Validating Vision-Based Trail Counting in Real-World Conditions

A trail counter is only useful if its counts can be defended. For a vision-based counter, that depends on whether the sensor can reliably see and classify every trail user who passes through its field of view — the person walking a dog right in front of the camera, the cyclist crossing at the far edge of the scene, and the group of three walking side by side.

This study asked: across the distances, positions, lighting conditions, and user types a typical trail installation actually encounters, does VisionAI One detect and classify trail users reliably enough to produce defensible counts?

VisionAI One units fitted with a 120° wide-angle lens were mounted beside active multi-use paths and captured frames of uncontrolled public traffic over multiple days, under clear sun, overcast skies, backlit afternoon light, and dusk.

Every frame was processed through the same production vision pipeline used in Waypoint deployments. That pipeline detects people, bicycles, vehicles, and animals; fuses riders with their bicycles or motorcycles; applies the size and shape gates that decide whether a detection becomes a counted passage; and estimates body pose and direction of travel.

Each frame was then reviewed with its detection overlay so that counted, discarded, and missed subjects could be inspected directly.

Detection performance was strongest within 100 ft and declined beyond it. Subjects that appear smaller than the pipeline's minimum size threshold are intentionally not counted, because counting them reliably is not possible at that resolution — this is a deliberate trade against false positives.

Strong backlighting and blown-out sky lowered detection confidence in some frames, and very low light produced few usable frames. Seated or stationary people are not counted as passages, which is the intended behavior for a trail counter but means the system is not a crowd-occupancy tool on its own.