Accuracy method
How we measure our own error
Automated counting accuracy can be stated two ways. Both are legitimate, they answer different questions, and quoting only one can mislead. This page states both, names who produced the numbers we were measured against, and explains where ours degrades.
Weighted class
92.1%
Did every vehicle land in the right column?
Volume
95.0%
Total counted against total present.
Two ways to state it
Volume, and weighted class
Volume accuracy
Compares the total count against the manual count. This is the figure that matters for capacity studies, toll modelling and PCU-based design, where aggregate flow is the deliverable.
Weighted class accuracy
Averages per-category accuracy, weighted by each category’s manual count. The stricter figure: it asks whether every vehicle landed in the right column, not merely whether the total is right.
What we were measured against. All 16,560 vehicles were counted by hand by an independent survey firm — a separate company, not a separate process of ours. Every figure on this page is stated against those counts. None of it is validated against another model’s output, which would only tell you how closely two pieces of software agree.
Why they differ. Volume accuracy allows opposite errors to cancel. If a site under-counts 40 two-wheelers and over-counts 40 cars, the total is perfect while both categories are wrong. Weighted accuracy does not permit that cancellation, so it is always the more conservative number.
Three sites, Indian arterial roads, live uncontrolled traffic
11 August 2026
| Site | Manual | Counted | Volume | Weighted |
|---|---|---|---|---|
| A | 4,579 | 4,595 | 99.7% | 93.0% |
| B | 5,824 | 5,375 | 92.3% | 91.2% |
| C | 6,157 | 5,759 | 93.5% | 92.3% |
| All | 16,560 | 15,729 | 95.0% | 92.1% |
Manual — hand-counted by an independent survey firm Counted — by the software
We lead with 92.1% weighted, not 95.0% volume. Weighted accuracy asks whether every vehicle landed in the right column. Volume accuracy lets over- and under-counts cancel.
What actually governs it
Pixels on target at the counting line
Across the three surveys the software, model and settings were identical. The only variables were camera position, mounting height, viewing angle and traffic composition. Accuracy varied by a wide margin — and it varied most for the smallest vehicles.
| Survey | Accuracy | Median height |
|---|---|---|
| Site A | 94.2% | 94 px |
| Site B | 88.4% | 89 px |
| Site C | 88.9% | 86 px |
The relationship is direct and physical: the more pixels a vehicle occupies as it crosses the counting line, the more reliably it is detected and classified. An eight-pixel difference in median two-wheeler height between the best and worst site corresponds to roughly six percentage points of two-wheeler accuracy.
This is not a limitation peculiar to one product. It is the governing constraint of every camera-based counting system. A vehicle occupying too few pixels at the measurement point cannot be resolved reliably by any detector, and no amount of software tuning creates image detail the camera did not capture.
Two-wheelers are the sensitive case because they are the smallest class and, on Indian arterials, by far the most numerous — in these surveys, 60% of all traffic. A site that resolves two-wheelers well resolves everything well.
Before a frame is processed
Site requirements for best accuracy
Accuracy is largely decided by where and how the camera is installed. We are glad to review a proposed position before a survey starts.
- Camera height and angle
- A higher mounting with a steeper view separates vehicles travelling side by side. Shallow, near-horizontal views cause vehicles to overlap in the image — the most common cause of under-counting in dense two-wheeler traffic.
- Distance to the count line
- The counting line should fall in the nearer half of the field of view. Vehicles measured far up the road occupy few pixels and are intrinsically harder to classify.
- Full carriageway coverage
- The camera must see the entire width being surveyed. Where a median divides the carriageways, each direction is best served by its own view.
- Resolution and encoding
- 1080p or above, with a bitrate high enough to avoid compression smearing on small, fast-moving objects. Excessive compression removes exactly the detail two-wheeler detection depends on.
- Stable mounting
- Vibration and slow drift degrade tracking continuity. A rigid mount materially improves consistency over a long survey.
- Lighting and obstruction
- Avoid direct sun into the lens at survey hours, and keep foliage, poles and signage clear of the counting zone.
Reading the per-class figures
Thin categories are excluded, not averaged in
Categories with very low observed volumes show large percentage swings that carry no statistical meaning. Where a category records only one or two vehicles across an entire survey, a single classification difference moves its accuracy by 50 or 100 percentage points.
For this reason any category with fewer than 50 manually counted vehicles is marked statistically unreliable and excluded from the headline. The headline is carried by the categories that carry the traffic — in these surveys, six categories accounted for 94% of all vehicles.
| Category | Manual | Counted | Accuracy |
|---|---|---|---|
| Car / Van / Jeep | 2,946 | 3,036 | 96.9% |
| 3-Wheeler | 2,717 | 2,633 | 96.9% |
| Bus | 150 | 161 | 92.7% |
| 2-Wheeler | 9,921 | 8,941 | 90.1% |
| Cycle | 273 | 302 | 89.4% |
| Mini LCV | 364 | 414 | 86.3% |
How each figure is computed
- Volume accuracy
- 100 − |counted − manual| / manual
- Weighted class accuracy
- per-category accuracy, weighted by manual count; categories below 50 excluded
All figures derived from production runs dated 11 August 2026, measured against counts produced by hand by an independent survey firm. Both measures are disclosed. Two-wheelers were 60% of all traffic across the three sites. Site names and clients are withheld; the conditions that determine whether these figures transfer to your road are not.
Test it yourself
Send 15 minutes and check our arithmetic
The fastest way to evaluate a counting claim is on your own footage. We will run it, return the classified count and the annotated verification video, and tell you where the camera position is costing you accuracy.
Send us footage