Skip to content
See the work
All work
Urban Edge Vision Analytics: cover image

Edge inference

Active

Urban Edge Vision Analytics

One edge box watching several street cameras, turning what a vision-language model sees into traffic events an operator reviews. Runs end to end on live cameras.

Read the case study View the code on GitHub

Measured results

  • Live camerasRunning end to endThor and RTX 5090 hosts
  • Use-case packs4 per cameraStop sign, speed between gates, vehicle count, moving object
  • Latency, power, pack accuracyNot yet measuredField validation in progress
How the system fits together
  1. PhysicalStreet cameras
  2. Edge computeCapture at the camera's native rate
  3. Edge computeVision-language model on its own cadence
  4. Edge computeTracker and use-case packs
  5. PhysicalOperator review queue

Hardware

A Jetson AGX Thor runs Cosmos-Reason2-2B on vLLM at the edge. An RTX 5090 workstation runs Gemma 4 on Ollama or vLLM. Cameras connect over RTSP, HTTP MJPEG or a webcam, with ten vendor profiles.

Workload

Each camera carries one or more packs drawn by the operator over a live snapshot: stop-sign behaviour, speed between two gates, vehicle counts by direction and class, and moving-object tracking. Speed and stop sign are refused on the same camera, because they need different sight lines.

How it runs

Video never waits on the model. One capture thread per camera keeps the newest frame; the browser shows it at the camera’s own rate, and the model samples it on its own cadence. Events persist with the frame the model saw, and flagged ones wait for an operator. Nothing is enforced automatically.

What is not yet shown

It runs end to end on live cameras, but latency, power and pack accuracy have not been measured yet. Those numbers will be published here when they exist.