What's actually happening in your stores, measured from camera data

The Problem

Online retailers know everything about their visitors: where they came from, what they looked at, what they bought, what they almost bought. Physical retailers know how many transactions happened at the register. Everything else is guess work.

The gaps are big and expensive. Which entrances drive most traffic? Which aisles get walked vs ignored? Does the new end-cap promo actually pull visitors past the price-comparison shelf? Does an in-mall location near the food court convert better than one near the parking entrance? How does a 5pm rush differ from a 1pm lull?

The traditional alternatives, manual headcount, mystery shoppers, customer surveys, give you small samples at high cost. Mall operators sometimes have basic entry counters but no in-store behavior data. Brand-side marketing teams running co-marketing campaigns inside retailers have almost zero visibility into whether the campaign material is actually being seen.

What the Solution Does

A camera-based analytics layer that turns existing or newly installed cameras into a data source comparable to web analytics for a website. Five things the system tracks, configurable per deployment:

1. Entry / exit counts, how many visitors per hour, per day, per week, per location.
2. Pass-by vs entry, how many passers-by don’t enter (the equivalent of web “bounce” outside the front door).
3. Demographic segmentation, gender, age bucket, optionally other attributes.
4. In-store heatmaps, where people walked, where they lingered, which zones got ignored.
5. Conversion attribution, visit, then in-store path, then transaction, optionally linked to POS data.

Output is a dashboard tuned for the team that uses it, marketing for campaign attribution, ops for staffing decisions, real estate for location performance.

Where It Fits

This makes sense if you…

• Operate retail at 5+ locations and want comparative analytics across them
• Run marketing campaigns whose ROI depends on in-store behavior
• Allocate store layout, fixtures, or promo materials based on traffic data
• Operate a mall and need to demonstrate value to tenants beyond entry counters
• Have a brand-side relationship with retailers and want better visibility into your in-store presence

This is probably not the right time if you…

• Have a single store at small scale, the cost of CV exceeds the marginal value of analytics
• Operate in a jurisdiction with strict customer-imaging restrictions you haven’t validated (GDPR / regional privacy)
• Are not ready to invest in acting on the analytics, the data only becomes valuable when it changes decisions
• Need real-time identification of specific individuals (this is consciously not a face-recognition product)

Business Value

Marketing attribution. Online marketing has UTM tracking; physical retail has had nothing comparable. Foot-traffic counts segmented by entry, time, and demographic let you measure offline campaigns: flyer drops, billboard placements, sponsored mall events, in-store promo end-caps. You read them against a funnel that mirrors the one you measure for web traffic.

Layout optimization. Heatmaps surface what gets walked and what gets ignored. The expensive end-cap that nobody walks past is the expensive end-cap that gets a layout revisit. Conversion lift after layout changes informed by heatmap data is typically 5-15%, depending on how aggressive the changes are.

Staffing decisions. Foot-traffic patterns by hour and day make staffing schedules data-driven. A 5pm spike in a category that needs sales-rep interaction is the kind of pattern that justifies an extra shift on that aisle.

Location performance benchmarking. Chain-wide foot-traffic and conversion data lets head office compare locations on one shared metric. Stores get judged on conversion rate (transactions per visit), which separates the “low traffic but high conversion” location from the “high traffic but missing the close” location. Absolute revenue alone hides that difference.

How It Works

The pipeline below is what we proposed for a multi-store media operator running in-mall advertising attribution. It is also what we built for a multi-store furniture retailer running sales-rep and visitor interaction analytics.

1. Camera infrastructure

We don’t pretend any deployment “just uses existing cameras”. For meaningful analytics, you need:

• Entry cameras, fronts of stores, narrow field-of-view, demographic-classification angle. 1-2 per entry point.
• Sales-floor fisheye cameras, 12MP IP fisheye, mounted at ceiling, full-store coverage. Roughly 10-16 cameras per 1,000m² with overlap.
• Optional category cameras, narrow field-of-view on specific zones (premium displays, promo end-caps, checkout).

For new installations we spec everything during the test phase. Existing CCTV often works for entry counting but not for full-store heatmaps.

2. Edge processing, Jetson Orin

Each store gets one or two Jetson Orin units (Nano or AGX depending on camera count). The Jetson handles real-time video stitching across fisheye cameras (overlapping fields combine into a single panorama), preliminary motion detection, and bandwidth reduction before upload.

3. ML pipeline, five concurrent tasks

• Visitor and staff detection, two-class detector. Knowing which is which matters for interaction analytics (“how many visitors got a staff interaction?”).
• Entry analysis, separate model on the entry camera: detection, gender classifier, and age-bucket classifier.
• Unique-visitor counting, within-camera tracking, then cross-camera re-identification (Re-ID) to follow a visitor across the store. Built without face recognition, we use clothing, gait, and body-shape embeddings, which preserves privacy.
• Heatmaps, projection of detected trajectories onto your floorplan. Aggregatable by hour / day / week / month, filterable by demographic.
• Optional: POS / transaction linkage, connecting visit trajectories to register transactions to compute true conversion. This is a separate integration with your POS system.

