Optimizing Sensor and Data Selection on Lower Limbs via Deep Learning for Real-Time Human Activity Recognition

Area of Science:

• Robotics
• Biomedical Engineering
• Machine Learning

Background:

• Real-time human activity recognition (HAR) is vital for adaptive control in lower limb exoskeletons.
• Current systems face challenges in balancing accuracy, latency, and sensor complexity.
• Deep learning models offer potential for advanced HAR but require careful sensor selection.

Purpose of the Study:

• To systematically evaluate sensor combinations and data modalities for real-time HAR.
• To assess the trade-offs between accuracy, latency, and model complexity using deep neural networks.
• To establish design principles for HAR-driven control in assistive robotics and mobile health.

Main Methods:

• Utilized a dataset from 21 subjects performing six locomotion activities.
• Employed deep learning models including MLP, LSTM, and CNN-LSTM with 50 ms sliding windows.
• Assessed various sensor combinations: joint angles, derived angular velocities, and inertial measurements.

Main Results:

• Bilateral joint angles (hip, knee, ankle) achieved 98.98% accuracy, outperforming unilateral setups.
• Adding a thigh-mounted IMU increased accuracy to 99.23% through multimodal sensor fusion.
• Derived joint angular velocities enhanced accuracy by up to 15%, with minimal configurations (bilateral hip + velocities) reaching over 94%.

Conclusions:

• Bilateral sensor configurations, especially with derived angular velocities, provide high accuracy for HAR.
• Multimodal sensor fusion, including IMUs, further improves recognition performance.
• Findings offer practical solutions and design principles for low-power, HAR-driven assistive robotic systems.