A drone can reveal why vehicles are queuing where a roadside camera sees only the front of the line. That wider view is useful, but it does not make counting automatic. A moving aircraft changes scale and perspective, parked vehicles resemble slow traffic, and a car can cross the same line more than once. A credible traffic pilot has to define the question before reporting a number.
Choose a unit of measurement
A frame count asks how many vehicles are visible in one image. A flow count asks how many distinct vehicles crossed a line during a period. Occupancy asks how much of a road or parking area is filled. These are different measures. Do not present a frame count as vehicles per hour or infer travel speed from image motion without a calibrated ground scale and stable timing.
For a gate, a line-crossing count may be the right metric. For a junction, queue length and turning movement might be more useful. For an event car park, a snapshot of occupied space may matter more than individual trips. Agree on the unit and the reporting interval with the traffic team first.
Control the camera geometry
Record altitude, zoom, camera angle, and the intended observation area. If the aircraft shifts or the camera pans, a fixed image polygon may no longer represent the same road. Use a stable hover or a repeatable viewpoint for comparisons over time. Review occlusions caused by trees, buses, flyovers, and shadows. A model cannot count vehicles that the camera never saw.
Measure errors by class and condition
- Hand-label a sample from the actual site, including two-wheelers, auto-rickshaws, cars, buses, and trucks if those classes matter to the decision.
- Compare model output with the labelled sample by class. An overall accuracy figure can conceal poor detection of the smaller or less common vehicles.
- Separate daylight, dusk, rain, glare, and heavy occlusion. Report the conditions under which a count is withheld or marked uncertain.
- For flow counts, inspect tracks that split, merge, or return across a counting line. Duplicate counts can dominate an apparently plausible total.
Treat heatmaps as summaries, not photographs
A heatmap can be based on detections, dwell time, flow, or queue length. The legend must say which one it represents, along with the time window and camera coverage. Otherwise a red patch may be read as congestion when it only shows that the aircraft spent longer watching that location. Pair the heatmap with source frames and a simple timeline so a reviewer can check the interpretation.
Use the output to answer an operational question
A useful report might compare inbound volume at two gates in fifteen-minute windows, note when a queue reached a marked point, and show the source images. It should state periods with no coverage. The traffic team can then decide whether to redirect staff, change signage, or investigate a bottleneck. The software supplies evidence and a consistent measure; the team owns the road decision.
Before scaling to more sites, repeat the validation with different road widths, viewpoints, and vehicle mixes. One well-calibrated junction does not establish performance across a city.