4. Analytics layer

Output flows into your existing BI tool (Metabase, PowerBI, Looker) or a custom dashboard. Standard widgets: entry/exit count, conversion funnel, demographic breakdown, heatmap overlay on floorplan, comparative views across locations.

Stack

YOLO-family for detection and tracking, ResNet-class encoders for demographic classification, custom Re-ID models, video stitching on Jetson Orin (TensorRT optimized), Datapipe for the data pipeline into the warehouse, Metabase / PowerBI / Looker for dashboards. Backend: CPU and GPU (V100-class) cloud or on-prem.

What You Need to Make This Work

Data. Store floorplans for heatmap projection. Access to existing video infrastructure or budget for new camera installation. POS data export (optional, for conversion attribution).

Integrations. Camera management (RTSP feeds or API). BI tool of your choice. Optionally, POS system for the conversion-attribution layer.

Hardware. Cameras (we spec during test). Jetson Orin per store (1 per 10 cameras). Cloud GPU for centralized processing OR on-prem GPU server for stores that require local-only data.

Team. A marketing / category lead who knows what questions they want answered. A store-ops contact for installation logistics. IT contact for network access. POS integration (if included) typically requires a developer (~20-30 hours).

Implementation Roadmap

1. Test (4-6 weeks)

Single store. Spec camera coverage; install or repurpose cameras; deploy Jetson Orin; build the first version of the detection pipeline; validate counts against manual ground truth (turnstile, manual counters, or staff observation). Output: a written report on detection accuracy, demographic accuracy, heatmap quality, with recommendations for the multi-store rollout.

2. Pilot (2-3 months)

Deploy across 3-5 stores with format / location diversity. Build dashboards tuned for the consuming team (marketing, ops, real estate). Integrate POS if in scope. Output: working dashboards in use, documented business outcomes (campaigns measured, layout changes proposed, staffing schedule adjusted), go/no-go on full rollout.

3. Scale (4-9 months)

Add stores progressively. Hardware installation is the rate limit at this stage. Ongoing model retraining as new store formats come online. By the end your team owns the analytics; we stay on for retraining and edge cases.

Keep in Mind

Where it breaks, and one important non-feature:

• This is consciously not a face-recognition system. The system tracks visitors anonymously using clothing and body-shape embeddings. It never reads faces. It cannot identify specific individuals, and we recommend against architectures that could. This is a privacy posture and a regulatory one: most jurisdictions either restrict or prohibit face-recognition retail analytics.
• Demographic classification has known limits. Gender classification reaches 90%+ in typical conditions. Age-bucket classification (under-25 / 25-45 / 45-65 / 65+) reaches 75-85%. Finer age estimation is less accurate. We report the measured numbers during pilot.
• Heatmap quality depends on camera coverage. A store with bad coverage (single camera, narrow angle) produces a heatmap that looks coarse and isn’t trustworthy. Either invest in coverage or scope the analytics to the zones cameras actually see.
• “Why” is your job. The system tells you what happened; you supply the cause. It tells you that traffic dropped 15% on Tuesday afternoon. It can’t tell you that a competitor opened a flash sale next door. The analytics gives you data. Your team interprets it.
• POS integration is a separate project. Linking visits to transactions requires POS access, timing reconciliation, and a privacy review. We can do it; it adds 4-8 weeks to scope.
• Privacy compliance is jurisdictional. GDPR, CCPA, and regional rules all take different stances on retail video analytics. We work through compliance during the test phase and can deploy in modes ranging from fully cloud to fully on-prem-anonymized.

FAQ

Does this use face recognition?

No. The system tracks anonymous visitors using body shape, clothing, and gait. It never reads faces. Cross-camera tracking (Re-ID) uses clothing and body embeddings. This is a deliberate privacy choice and a regulatory one in most jurisdictions.

How accurate is the demographic classification?

Gender: 90%+ in typical conditions. Age bucket (under-25 / 25-45 / 45-65 / 65+): 75-85%. We surface the measured numbers per project during pilot.

Can we integrate with our POS to link visits to purchases?

Yes, this is the conversion-attribution layer. POS integration adds 4-8 weeks to scope and requires your POS vendor’s cooperation. We’ve integrated with most common POS systems.

How does this compare to mall-operator turnstile counters?

Turnstile counters give you entry counts. The CV system gives you entry counts, pass-by counts, demographic segmentation, and in-store path. If you already have turnstile counts and they’re working, the value-add is everything that happens after entry.

Can we use this for queue management at checkout?

Yes, that’s a related use case (smart queue analytics). It runs on one shared camera setup with a different focus.

What about privacy regulations?

We work through compliance during the test phase. The architecture is designed to support the most restrictive jurisdictions: on-prem anonymized processing, no individual identification, configurable retention. We discuss requirements specific to your locations.

Ready to Discuss?

If you operate retail at scale and the gap between online-style analytics and your physical-store data is a real strategic problem, this is the right conversation. We’ll talk through your current locations, your camera infrastructure, and the questions you most want answered. Then we’ll tell you what a single-store pilot would cost and what to expect from it.

Discuss your project