METHODOLOGY FOR CONTACTLESS MEASUREMENT OF HUMAN MOTION KINEMATIC PARAMETERS USING MULTIMODAL AI SYSTEMS

Abstract

Under conditions of rapid digitalization of measurement processes, the issue of metrological support for hardware and software systems that use artificial intelligence methods for the quantitative and qualitative evaluation of physical quantities becomes increasingly relevant. This paper presents the results of a study of the metrological characteristics of a multimodal measurement system designed for markerless motion capture and analysis of human kinematic parameters. The game of table tennis, characterized by highly dynamic changes in measured variables, was selected as the test object for validating the methodology. A structural model of error generation within the measurement channel based on a monocular camera and OpenPose neural network algorithms was developed. The components of the uncertainty budget were evaluated, including instrumental, methodological, and algorithmic contributions. An approach to validating the system’s classification decisions is proposed through comparison with reference values obtained using expert qualimetry with a verified inter-rater agreement coefficient (Fleiss’ kappa statistic). The results of the experimental study demonstrate an achievable accuracy of kinematic pattern identification at the level of 73–82%, provided that the effects of occlusions and motion blur are minimized. Furthermore, uncertainty budget analysis identifies high-speed motion blur and low sampling rates as the main limiting factors, suggesting that increasing frame rates to 60-120 fps is essential for highly accurate trajectory approximation in dynamic sports environments. This work lays the foundation for the development of standardized calibration methodologies for intelligent sports telemetry systems.