Glossary / Retail analytics
What is dwell time?
Dwell time is how long a person stays inside a defined zone, computed by timestamping when a tracked individual enters and leaves it. It is used to measure engagement with a display, queue duration or congestion. The number depends as much on configuration as on the footage, which is why vendors disagree.
Conceptually dwell time is trivial: note when someone enters a zone, note when they leave, subtract. Every complication comes from the fact that video tracking is imperfect and a single visit rarely produces a single clean track.
How it is actually computed
- Detect people in each analysed frame.
- Track each person across frames, assigning a persistent track identifier.
- Test whether the track's reference point falls inside the zone polygon.
- Timestamp the first and last frames where that test passes.
- Apply rules for fragmentation, re-entry, minimum duration and exclusions.
Step five is where two vendors analysing identical footage produce different numbers, and it is almost never disclosed in a report.
Why one visit becomes three
Worked example: track fragmentation
A shopper spends 90 seconds at a display. During that time another shopper passes in front of her twice, each time hiding her for about two seconds. The tracker loses the track at each occlusion and starts a new one when she reappears.
Tracks produced: 3 tracks of 40 s, 30 s, 20 s
Reported (naive): 3 visitors, mean dwell 30 s
The system has tripled the footfall and cut mean dwell to a third, from one accurate piece of footage. Fixing this needs re-identification to recognise that the three tracks are one person, or a re-entry merge window that stitches tracks separated by a short gap in a similar position.
Fragmentation is the dominant error in busy zones, and busy zones are exactly the ones you care about. Any dwell figure quoted without saying how fragmentation is handled should be treated as unverified.
The six choices that change the number
| Choice | Options | Effect on the reported number |
|---|---|---|
| Reference point | Bounding box centroid or foot position | Centroid places a standing person inside a zone earlier and later than their feet do. Differences of several seconds at zone edges. |
| Re-entry window | None, or merge gaps under N seconds | The main fragmentation control. A 5 s window recovers most occlusion splits. |
| Minimum dwell floor | Typically 2 to 5 s | Excludes pass-throughs. Raising the floor increases mean dwell mechanically, without anything changing on the floor. |
| Staff exclusion | Included or excluded | Staff dwell far longer than shoppers. Including them inflates mean dwell substantially in small zones. |
| Statistic reported | Mean or median | Dwell distributions are heavily skewed. Mean is pulled up by a few long visits; median is more stable. |
| Zone geometry | Polygon in image space or ground plane | Image space zones distort with camera angle, so the physical area measured differs from the area drawn. |
Six binary or near-binary choices produce dozens of defensible answers from one video file. None of them is wrong. What is wrong is comparing a number produced under one configuration to a number produced under another, which is what happens whenever a vendor is switched and dwell appears to jump.
Using dwell without fooling yourself
Two habits make dwell useful:
- Treat it as relative, not absolute. Dwell at endcap A versus endcap B, measured the same way in the same store, is a sound comparison. Dwell in your store versus a published industry figure is not, because the configuration behind that figure is unknown.
- Freeze the configuration. Record the six settings above alongside the numbers. When someone adjusts a zone polygon or a minimum floor six months later, the trend line moves for reasons that have nothing to do with the store, and without a record nobody will work out why.
Long dwell is also ambiguous by nature. A shopper studying a display and a shopper unable to find what she wants both register as high engagement. At a service counter, high dwell is usually a queue problem rather than an interest signal. Pair dwell with conversion or transaction data before drawing conclusions from it.
Privacy
Dwell measurement does not require knowing who anyone is. Where the output is aggregate counts and durations, the appropriate design is to discard tracking data once the visit closes and retain only the aggregate. Processing at the edge makes that enforceable, since nothing at individual level ever leaves the site. Where re-identification is used to stitch fragmented tracks, the embeddings involved should be treated as personal data and retained for hours rather than days.
Frequently asked questions
Why do two vendors report different dwell times for the same footage?
Because dwell depends on configuration as much as on video. The reference point used, the re-entry merge window, the minimum dwell floor, whether staff are excluded, whether mean or median is reported, and how the zone polygon is defined all change the number. Ask for those six settings before comparing figures from different systems.
What causes track fragmentation?
Occlusion, mainly. When a person is hidden behind another shopper, a fixture or a pillar, the tracker loses them and assigns a new identifier when they reappear. One visit becomes several short tracks, which simultaneously inflates footfall and deflates mean dwell. A re-entry merge window or re-identification recovers most of these.
Is high dwell time good?
It depends entirely on the zone. At a product display, high dwell suggests engagement. At a service counter or checkout, it usually means a queue. At a shelf, it can mean interest or it can mean a shopper unable to find what she came for. Dwell needs pairing with conversion or transaction data before it means anything.
Should I use mean or median dwell time?
Median for most purposes, because dwell distributions are heavily skewed by a small number of very long visits, often staff or people waiting for someone. Mean is more volatile and moves for reasons unrelated to shopper behaviour. Whichever you choose, use it consistently, and state which one a report uses.
Does dwell time measurement need face recognition?
No. Dwell needs to know that a person is in a zone and for how long, not who they are. Detection and tracking are sufficient within a camera view, and re-identification handles stitching across occlusions or cameras. Face recognition adds identity the measurement does not require, along with obligations it does not need.
What minimum dwell threshold should I set?
Commonly two to five seconds, to exclude people simply walking through. Be aware that raising the floor increases mean dwell mechanically without anything changing in the store, so the threshold has to stay fixed for a trend line to mean anything. Record whatever you choose alongside the data.
Getting dwell numbers you do not trust?
IndoAI computes dwell on site with the configuration exposed rather than hidden, so you can see exactly how a number was produced. Send us a zone layout and we will walk through what your current setup is likely reporting.
Talk to an adviserReviewed by Dr. Vivek Gujar, Chief Strategy Officer at IndoAI Technologies Pvt. Ltd., a Pune-based edge AI camera platform founded in 2021. He reviews IndoAI's published technical claims for accuracy. Profile