Pendulum-Based Characterization of a Commercial IMU Sensor and Real-Time OpenSim Integration for Upper-Limb Motion Analysis

Research on human motion representation commonly investigates portable, wearable, and ergonomic sensing systems. Cameras, infrared sensors, and inertial measurement units (IMUs) are widely used to reproduce and validate human movement. Known limitations persist, including increased error during slow movements, the gimbal lock effect in Euler space, and the requirement for one sensor per joint. The objective of this work is twofold: first, to characterize the measurement accuracy of a commercial IMU sensor (BWT901BLE) under controlled conditions using a fixed-arm pendulum model that replicates the single-degree-of-freedom planar kinematics of elbow flexion–extension, comparing angular position, angular velocity, and angular acceleration outputs against a video-based reference system; and second, to describe and publish a complete data processing pipeline—from raw sensor readings to real-time biomechanical motion visualization within OpenSim—demonstrated through upper limb motion recordings from 6 participants, whose data were used to generate motion files and estimate muscle fiber lengths and activation patterns within OpenSim. Regarding sensor characterization, experiments compared sensor data against the video-based reference. The inter-sensor angular position mean error was 0.765° (100 Hz) and 0.445° (200 Hz); angular velocity mean error was 0.124°/s (100 Hz) and 0.277°/s (200 Hz). Direct Euler angle measurements outperformed quaternion-to-Euler conversion (mean RMSE 5.69° vs. 53.1° at 100 Hz; 5.08° vs. 41.8° at 200 Hz). Angular velocity showed the highest agreement with the video-based reference (mean RMSE 0.60 rad/s at 100 Hz and 0.43 rad/s at 200 Hz; mean R = 0.982 and 0.991). Raw accelerometer output showed negligible correlation with the video-based angular acceleration reference (mean R ≈ 0.00–0.05); however, acceleration derived from angular velocity differentiation achieved high accuracy (mean RMSE 4.43 rad/s2 at 100 Hz and 3.06 rad/s2 at 200 Hz; mean R = 0.976 and 0.989). Regarding the OpenSim integration, the real-time visualization pipeline achieved an effective frame rate of 40–50 fps with an estimated end-to-end latency of 35–50 ms, and the recorded motion data were used to estimate muscle fiber lengths and activation patterns through OpenSim’s analysis tools. These findings confirm that angular velocity is the most reliable output of this sensor class.