Computer Vision · AI Research
TrafficVision
Computer Vision · AI Research
A computer-vision platform that ingests traffic video and turns it into structured, real-time intelligence about vehicles and road conditions — built from the detection model up on self-owned infrastructure.
- 9
- Technologies
- 2
- Domains
- 5
- Capabilities
- 1
- Public links
Who it's for
Traffic authorities, researchers and smart-city teams who need structured intelligence from raw road video — counts, congestion, lane-level behaviour.
Where it fits
Research and evaluation settings today; architected for RTSP/CCTV ingestion and multi-GPU scale as the detection model matures.
The hard part
The problem
Raw traffic video is just pixels. Turning it into decisions — counts, congestion, signal optimisation — needs both a detector and a model of the physical road network that gives each detection spatial meaning.
Architecture overview
The platform models the physical network (city → intersection → road → lane) with camera placements, regions-of-interest and virtual counting lines. A self-trained FCOS detector, exported to ONNX, localises vehicles per frame; a Celery-driven GPU engine runs video off the request path with GPU allocation, heartbeats and automatic crash recovery. Detections become typed, confidence-scored records in a canonical contract.
Nested spatial model
City
Intersection
Road
Lane · ROI · counting line
How it works
What's actually in it.
- 01Traffic-scene modelling: city → intersection → road → lane, with ROIs and virtual counting lines.
- 02Self-trained FCOS detector exported to ONNX for fast inference.
- 03GPU-managed Celery processing sessions with allocation, heartbeats and crash recovery.
- 04End-to-end training pipeline: leakage-safe splits, interruptible multi-day training with checkpoint/resume, honest evaluation (IoU precision/recall, mAP).
- 05Structured, validated, confidence-scored output in a canonical contract.
Engineering decisions
Train-equals-serve preprocessing
A model is only as good as the consistency between how it was trained and how it runs in production. TrafficVision uses identical preprocessing on both sides so results stay consistent from training to serving — and exports to ONNX for fast, portable inference.
The trade-off
Training a detector from first principles on self-owned hardware is far slower to a first result than calling a hosted vision API — but it keeps the model, the data and the infrastructure fully owned.
Technical highlights
- Detection model trained and evaluated on self-owned infrastructure.
- Modular-monolith architecture ready for RTSP/CCTV ingestion and multi-GPU.
Claims & evidence
Detection model is in training; the platform is an active independent research project, not a production deployment.
Qualitative · TrafficVision status
Every statement above carries its evidence class — nothing is invented.
Tech stack
- Python
- PyTorch
- FCOS
- ONNX Runtime
- Django REST Framework
- Celery
- Redis
- PostgreSQL
- Next.js
Want to go deeper on TrafficVision?
I'm happy to walk through the architecture and the decisions behind it.