Comparison of kinematic and dynamic sensor modalities and derived features for human motion segmentation

Abstract

Human motion segmentation aims to extract individual motion repetitions from a continuous stream of data, typically using a single sensor modality. However, with the numerous sensor modalities available for motion measurement, it can be difficult to determine which modality is the most suitable. This paper investigates how segmentation accuracy is affected by the choice of sensing modality. Motion capture joint position, kinematic, force plate ground reaction force, centre of pressure, and joint torque features were considered, and their segmentation accuracy compared using classifier-based segmentation. It was found that joint position, joint angle, and ground reaction force produced similar accuracy values at 96%. These results suggest that raw motion capture and force plate sensor data can provide comparable accuracy to joint angles, reducing the need for computationally expensive inverse kinematic/dynamic computation and difficult parameter estimation.