OrientationNN: a physics-informed lightweight neural network for real-time joint kinematics estimation from IMU data

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Accurate joint kinematics are crucial for human movement analysis but traditional optical motion capture is costly and complex.
• Inertial Measurement Units (IMUs) offer a portable alternative, but accurate kinematic estimation remains a challenge.

Purpose of the Study:

• To develop a lightweight, physics-informed neural network for real-time joint kinematics estimation using IMUs.
• To ensure biomechanically consistent estimations through integrated physical constraints.

Main Methods:

• Developed OrientationNN, a compact multi-layer perceptron integrating orientation-based physical constraints.
• Evaluated OrientationNN on a public dataset, comparing it against OpenSense, MLP, LSTM, CNN, and Transformer models.

Main Results:

• OrientationNN achieved average joint angle estimation errors below 5° during ambulatory motion.
• The model outperformed the physics-based OpenSense framework across all kinematic variables.
• OrientationNN demonstrated high computational efficiency with only 4.9 × 10³ FLOPs per frame and 10.8 KB of parameters.

Conclusions:

• OrientationNN provides accurate and computationally efficient joint kinematics estimation from IMU data.
• The model presents a cost-effective and scalable solution for wearable biomechanical and motion analysis applications.