Every year 274,000 pedestrians are killed on the world's roads.
Pedestrians are 23% of the 1.19 million people killed on the world's roads each year (WHO, 2023); in the EU alone 20,418 people died in traffic in 2023 (ETSC). We show a real pedestrian crossing LIVE: the system counts pedestrians, cyclists and vehicles, AI judges every conflict situation, and you verify its verdicts. The same open technology could watch any crossing — including yours.
Live camera — AI analyses in real time
○ OFFLINEPublic crossing camera (PL) · 24/7
How it works: 3 models that cooperate
Together they approach an expensive "unlimited" AI agent — for pennies. An event is created ONLY when a moving vehicle meets a pedestrian in the crossing zone (a car stopped at a red light is not an event).
Check the AI — click an event, watch the clip fullscreen, step through the rest
Loading events…
Crossing statistics — live from the database
Traffic per hour (48h)
Speeding by time of day
Vehicle speed distribution (indicative)
Episodes & AI violations per hour
Download the full crossing analytics: HTML report CSV data
Why this matters to government
pedestrians die on the world's roads — one every two minutes. One monitored crossing can save lives.WHO Global Status Report 2023
of all EU road deaths are pedestrians — and most happen in cities, at crossings.European Commission / CARE
the global and EU target: halve road deaths by 2030, zero by 2050 (Vision Zero / Safe System).UN Decade of Action · EU Road Safety Framework
we supply the objective data missing from infrastructure decisions: a real camera, open AI, public verification.Bezpieczne Przejścia / SafeCross
Open source · for researchers and integrators
The whole project is open (Apache-2.0) — anyone can reproduce, verify or extend it. It is a deliberately cheap, multi-model system (local detector + cheap LLM + human verification) that approaches the quality of an expensive "unlimited" AI agent for pennies.
What's in the repo
- CV service (YOLOX ONNX, tracking, zones, episodes, clips).
- AI layer: scene-context + per-event analysis (cheap LLM).
- Database, API, camera admin panel.
- Frontend, reports, docker-compose with resource caps.
For researchers
- We collect anonymous AI explanations + human ratings → a dataset for surrogate-safety and "did-not-yield" research.
- The method ports to any crossing (describe the scene → analyse).
- Honest error profile and limits (see "Accuracy").
Collaborate
Andrii Shramko — computer vision / VR / 3D (Poland).
Deployments, consulting, joint research and grants.
✉ zmei116@gmail.com
in linkedin.com/in/andriishramko
⌥ github.com/AndriiShramko/bezpieczne-przejscia
About the project
Bezpieczne Przejścia (“Safe Crossings”) is a personal project of Andrii Shramko — a VR, volumetric-video and computer-vision engineer. No grant, no company, no budget: one small CPU server, open code and a public city webcam.
The goal is simple: to show that road-safety monitoring can be transparent, cheap and open. An AI watches a live pedestrian crossing, flags dangerous situations and explains its decisions — and any visitor can verify and vote on those verdicts. Faces and license plates are blurred; only anonymized statistics ever touch the disk. All code is public (Apache-2.0).
Cities, road authorities, companies and researchers are invited to collaborate: share a camera over a crossing, pilot the system in your municipality, or build on the code. Code on GitHub · get in touch via the contact form.
Frequently asked questions — no marketing, honestly
How does the system detect violations?
The system does not guess from a single frame. An episode is created only when local kinematics shows two things at once: a pedestrian is walking on the crossing and a vehicle is moving through the same zone (a car stopped at a red light is not an episode). A sequence of up to 8 frames then goes to the AI, which issues a verdict with an explanation. Humans have the final say — every verdict can be confirmed or refuted.
Does the AI make mistakes?
Yes — and we don't hide it. That is why every verdict has ✓/✗ buttons: any visitor can confirm or refute the AI's judgement. From those votes we compute AI↔human agreement and publish it live on the home page. We treat events as candidates, not as facts.
How is vehicle speed measured?
It is a monocular estimate: pixel displacement between frames converted through a known lane scale. The result is indicative — the error can reach ±30%. It is not a radar or a measurement instrument, and it is not legal evidence of anything.
What about privacy and GDPR?
The system identifies no one: tracking uses ephemeral in-memory IDs, and only aggregate counters and short event clips reach the disk. Blurring of faces and licence plates is available. This is a research demonstrator, not surveillance or enforcement — no register of offences is created. Details on the Compliance / GDPR page.
Why does the video lag a few seconds and run at a low FPS?
The source is an HLS stream, which by design splits video into multi-second segments — hence the delay. The analysis runs on a CPU with a limited compute budget, so we process a few frames per second instead of the full 25. A smoothing queue keeps the preview fluid instead of a choppy "live at any cost" picture.
Are bicycles counted?
Yes — bicycles are a separate detection class with separate statistics, because their behaviour at a crossing differs from pedestrians and cars. Motorcycles are counted as vehicles.
Where do the cameras come from?
From public webcams made available by municipalities. The source of each camera is credited next to the preview. We do not install our own cameras and we do not use closed CCTV systems.
Can I run this at my own crossing?
Yes — the whole project is open source under Apache-2.0: github.com/AndriiShramko/bezpieczne-przejscia. You need a public camera stream and a small server — CPU is enough, no GPU. The method is portable: the AI describes the scene once (zones, lanes, directions), then the analysis runs automatically.
How much does it cost?
Almost nothing. Detection runs locally on a CPU, AI verdicts fit within the free daily tier, and hosting is a small VPS. The code is open — the only real cost is the server and your setup time.
What does the AI do, and what does YOLO do?
It is a cooperation of three "models". YOLOX is perception: it finds pedestrians and vehicles in every frame, locally and for free, but it doesn't "understand" the scene. Gemini is understanding: it describes the crossing topology once, then judges each episode from a frame sequence. Humans are the final check — they confirm or refute the verdicts, and true accuracy is computed from that.
Let's talk about your crossing
For a municipality, road authority, company or research project. I reply within 1–2 business days.