GÖZCÜ: smart highway accident detection system
A domestically developed AI platform that analyzes footage from existing highway cameras on site, detects accidents within seconds and mobilizes response teams instantly.
- 100%recall (test)
- <100 msper frame
- 6incident types
Sees the accident within seconds, dispatches the team
Smart highway accident detection: a domestically developed platform that detects accidents within seconds using AI and mobilizes response teams instantly.
Normal flowIncident, vehicle tracking
The problem
Accident reporting relies on calls and manual camera monitoring; critical minutes are lost before first response and the risk of secondary accidents grows.
Works with existing IP camerasMinimal additional infrastructure; on-site analysis with an edge device.
Recognizes 6 incident typesEach incident type is classified and prioritized separately.
AI-generated incident summaryTells the operator what happened and the recommended action instantly.
Zero missed accidents, high precision
Recall
No accident missed. Target: >85%Precision
Accuracy of alarms. 52 TP, 3 FPF1 score
Balanced performance metric. Target: >0.95Hard negatives
Correct rejection of accident-like scenes. 19/20 scenariosPer-frame processing
Real-time analysis. Target: <100 msEvaluation: end-to-end test on 51 labeled videos, 52 accident events and 5 accident types (ML threshold 0.80, 2-frame confirmation).
IP cameras
- Highways and junctions
- Existing CCTV infrastructure
- Day & night
Edge AI device
- NVIDIA Jetson / mini PC
- YOLOv8/v11 object detection
- ByteTrack vehicle tracking
- Optical flow, ML classifier
GÖZCÜ core
- FastAPI microservices
- LLM incident summary
- PostgreSQL + TimescaleDB
- Redis, MinIO (S3)
Control center
- Live map & alarms
- Analytics & reports
- Role-based access
- Real-time notifications
Processing at the edge: video is analyzed on site without being sent to the center; low bandwidth, high privacy and alarms within seconds.
Live KPIs
Active alarms, resolution rate, average response time.
Trend analysis
24-hour and weekly alarm volume.
Severity prioritization
Critical, high, medium and low levels.
Map & weather
Location-based incidents, live weather data.
Privacy & data protection
- Automatic face blurring
- License plate anonymization in non-alarm footage
- Automatic data deletion after 30 days
- User access logs (audit trail)
- Role-based authorization
Deployment model
- DiscoveryCamera inventory and critical locations
- Pilot10-camera field pilot, edge installation
- CalibrationSite-specific thresholds and operator training
- Scale-upNetwork-wide rollout
Target organizations
- General Directorate of Highways
- Motorway operators
- Metropolitan municipalities
- Insurance companies
- Organized industrial zones
- Smart city projects
Questions about GÖZCÜ
If you cannot find the answer here, send us your question and the relevant specialist will get back to you.
Ask a questionDoes GÖZCÜ work with our existing cameras?
Yes. GÖZCÜ takes the RTSP stream from existing IP / CCTV cameras on highways and junctions. Additional investment is limited to the edge device (NVIDIA Jetson or mini PC) that processes video on site.
Which incidents can it detect?
It recognizes 6 incident types: collision, rollover, wrong-way driving, pedestrian on the road, smoke / fire and stopped vehicle. Each is classified separately and prioritized by severity.
How accurate is detection?
In an end-to-end test on 51 labeled videos with 52 accident events, it achieved 100% recall (no accident missed), 94.5% precision and an F1 score of 0.972. Correct rejection of accident-like scenes was 95%, with under 100 ms of processing per frame.
How are personal data and privacy protected?
Video is analyzed on site without being sent to the center. Faces are blurred automatically, license plates are anonymized in non-alarm footage, data is deleted automatically after 30 days, and access is role-based with an audit trail.
Can it run on a network without internet access?
Yes. The AI incident summary sent to operators can be generated by a local (air-gapped) large language model or a cloud model.
How does deployment start?
Discovery (camera inventory and critical locations), a 10-camera field pilot, site-specific calibration and operator training, then network-wide scale-up. The first field pilot is planned for Q4 2026.
Other solutions
Kingstur
A global booking platform bringing flights, buses, hotels, tours and transfers under one roof.
IoT, Industry 4.0Smart Factory
Real-time data from injection molding machines, MRP and advanced reporting.
R&D, AIMoldIQ
Automated mold and injection design with machine learning.
Let's talk about GÖZCÜ
Tell us what you need, and we will get back to you with a solution architecture and proof of concept (PoC) tailored to your organization.