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

Real-time human activity recognition (HAR) is crucial for adaptive control in lower limb exoskeletons, yet achieving an optimal balance between accuracy, latency, and sensor complexity remains challenging. This study’s key idea is to systematically evaluate sensor selection and data modalities for real-time HAR using deep neural networks (DNNs) with ultra-short 50 ms sliding windows. Leveraging a dataset from 21 subjects performing six locomotion activities and employing three deep learning models [multilayer perceptron (MLP), LSTM, CNN-LSTM], we assess multiple sensor combinations, including joint angle data, derived angular velocities, and inertial measurements, and quantify their trade-offs across accuracy, latency, and model complexity. Our analysis reveals that bilateral joint angles from hip, knee, and ankle achieve 98.98% accuracy, significantly outperforming unilateral sensor setups. Adding a thigh-mounted inertial measurement unit further elevates accuracy to 99.23%, highlighting the advantages of multimodal sensor fusion. In addition, incorporating derived joint angular velocities enhances accuracy, with up to 15% increase when using single-joint bilateral inverse kinematics data. Even a minimal configuration, such as bilateral hip joints with derived angular velocities, achieves over 94% accuracy, offering practical solutions for low-power wearable systems. These findings establish actionable design principles for HAR-driven control in assistive robotics and mobile health applications.