The short answer: a camera-fed perception model measures the whole approach continuously, where inductive loops and fixed-timing plans report presence at a point on a schedule. The reported result is usually a 10–25% reduction in intersection delay against fixed-timing plans (as reported in vendor and pilot studies), with associated throughput gains and fewer stops per vehicle. That figure is only meaningful when it comes attached to a second number — the share of operating hours in which the perception layer verifiably held its detection-quality threshold.

Most published answers stop at the first half. “AI sees more than a loop detector, therefore AI is better” is not wrong. It is simply unusable to a transport authority trying to work out what changes on the ground the week after commissioning.

What can perception measure that a loop detector cannot?

Video-based systems count vehicles in four lanes simultaneously while inductive loops report only presence in a single sensor footprint. Everything a signal plan does downstream is inferred from a chain of those point answers.

A perception model on the same approach reports:

Quantity	Inductive loop / fixed plan	Camera-fed perception
Vehicle presence at a point	Yes	Yes
Queue length across the approach	Inferred from occupancy	Measured directly
Mode mix (car, HGV, bus, bike)	No	Yes
Pedestrian and cyclist presence	Push-button only	Detected passively
Turning demand per lane	Requires per-lane loops	Measured per lane
Update cadence	Fixed plan / schedule	Continuous

The advantage is not “more data”. It is that the controller stops working from a schedule written months ago and starts working from what the approach currently contains.

Which gains actually move delay — and which do not

Prediction accuracy above 80% (illustrative threshold) still produces net-negative outcomes when false positives trigger unnecessary signal-phase extensions during peak hours.

The operating condition nobody puts in the headline

Adaptive signal control requires two data streams: real-time occupancy and short-horizon demand forecasts. The advantage exists only for the hours the perception layer is verifiably detecting correctly. Night, heavy rain, low-sun glare, seasonal foliage growth over a camera’s field of view, and a lens that slowly hazes — each degrades detection without producing an obvious alarm. The controller keeps acting on the model’s output, and the output is now wrong.

That is a monitoring problem, not a modelling problem. It is not solved by a better detector; it is solved by drift telemetry, a perception-specific evaluation suite that covers the hard conditions, and a release-readiness gate that a corridor has to pass before its numbers are accepted. We treat that package as part of the deliverable rather than a follow-on, which is the argument we develop in our work on production AI reliability.

So a 20% delay reduction (illustrative figure) verified across 96% of operating hours is a different asset from the same 20% measured at a peak-hour daylight commissioning window. Same headline, different procurement decision.

What to ask a vendor for

Combine mean wait time per approach with detection-zone coverage percentage to assess whether an upgrade justifies capital expenditure.

• Detection performance broken out by night, precipitation, and glare — not a single aggregate.
• The drift-monitoring mechanism and what it triggers when a camera degrades.
• The evaluation set: was it collected at this corridor, or borrowed?
• The condition under which the vendor would recommend keeping the existing controller.

That last one matters. On a low-volume, single-phase junction with stable demand, a well-tuned traditional controller can be the correct engineering answer, and a vendor unwilling to say so is telling you something.

The interesting question for a smart-city programme is not whether AI beats fixed timing — it usually does, under observation. It is whether your pilot is structured so the number it reports survives its first winter.