How to read a traffic counting accuracy claim

1. 1

   Is it volume accuracy or class accuracy?

   Volume accuracy compares the total counted with the total present, so opposite errors cancel: a count that misses 40 two-wheelers and adds 40 cars has a perfect total and two wrong classes. Class accuracy asks whether each vehicle landed in the right column, and weighted by each class’s count it does not allow that cancellation.

   Our answer

   95.0% by volume and 92.1% weighted by class, on the same 16,560 vehicles and the same footage — 2.9 points apart from the choice of measure alone. We lead with the lower one.

2. 2

   What was it measured against?

   A figure is only as good as the count it was checked against. The strongest reference is a hand count by a separate organisation. A vendor’s own review is weaker, and agreement with another model’s output only shows how closely two pieces of software agree.

   Our answer

   All 16,560 vehicles were counted by hand by an independent survey firm.

3. 3

   How many vehicles, at how many sites?

   Accuracy moves from site to site even when nothing in the software changes, so a figure from one favourable site says little. Ask for the number of hand-counted vehicles, and for the per-site figures rather than only the pooled one.

   Our answer

   16,560 vehicles at 3 sites. With identical software, model and settings, volume accuracy ranged from 92.3% to 99.7% between them.

4. 4

   How many classes was the traffic split into?

   Fewer classes leave fewer ways to put a vehicle in the wrong one. A figure measured on a handful of classes and a figure measured on a full survey schedule are not the same test, even when the percentages look alike.

   Our answer

   19 class columns — 18 named vehicle categories plus Other. The current model has no class for articulated trucks or cycle rickshaws; in this test the hand count recorded 14 cycle rickshaws, and the software counted none.

5. 5

   Were thin classes excluded, or averaged in?

   A class with two vehicles in the whole survey swings 50 points on a single mistake. Averaged in, it adds noise; quietly dropped, it can hide a weakness. Ask what the rule is.

   Our answer

   Any class with fewer than 50 hand-counted vehicles is marked statistically unreliable and excluded from the headline. The 6 classes that remain carried 98.9% of all hand-counted vehicles.

6. 6

   Under what traffic and camera conditions?

   Two-wheelers are the smallest class and the hardest to resolve, so the traffic mix moves the figure. So does the camera: how many pixels a vehicle occupies at the counting line governs every camera-based counter. Ask for the traffic composition, the camera geometry, and whether the traffic was live and uncontrolled.

   Our answer

   Three sites, Indian arterial roads, live uncontrolled traffic; two-wheelers were 60% of all traffic. Two-wheeler accuracy was 94.2% where the median two-wheeler stood 94 px tall at the line, and 88.9% at 86 px — 5.3 points for 8 pixels.

7. 7

   Can a single fifteen-minute figure be audited?

   A survey is reviewed interval by interval. Ask whether any figure in the deliverable can be traced to the vehicles behind it, or whether the total has to be taken on trust.

   Our answer

   Every counted vehicle is logged with its timestamp, class and confidence, and an annotated verification video ships with the count, so any interval is checked by scrubbing to it and counting.

The full method — both measures, who produced the hand counts, the per-class figures and the site requirements behind the camera answer — is on how we measure our own error. Where manual counting is still the better instrument is set out in automated against manual counting.