Computer-vision warehouse access control
Replaced a manual gatehouse process that bottlenecked at peak hours — peak-hour gate cycle time went from minutes per truck to seconds, gatehouse staff redeployed, every entry auditable.
Context
Peak hours saw queues of inbound trucks waiting for the gatehouse to manually verify each driver and load. Off-the-shelf ALPR products solved part of the problem but were closed black boxes with per-camera licensing and no driver-identification capability — and they couldn't talk to the existing yard-management software without expensive integration work.
Problem
The real requirement wasn't just plate recognition — it was 'this plate, this driver, this expected load, this gate' all matched up against the dispatch record before the boom opens. No vendor sold that. Custom integration with a closed product would have cost the same as building it; building it gave us ownership.
Approach
The pipeline runs at each gate on commodity edge hardware. A first-stage detector finds vehicle and driver regions; downstream models do plate OCR, vehicle-type classification, and driver identification. Confidence thresholds are tuned per stage so failures fall back to the gatehouse rather than to bad decisions. Integration with yard-management is a small REST contract — the existing dispatcher UI didn't change.
Outcome
Peak-hour gate cycle time went from a multi-minute manual process to seconds-per-truck for the verified path. Gatehouse staff were redeployed to higher-value yard coordination. Auditability — every entry has a stored record with the input frames — turned out to be a side benefit nobody had asked for but everyone now uses.
What I learned
Edge over cloud was the right call: latency requirements at the gate (open the boom in under a second) are not compatible with a round-trip to a regional cloud. Owning the model also meant we could fix recognition issues with new training data within a week instead of opening a vendor ticket.