleton-only analysis. The model uses YOLO11n-Pose to extract 17 skeletal keypoints per person, then a spatiotemporal graph network classifies the motion pattern. No faces. No identifying features. Just geometry.
- Instant Telegram alert. When the model fires, the on-call adult gets a message with the camera location, a confidence score, and a short clip preview. They confirm or dismiss with one tap. Confirmed events create an incident record.
Why pose-based detection instead of facial recognition?
Most off-the-shelf "school safety AI" products run face recognition. That model is illegal in many U.S. states for K-12 use (e.g., New York's 2023 moratorium, Illinois HB100), violates GDPR's Article 9 in the EU, falls under stricter consent rules under Russian 152-ФЗ, and creates a permanent biometric record of every minor who walks past a camera.
GuardianAI's pipeline is fundamentally different. The CTR-GCN classifier looks at how bodies move relative to each other, not who they belong to. The model literally cannot identify a student — it operates on (x, y, confidence) coordinates of joints, with the camera image discarded after inference. Concretely:
| Approach | Identifies people? | GDPR / FERPA / 152-ФЗ safe? | Survives lighting changes? | Survives camera angle? |
|---|---|---|---|---|
| Face recognition | Yes (always) | No (Article 9 / biometric) | Yes | Sometimes |
| Pixel-only CNN classifier | Sometimes | Indirect risk (raw video retained) | No | No |
| Pose-based (GuardianAI) | No (architecturally impossible) | Yes (no biometric storage) | Yes | Yes |
What's included with GuardianAI for schools?
| Component | What it does |
|---|---|
| On-prem GPU appliance | NVIDIA Jetson Orin Nano 8 GB or equivalent. Handles up to 50 cameras at 30 FPS each. ≈$1,200 one-time hardware cost. |
| Live dashboard | Single web view showing every connected camera with red-bordered overlays when violence is detected. Unlimited operator seats. |
| Telegram bot | Per-school chat groups for SROs, admins, counselors. One-tap confirm / false-alarm loop retrains thresholds over time. |
| Incident log | Searchable history with timestamps, cameras, and resolved / false-alarm labels. CSV export for board reports. |
| Analytics heatmap | Incident times and locations across the school day so you can adjust supervision (e.g., the 12:05–12:10 cafeteria rush, the 3:15 bus loop, back stairwell during passing periods). |
| Quarterly model retraining | Model is fine-tuned on confirmed incidents from your own deployment, raising precision per-school over the first 90 days. |
| 24/7 NOC monitoring | Appliance uptime, stream health, and Telegram delivery monitored from our side. Median MTTR on appliance issues: under 4 hours. |
How long does deployment take?
A typical 800-student building goes from contract signature to first live alert in 10–14 days. The bulk of that time is network configuration (RTSP credentials per camera) and one site visit to install the on-prem GPU box. No camera replacement, no rewiring, no curriculum disruption.
| Day | Milestone |
|---|---|
| Day 0 | Contract signature |
| Day 2 | School IT provides RTSP credentials and camera inventory (typical 30–60 cameras) |
| Day 5 | GuardianAI ships the pre-configured GPU appliance |
| Day 7 | Site visit: install appliance in network rack, connect to camera VLAN, validate stream ingestion |
| Day 10 | Designated staff phones onboarded for Telegram alerts; first live alerts begin |
| Day 14 | First weekly review; confidence threshold tuned per the school's tolerance |
What does GuardianAI cost per camera?
GuardianAI is sold as an annual per-camera subscription. $200–$600 per camera per year depending on volume tier. A 30-camera K-12 school typically pays $9,000–$15,000/year all-in. Hardware (Jetson Orin Nano appliance, ≈$1,200) is one-time. We don't charge per-incident or per-alert.
| Volume tier | Per-camera per year | Notes |
|---|---|---|
| 1–25 cameras | $600 | Single small site, full SLA |
| 26–100 cameras | $400 | Most K-12 deployments |
| 101–500 cameras | $300 | Mid-size school district |
| 501+ cameras | $200 | University campuses, district-wide |
Pilot programs are billed at a flat 2-month fee with full credit toward year one if converted. Full pricing details: /pricing. For a current quote: hello@guardianai.tech.
Frequently asked questions
Does GuardianAI work on existing analog cameras?
Yes, as long as the camera feeds into a DVR or NVR that exposes an RTSP stream. About 95% of the units installed in U.S. schools since 2018 do. Verified compatible: Hikvision, Dahua, Avigilon, Axis, Reolink, Hanwha Wisenet, Bosch.
Can it run if our internet goes down?
Yes. Detection is fully local. If the internet is offline, the dashboard continues to function on the school LAN; only the Telegram alert delivery is interrupted, and queued messages flush automatically when connectivity returns.
What about privacy law (FERPA, COPPA, GDPR)?
Because the system never stores or transmits identifiable images of minors, it sidesteps the disclosure rules that apply to traditional CCTV. We provide a full Data Processing Agreement and DPIA template on request. See our privacy policy for details.
Will it generate false alarms?
Yes — every detection model does. On internal validation across 14 K-12 deployments (≈47,000 hours of recorded footage), GuardianAI sustains 94% precision and 89% recall at 0.5 confidence threshold. The model is explicitly trained to ignore running, hugging, sports horseplay, and salam greetings — the most common false-positive triggers. The confirm/false-alarm feedback loop further reduces noise per-school over the first 90 days.
Is it appropriate for elementary schools?
Yes. The model was trained partly on the UBI-Fights dataset and a curated set of school cafeteria/playground clips, so it generalizes to younger and shorter children. We do recommend a higher threshold (0.65+) in elementary settings to filter out roughhousing.
Ready to see a live demo on your own footage? Email us or read about how GuardianAI is deployed on university campuses.