Heathrow T5 Computer Vision System
A real-time computer vision system for object detection and people-flow analysis deployed in Heathrow Terminal 5. Designed to support operational decision-making and safety monitoring in a high-throughput airport environment.
Python · PyTorch · TensorRT · RTSP · OpenCV · CUDA
Overview
This engagement involved designing and deploying a computer vision system within Heathrow Terminal 5 for The Network Group. The system provides real-time object detection and people-flow analytics across a distributed camera network, feeding into an operational dashboard used by terminal supervisors to monitor capacity, identify bottlenecks, and respond to safety events.
Given the constraints of the environment — high-throughput, regulated, with strict uptime requirements — the system design prioritised reliability and explainability over raw benchmark performance. Every detection inference is associated with a confidence score and a timestamp, allowing downstream analysts to audit system outputs and understand the basis for any flag.
Technical Approach
The detection pipeline is built on a YOLOv8 backbone fine-tuned on a proprietary dataset of terminal-specific images, covering the full range of lighting conditions, camera angles, and crowd densities encountered in the deployment environment. Inference runs on edge hardware with CUDA acceleration via TensorRT, achieving sub-30ms latency per frame at 1080p.
People-flow analysis is handled by a separate tracking module that associates detections across frames using a Kalman filter and Hungarian algorithm assignment. The tracker maintains trajectory histories for active entities, allowing computation of flow rates, dwell times, and queue length estimates for defined spatial zones.
Deployment and Operations
The system was deployed in a phased rollout across 12 camera zones over a three-month period. Each phase included a calibration step — aligning the model’s spatial coordinates with the physical terminal map — and an evaluation period during which operational staff validated outputs against ground-truth observations.
Post-deployment, false-positive rates on safety events (unattended baggage, access-zone violations) were held below 2%, a threshold agreed with the operations team as acceptable for the alert workflow. The system has been in continuous production operation since deployment.