Edge vision for people flow and licence-plate recognition
Privacy-preserving edge models deployed to street-level cameras across a metro district – cleaning raw footage into people-flow analytics and ANPR events on-device.
Challenge
A metro authority in Kuala Lumpur wanted real-time insight into pedestrian flow, dwell time, and vehicle movements across 140 street-level cameras – but the local data-protection regulator had made it clear that no raw footage could leave the device.
Their previous approach shipped full video streams to a central cluster for processing, which both breached the privacy boundary and buckled under bandwidth cost.
Approach
We built a lightweight edge pipeline that runs on the existing camera hardware. On-device models extract only the signals the authority actually needs – anonymised person-flow counts, dwell heatmaps, and validated licence-plate events – and publish them to a lightweight ingest endpoint.
Raw frames are pre-processed and cleaned on the camera (exposure correction, deblurring, rejection of low-confidence frames) before inference, so the downstream signal is clean without touching a central GPU.
A managed retraining loop lets operators flag false positives from the dashboard; those examples are queued for the next edge-model release, which is pushed OTA.
Outcome
Zero raw footage now leaves any device in the network, passing the regulator’s independent audit.
Licence-plate recognition holds at 97.8% precision on-device, and the authority cut monthly bandwidth costs by 86% versus the previous central-processing approach.
The platform is now powering the city’s live congestion signalling during event days and peak commutes.
Let's build what's next
Share your challenge – AI, data, or infrastructure. We'll scope your project and put the right team on it.