Project Overview
Our retail client wanted the same level of analytical insight in their physical stores that they already had online: who was walking in, where they went, what they stopped to look at, and how that behaviour translated into purchases. The catch was that everything had to run on the existing in-store hardware — standard cameras and modest compute — without shipping raw video off-site.
The Challenge
- Detect, count and track people entering, leaving and moving through the store in real time, on modest on-premise hardware.
- Classify broad demographic attributes like age range and gender, while preserving privacy and avoiding any individual identification.
- Stream video signals with minimal load so the system could coexist with existing store systems on constrained networks.
- Turn raw detections into a store-operator-friendly dashboard showing foot traffic, dwell times, heat maps and merchandising performance.
Our Approach
Edge-optimised detection
We chose and tuned detection models specifically for constrained in-store compute — dual-core CPU plus a modest entry-level GPU — making aggressive trade-offs between model size, resolution and throughput to hit real-time.
Lightweight tracking & anonymisation
A tracking layer re-identified individuals across frames only long enough to produce aggregate movement and dwell metrics, with no personally identifying data ever leaving the store.
Efficient video pipeline
Frames were sampled, pre-processed and encoded in a way that kept network and disk load low, so the solution fit comfortably alongside existing store systems.
Operator-facing dashboard
Detections were aggregated into a dashboard showing traffic by entrance, dwell time by aisle, product zone engagement and demographic mix — giving store managers the kind of analytics previously only available to e-commerce teams.
Technology Stack
The solution was engineered with a carefully chosen set of tools and frameworks, balancing maturity, performance and fit to the problem domain.
Results & Impact
+22% increase in sales
driven by data-informed store layout, merchandising and staffing decisions.
Real-time operational visibility
for store managers, with heat maps and traffic trends updating continuously.
Privacy-preserving architecture
that kept raw video and any identifying signal on-premise by design.
Scalable roll-out
across additional stores without requiring new hardware investments.
Conclusion
The interesting engineering in retail computer vision is rarely the model itself — it is the constraints: modest hardware, privacy, store operations that cannot tolerate downtime. By treating the constraints as first-class requirements, we delivered a system that physical retailers could actually deploy, and that turned everyday cameras into one of the most strategic assets in the business.