How Edge-Optimized Deep Learning Is Reducing Video Analytics Costs for a Retail Chain

The Challenge

The retailer's first video analytics pilot sent camera streams from a handful of stores to cloud GPUs. It worked technically but did not scale financially.

The existing approach made it difficult to:

• Expand the pilot to every store without a steep rise in bandwidth and cloud compute bills.
• Get queue and stock alerts to store staff quickly enough to act on them.
• Keep analytics running when a store's internet connection dropped.
• Limit how much identifiable video left the store, in line with the company's privacy policy.
• Update models across hundreds of store devices without site visits.
• Compare footfall, dwell time, and conversion across stores on a like-for-like basis.

From a technical perspective, the stores ran a mix of camera models, resolutions, and mounting angles, and lighting changed through the day. Any deep learning model had to fit on low-power hardware in a back-office rack and process several streams at once in real time. It also had to be updated remotely across hundreds of devices.

The Solution

InnovationM developed an edge-first video analytics system in which inference happens inside each store and the cloud handles only aggregation, reporting, and model management. The solution included:

• Store-specific data collection: Building a labeled dataset from sample footage across store formats, camera types, and lighting, with faces and identifying details blurred before annotation through data engineering.
• Compact detection & tracking models: Training lightweight object detection and multi-object tracking networks for people counting, queue length, and shelf-gap detection.
• Model compression: Applying knowledge distillation, INT8 quantization, and TensorRT optimization so several camera streams run on each edge device.
• Edge inference runtime: Running containerized models on in-store devices that process video locally and discard raw frames after analysis.
• Event-only uplink: Sending only counts, timestamps, and alert events to the cloud, with local buffering when the store connection is down.
• Central dashboard & alerts: Showing footfall, dwell time, and queue trends by store and hour, plus mobile alerts to store managers when a queue crosses an agreed threshold.
• Fleet model management: Delivering over-the-air updates, version tracking, and staged rollouts, with accuracy checks against privacy-filtered sample clips.

The key focus was not simply moving existing models closer to the cameras, but creating models that balance accuracy, speed, and hardware cost for the store environment.

The Impact

The implementation helped the retailer move from a costly cloud pilot in a few stores toward an edge-based system ready for the full estate. Operations and store teams now get insights in near real time. The solution supported:

• Lower bandwidth use per store.
• Reduced cloud compute costs for video analytics.
• Faster queue alerts to store staff.
• Analytics that continue through internet outages and sync once connectivity returns.
• Far less identifiable video leaving each store.
• A foundation for chain-wide, AI-driven store operations.

By bringing together compressed deep learning models, local inference, centralized data engineering, and fleet management, InnovationM showed how deep learning development services can turn existing cameras into a practical operations tool rather than an expensive experiment.