Accuracy Is Not One Number

A Measurement Assurance Framework for Camera-Based Human Movement

Abstract

Camera-based human movement systems can produce fundamentally different outputs, including event detections, repetition counts, temporal measures, two-dimensional kinematics, and reconstructed three-dimensional trajectories. Asking for a single platform-level “accuracy percentage” therefore conflates measurands with different measurement chains, sources of uncertainty, and validation requirements.

This scientific note presents the PoseIQ Measurement Assurance Framework, in which evidence is attached to a defined measurand, task, measurement configuration, and operating envelope, rather than assumed to generalise across an entire platform. Performance should be evaluated against an appropriate reference using metrics suited to the measurement question, including bias, absolute error, limits of agreement, repeatability, reliability, classification performance, and failure rate where applicable. Model confidence is distinguished from measurement accuracy, and unsuccessful observations are treated as part of performance evidence rather than silently excluded.

The framework introduces measurement-specific Evidence Cards that document the validation protocol, operating envelope, quantitative performance, failure criteria, limitations, and evidence status of a configured measurement. The objective is not to weaken the requirement for validation, but to make validation claims specific, reproducible, and bounded. Evidence established for one configuration should not be extrapolated to untested measurements, populations, or capture conditions without further evaluation.