A food-service operator was running on instinct — counting heads by eye, guessing peak hours, and reading table turnover from memory. Axlume turned the cameras already on the ceiling into a computer-vision system that counts customers, separates staff from guests, and tracks every table’s occupancy in real time.
Our client runs a busy food-service floor and needed it to stop being a guessing game. Owners couldn’t see real peak hours, managers couldn’t read crowd conditions without walking the floor, and analysts had nothing but a recording — hours of CCTV that captured everything and told them nothing.
Axlume built an AI layer on the cameras already installed: a vision pipeline that counts customers, separates staff from guests, and tracks every table’s state in real time — turning passive surveillance into live operational intelligence for three roles: the owner watching revenue, the manager watching the floor, and the analyst watching the trend.
Headcounts were done by eye and were wrong by lunchtime. Nobody could continuously watch every area of the floor, so peak rushes were spotted late and staffing during them was mostly guesswork.
There was no automated read on occupancy — which tables were full, how long they’d been full, how long they’d sat empty. Seating went under-used, traffic patterns went unmeasured, and the only “analytics” the CCTV produced was footage nobody had time to watch.
Every decision that mattered — when to add staff, how to lay out seating, when the rush really hits — rested on someone’s memory of a busy night.
The vision layer running on a live feed — tracking each customer and waiter by ID, boxing every table as occupied or empty, and keeping a running count on screen. This footage is the proof the rest of the platform is built on.
The footage was already being recorded. The job was to extract intelligence from it — a lightweight, scalable pipeline that runs on a real restaurant feed, not a lab clip.
We spoke to the people who feel the gaps — owners wanting automated occupancy, managers wanting faster visibility into crowding, analysts wanting trends instead of tape. The recurring theme: CCTV that only records, never reports.
We scoped a system around four outcomes: automate monitoring, extract business insight from the feed, cut manual workload, and put the numbers where a decision gets made — on a screen, in real time.
Frames are pulled from the stream; a detection model locates every person and classifies customer vs. staff; a tracker holds each identity across frames; and an occupancy layer reconciles people against tables to mark each one occupied or empty, with the clock running on both.
Built as a modular pipeline — detection, tracking, occupancy and analytics as separate stages — so it stays light enough for a real venue and scales from one camera to many without a rewrite.
Shipped as a working analytics layer: live counts and table states on the feed, plus the aggregated metrics that turn a night of service into a chart an owner can actually use.
Every person on the floor located per frame and held as a stable identity across frames — the foundation everything else counts on.
People classified as customers or waiters, so the live count reflects actual demand — not staff walking the floor — and staffing can be read against it.
Each table tracked as occupied or empty with the elapsed time on both — turnover, dwell time and idle seats become measurable instead of remembered.
Movement aggregated into traffic patterns and density over time — where guests cluster, which areas sit cold, how flow shifts through a service.
The live read for managers and the aggregated read for owners — peak hours, occupancy curves and utilization, surfaced as visuals rather than raw footage.
Historical performance and trend analysis generated without anyone scrubbing tape — the analyst’s view, produced on its own.
Pulled straight from the live feed — not a staged demo. Crowded floor, real lighting, real angles, the model holding identities and table states throughout.
Four stages, decoupled so each can scale or be swapped on its own — the feed comes in, the AI makes sense of it, the analytics layer turns it into numbers, and the output layer puts those numbers in front of a person.
The floor went from observed to measured. Manual monitoring dropped, the rush stopped being a surprise, and the decisions that used to ride on memory — staffing, seating, timing — got a live read behind them.
What changed operationally:
Detection accuracy and deployment metrics are held under NDA. Contact us directly for a detailed brief.
Real computer vision in a real venue is the hard part — detection and tracking that hold up under crowding, fisheye lenses, mixed lighting and constant motion, not a clean benchmark clip. Holding a stable identity for a customer who keeps disappearing behind a counter is the whole game.
It was built hardware-light on purpose: it runs on the cameras the venue already has, so the analytics are a software deployment, not a refit. A modular pipeline — OpenCV and a YOLO-based detector at the core, with the tracking, occupancy and analytics stages decoupled — keeps it light enough for one venue and scalable to many.
The stack is deliberately pragmatic: Python, OpenCV, YOLO-based detection, deep-learning models, NumPy and Pandas for the analytics. The result behaves correctly on a live restaurant floor — not just in a notebook.
You’ve seen our standard of engineering. Now, let’s build yours. From high-impact websites to complex enterprise applications, we specialize in transforming ambitious ideas into elite-grade digital reality. Whatever the platform, we bring the precision your vision demands.