Wearable and Intelligent Movement Technologies

We do

Design of wearable IMU and pressure-sensor systems to monitor gait, posture, and balance outside the lab.

Application of machine learning and neural models to derive biomechanical variables (e.g., center of mass control, cost of transport) from sensor data.

Translation of these tools into fall-risk screening, rehabilitation progress tracking, and digital health solutions.

Journal Publication

[1] Yu, C. H., Yeh, C. C., Lu, Y. F., Lu, Y. L., