Comparative Analysis of Markerless Motion Capture Systems for Measuring Human Kinematics

Abstract

To date, there are several measurement methods for evaluating human kinematics based on inertial sensors or vision systems. However, a comprehensive comparison has not been undertaken to determine which of these systems offers the most appropriate accuracy for clinical or sports evaluations. This study conducted a comparative analysis of different motion measurement systems: optoelectronic system (OS), inertial measurement units (IMUs), and vision-based methods, including deep neural network (DNN) and non-DNN approaches. Ten healthy subjects were involved, performing walking (W.) and running (R.) tests at various speeds (3.5, 5.0, and 7.0 km/h). The measurement of human kinematics was conducted by taking video images via two RGB cameras, together with an IMU-based system and an OS as the gold standard. Comparative analysis was conducted on a set of measurement methods, including IMU, a method based on blob analysis (BA), and DNN algorithms: Alphapose (AP), TC former (TC), RTMPose (RTM), and MediaPipe (MP). Data analysis involved triangulation and measurement of lower limb joint angles. Results showed that vision systems do not allow ankle joint measurement, and IMUs outperformed other methods in terms of RMSE and absolute error of range of motion ( $\varepsilon _{\text {ROM}}\text {)}$ . RTM and MP exhibited results similar to IMUs, especially for the hip and knee joints, with the minimum absolute error reporting values of ( $3.1^{\circ }~\pm ~1.8^{\circ }\text {)}$ and ( $3.5^{\circ }~\pm ~1.9^{\circ }\text {)}$ for the hip joint and ( $4.0^{\circ }~\pm ~3.7^{\circ }\text {)}$ and ( $4.8^{\circ }~\pm ~4.3^{\circ }\text {)}$ for the knee joint, respectively.