In most shops the CCTV system does one thing: it records, in case of theft. AI video analytics uses cameras to measure how the store works. It can count people entering, show how many are inside at once, time how long queues are, and map which areas customers visit and which they skip. Those figures answer questions a till report cannot: how many visitors left without buying, whether Saturday staffing matches Saturday traffic, whether the new display is being noticed. The technology is mature for counting and queues, less dependable for anything involving individual behaviour, and in Singapore it must be set up so that it measures crowds without identifying people.
What can it measure?
| Measure | What it tells a retailer | Reliability |
|---|---|---|
| Footfall at the entrance | Visitors by hour and day; conversion rate when combined with sales | High, with a camera placed for counting |
| Occupancy | How many people are inside now | High |
| Queue length and wait time | When to open another till | Good |
| Heatmaps | Which zones attract or lose visitors | Good as a trend, not a precise count |
| Dwell time at a display | Whether a fixture holds attention | Moderate |
| Path through the store | Whether the layout works | Moderate; harder in crowds |
| After-hours intrusion, back-door activity | Security alerts | High |
| Staff presence at a counter | Coverage of service points | Good |
Reliability depends heavily on camera position, lighting and crowding, and differs between products. Treat vendor accuracy claims as best-case and test in your own store.
Why does footfall matter so much?
Because sales alone hide the cause. CALC. A shop takes 120 transactions on Saturday and 90 on Sunday. Counting shows 800 visitors on Saturday and 450 on Sunday. Conversion is 120 / 800 = 15 percent on Saturday and 90 / 450 = 20 percent on Sunday. Saturday looked like the better day; it is the day more visitors left empty-handed, which points to queues, staffing or stock, not marketing. Raising Saturday's conversion to Sunday's 20 percent would mean 800 x 0.20 = 160 transactions, 40 more, from the same visitors. The figures are an example to show the method.
How does it work?
- Cameras view the entrance, queue and shop floor. Counting works best from a camera mounted overhead, looking down.
- A vision model detects and tracks people as shapes, frame by frame.
- The system turns tracks into numbers: crossings of a line, people in a zone, time in a zone.
- Numbers go to a dashboard, and to alerts such as a queue exceeding five people.
The analysis can run in three places:
| Where | Suits |
|---|---|
| In the camera | Line-crossing counts and zone rules on one camera at a time |
| In the recorder or NAS | Counting and reports across several cameras, where the model supports it |
| On an edge AI computer in the store | Heatmaps, paths, queues and custom measures across many cameras |
Processing in the store, not in the cloud, avoids uploading video continuously and keeps footage on premises. See edge AI vs cloud AI.
Can existing CCTV cameras be used?
Sometimes. Modern network cameras can often feed an analytics system. But cameras positioned to see faces at the door are poorly placed for counting, and old analogue or low-resolution cameras may not be usable. Expect to add one or two purpose-placed cameras for counting, and reuse the rest for zones and security.
The useful retail questions are about crowds: how many, where, how long. A system that answers them never needs to know who anyone is.
What about PDPA and customer trust?
Video of identifiable individuals is personal data under Singapore's Personal Data Protection Act. For analytics, the design choices that keep risk low:
- Measure, do not identify. Counts, zones and durations. No face recognition, no linking to loyalty accounts or payment data.
- Aggregate. Keep statistics, not per-person records. Do not keep individual tracks longer than needed to produce the count.
- Process locally, and know exactly what, if anything, leaves the store.
- Give notice. Signs stating that cameras are used for security and for store analytics.
- Limit the purpose. Do not quietly extend it to monitoring individual staff performance; that is a separate decision with its own obligations.
- Be cautious with demographic estimates. Age and gender estimation moves towards profiling and needs a clear reason and advice.
- Check your vendor. If analytics run on a provider's platform, they are processing data on your behalf.
This is general information, not legal advice; check current guidance from the Personal Data Protection Commission for your situation. See CCTV and PDPA and, for screens that react to audiences, AI digital signage and audience analytics.
Where should a retailer start?
- Pick one question. Conversion rate, or queue time, not everything.
- Start with the entrance. A counting camera gives the most useful single number.
- Check accuracy by hand. Count manually for an hour and compare.
- Run it for four weeks before drawing conclusions; retail traffic varies by week.
- Act on it. A roster change or a moved display, then measure again.
- Extend to queues and zones once the first number has paid its way.
Buying in Singapore
Cameras with on-device detection: the Synology BC510 and Synology TC510 (S$276 each). Recording and management: the Synology DS925+ (S$1,260) with Seagate SkyHawk drives from S$158. For developing custom in-store analytics, the NVIDIA Jetson AGX Orin 64GB Developer Kit (S$5,871.60) (listed 12 Nov 2026; prices change; ask for a current quotation). TP-Link VIGI cameras such as the VIGI C485 are price on request. Retail analytics software is chosen separately; confirm that the cameras and recorder you pick support the measures you want. To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is AI video analytics in retail? It is the use of cameras and AI to measure store activity, such as visitor numbers, occupancy, queue lengths and which areas customers visit, instead of only recording video for security.
How accurate is people counting with cameras? With a camera positioned for counting, typically overhead at the entrance, it is generally reliable. Accuracy falls with poor placement, bad lighting or dense crowds, so test it against a manual count.
Can existing CCTV cameras be used for retail analytics? Often in part. Modern network cameras may work for zones and security, but accurate counting usually needs a camera placed specifically for it.
Is retail video analytics allowed under PDPA in Singapore? It can be, when it measures aggregate activity without identifying individuals, processes video locally, limits retention and gives clear notice. Seek advice for your specific deployment.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
