ZenariStart a conversation

← All case studies

Aviation — 2025

Seeing boarding bottlenecks before the staff do

A real-time computer-vision pipeline that detects passenger-flow bottlenecks at boarding gates and alerts staff early enough to protect on-time performance.

Role
AI lead — architecture and delivery
Duration
5 months
Outcome
Working pipeline on commodity CPU hardware; deployed to pilot gates

Stack

YOLO pose estimationByteTrackPythonWebSocketsDockerCamera calibration

The problem

A delayed departure usually starts as a queue nobody noticed. Gate staff are busy with the desk; by the time congestion is visible from where they stand, the aircraft is already going to miss its slot.

The constraint that shaped everything

No GPUs at the gate. The system had to run on the consumer-grade hardware already installed, which meant taking a 30 FPS camera feed and doing useful work at 3–5 FPS on CPU — and being honest about what that costs in accuracy.

How it works

Independent Docker services connected by WebSocket streams, so any stage can be replaced or scaled without touching the others:

  1. Calibration maps camera pixels to floor coordinates, so “crowded” means a density in square metres rather than a number of boxes in a frame.
  2. Pose detection rather than plain bounding boxes, because orientation distinguishes a queue from a group standing around.
  3. ByteTrack maintains identities across frames, which is what turns detections into dwell time and flow rate.
  4. Metric extraction produces queue length, throughput and dwell time per zone.
  5. Alerting fires on trend, not threshold — the useful signal is a queue that is growing, several minutes before it is a problem.

What it produced

A pipeline that holds real-time performance on the hardware that already exists at the gate, and flow metrics specific enough for staff to act on: which zone, how long, getting better or worse.