Most people counting projects that disappoint were not let down by the camera. They were let down by a decision made before the camera was chosen: the buyer compared datasheets instead of describing the entrance. Two stores on the same street can need different devices because one has a wide glass storefront with a corridor behind it and the other has a narrow single door with a ceiling two metres above head height.
This checklist walks through the order in which the decisions should be made. Work top to bottom; skipping to the product comparison is the most common mistake.
Step 1: Define the question before the device
Write down what the number will be used for. The device follows from the purpose:
- Staffing and opening hours. You need reliable total traffic by hour and day.
- Conversion rate. You need traffic joined to POS transactions, which means clean entrance counts and a stable reporting period.
- Marketing and promotion evaluation. You need to compare periods and stores on the same basis, plus enough history to see a baseline.
- In-store experience. You need zone-level data: where people go, how long they stay.
- Occupancy limits. You need a live count, not a daily total, and you need it to be trustworthy in a crowd.
- Chain reporting. You need consistent definitions across sites and an export or API.
A project that only needs daily totals can succeed with a simple entrance counter. A project that needs zone dwell and chain-level comparison needs a different class of device. Deciding this first prevents two of the most expensive mistakes: paying for analytics nobody uses, and discovering the analytics you need are not supported.
Step 2: Describe every entrance honestly
Walk the site with a tape measure and a notebook. For each entrance, record:
- Clear opening width. A wide mall storefront and a single door are different engineering problems.
- Mounting height and available mounting points. Ceiling height above the counting plane, and whether a ceiling mount, wall bracket, or pole is possible.
- What the ceiling is made of and how you will get cable there. Practical access decides more installations than specifications do.
- Cross-traffic. Does a corridor, escalator, or adjacent shop funnel people past the door without entering? This is the single most common source of inflated counts.
- Adjacent entries and exits. Fire doors, side doors, staff entrances, and connecting internal doors. Each needs its own decision: count, ignore, or net out.
- Lighting through the day. Direct sunlight through glass creates hard shadows and blown highlights. Evening lighting changes the scene again.
- Floor surface and reflections. Polished stone floors create reflections that confuse some sensors.
- Traffic character. Groups walking abreast, parents with strollers, wheelchairs, delivery staff, children. Note what applies.
- Peak conditions. The counting problem is hardest at the busiest five minutes of the week, not on an average Tuesday.
This record becomes the basis for the acceptance test later. Without it, "accuracy" is a number nobody can verify.
Step 3: Match the technology to the conditions
The three main families behave differently in the situations you just recorded.
2D AI cameras detect people in a single image and track them across frames. They cover wide entrances with one device, reuse existing network cabling, and can also produce zone-level data. They are the practical default for most retail entrances. Their weakness is depth: in dense crowds, with occlusion, or when people carry large objects, inference from a flat image becomes ambiguous.
3D stereo (binocular) cameras calculate distance from two viewpoints and produce a depth map. That extra dimension makes counting more stable in dense traffic and improves tracking when people overlap. The trade-off is cost, calibration discipline, and a defined working depth range.
Thermal and fisheye devices solve specific conditions: thermal sensing handles difficult lighting and can help where visible-light detection struggles, while fisheye panoramic cameras cover a large area or several zones from one mounting point. They are choices driven by scene geometry, not general upgrades.
If your entrance is narrow, low-traffic, and single-file, simpler and cheaper technologies may genuinely be adequate. Over-specifying is as common a failure as under-specifying.
Step 4: Read the spec sheet for the right things
Datasheets are written to be compared, which is why they emphasize numbers that matter less. Prioritise:
- Field of view and the mounting height range it supports. The pair together determines whether the device can see your entrance properly.
- Minimum and maximum working distance or depth range. Beyond the range, stereo performance degrades toward 2D behaviour.
- Lighting and operating conditions. Rated operating temperature and low-light behaviour matter in unconditioned entrances.
- Network and power. PoE support, whether the device can buffer results when the network drops, and what happens on reconnection.
- Storage and buffering. Local storage capacity, and how long data survives an outage.
- Interfaces and protocol support. ONVIF, RTSP, API availability, and whether outputs can be pushed to your platform.
- Mounting and enclosure. Whether an indoor-rated device will survive a covered outdoor entrance.
Treat resolution as one input among many rather than the headline. Beyond a certain point, more pixels do not fix a bad angle, and every vendor's accuracy is dominated by installation quality.
Step 5: Demand a verifiable accuracy test
Ask for the accuracy claim in writing, then define how you will prove it. A workable acceptance test looks like this:
- Pick the busiest entrance, not the quietest.
- Record the camera's count for a defined period.
- Count by hand, with two people if possible, for the same period.
- Compare, then repeat at the busiest hour and at night.
- Ask the vendor to state how errors are expected to change with crowd density and lighting.
If the results fall short, the cause is often fixable — mounting height, angle, the counting line position, or which door is being netted against which. Give the integrator one tuning cycle before judging the hardware. If a vendor refuses to participate in any verification, treat the accuracy figure as unproven.
Step 6: Settle the data and integration questions
Counting data is only useful where people can see it. Before purchase, confirm:
- Where data lives. On the device, on a local server, or in a vendor cloud. This affects both privacy posture and recurring cost.
- Data ownership and export. Can you export raw events, and in what format?
- APIs. Is there a documented API, and can you read it before signing?
- Multi-site management. Can a chain configure and monitor all sites centrally?
- POS or BI integration. Whether conversion-rate reporting is possible without manual spreadsheet work.
- Retention and privacy configuration. What is stored, for how long, and how access is controlled.
Step 7: Understand the cost drivers
Purchase price is one component of the total. The drivers that typically move the total most:
- Technology class. Stereo and thermal devices cost more than 2D cameras, and the gap is real rather than marketing.
- Number of devices and entrances. Counting every entrance is more expensive than counting the main one.
- Installation difficulty. Cable routes, conduit, lifts, and working outside opening hours all add labour.
- Analytics licensing. Some platforms charge per camera, per site, or per analytics module.
- Integration work. POS connectors, BI dashboards, and custom API work are project costs, not product costs.
- Support and warranty terms. Service level, spares, and firmware support over the product's life.
Ask for a total cost over three years with these lines separated. Unit price comparisons without them are not comparable.
The one-page checklist
- Purpose and required analytics written down
- Every entrance surveyed with dimensions and conditions
- Cross-traffic and secondary doors accounted for
- Technology class chosen to fit the conditions, not the budget alone
- Mounting, power, and cabling confirmed as feasible
- Accuracy claim documented and an acceptance test scheduled
- Data location, export, and API confirmed before purchase
- Retention and privacy configuration scoped
- Three-year total cost with components separated
- Support, warranty, and firmware policy in writing
FAQ
How do I know if I need 2D or 3D?
If entrance traffic is dense, people frequently walk abreast, or occlusion is common, stereo is worth the cost. For moderate traffic through a defined entrance with good mounting geometry, a well-installed 2D AI camera can be entirely adequate. The deciding factor is crowd behaviour at your busiest moment.
Should I count every entrance or just the main one?
Counting only the main door inflates conversion rate if customers use side doors or a connected mall corridor. Either count all customer entrances, or explicitly document which traffic is excluded so the metric stays comparable over time.
Is a higher-resolution camera automatically more accurate?
No. Accuracy is dominated by mounting geometry, lighting, crowd density, and how well the model handles your specific scene. Resolution helps with detail at distance, but it cannot correct a bad angle or heavy occlusion.
Can one camera cover an entrance and a zone?
Sometimes, if the field of view covers both with an acceptable angle for each. Performance in each role is usually slightly worse than with a dedicated device, so weigh the saving against the data quality you need.
What should I test before a full rollout?
One site, two weeks, a manual count comparison at peak and off-peak, and a check of the reporting your team will actually use. Most problems surface in that window rather than in the first three days.
---
The product center lists the camera families HooViz builds across the 2D, stereo, and panoramic categories, and the blog covers comparisons between them. If you are sourcing for a chain or an integration project, contact our team with your entrance survey and we can recommend a configuration.