RoadSight
Finds traffic events in fixed-camera CCTV video and scores accident risk frame by frame. Fully offline, on one GPU.
Give RoadSight one road-camera MP4. It returns every event as [start, end, label],
from stopped vehicles and congestion to wrong-way driving, jaywalking and collisions. It also emits a risk score
for every frame that uses only the frames seen so far.
Problem → Solution
The problem
Cities run many CCTV cameras at junctions, but operators can watch only a small part of the footage. Red-light running, cars that do not yield to pedestrians, vehicles stopped in a lane and near misses are noticed late or never, and reviewing one hour of video by hand costs about an hour of an operator's time.
Our solution
RoadSight watches the camera for you. For every video it returns each traffic event with its type, start and end time (14 event types, from red-light running and jaywalking to collisions), plus a risk score for every frame that rises before dangerous situations. It runs fully offline on one computer, so no footage leaves the city's network.
Why our team can solve it
- It already works on real footage. Our prototype analyses 4K video of a Tashkent intersection and finds red-light runs, stop-line violations, stopped vehicles and drivers not yielding at crossings.
- We measure honestly. We checked our results frame by frame against the event definitions, tested on real crash videos and publish our failure cases next to our successes.
- We cover the whole product. Computer vision, traffic rules, the risk model, the website and the live demo were all built by the team, and the system runs inside the time limits on ordinary hardware.
- We know the local roads. We build for the junctions and traffic we see every day, including dense queues, buses and pedestrians crossing between cars.
Roadmap
- done
Idea
Automatic traffic-event detection for fixed city cameras.
- we are here
Prototype
Working system online: 14 event types, risk score, live demo with upload, tested on real intersection and crash videos.
- next
MVP
Pilot on 2–3 real city cameras with operators: live stream input, an operator dashboard, alerts, and a detector fine-tuned on local footage.
- later
Launched
A service for a city traffic centre: many cameras, reports per junction and per hour, and integration with existing monitoring systems.
How we will build it
| Stage | What we do | Technologies and AI tools |
|---|---|---|
| 1. Prototype (done) | Detect and track road users, turn trajectories into 14 traffic rules, causal risk score, website and live demo | YOLO11 (object detection), ByteTrack (tracking), OpenCV, ffmpeg, Python, PyTorch, FastAPI, nginx |
| 2. Better accuracy | Label our own videos, fine-tune the detector on local cameras, tune each rule on labelled data | Fine-tuned YOLO on UA-DETRAC and our labels, our evaluation tools |
| 3. Accident anticipation | Replace the hand-made risk score with a learned model that predicts a crash seconds ahead | Video and trajectory models trained on public crash datasets (DoTA, CCD, ACCIDENT) |
| 4. MVP pilot | Process live camera streams, operator dashboard with alerts and statistics | RTSP stream input, GPU server, web dashboard, notifications |
| 5. Launch | Many cameras, reports per junction and per hour, integration with the city's systems | Docker, multi-GPU deployment, REST API |
During development we also use AI coding assistants for writing and reviewing code; every model that runs in RoadSight is open and runs on our own hardware, with no paid cloud AI.
How a video flows through it
Green boxes are learned; everything else is explicit, tunable rules. Read the approach.
What it measures on the sample videos
Loading sample results.
See it on real footage
Annotated sample videos with boxes, track IDs, active events and the risk curve. Open the results.
Run your own clip
Upload an MP4 of up to two minutes and get the event table, timeline and annotated playback. Open the live demo.
What worked and what did not
The one-page technical report, with measured runtimes and ablations. Read the report.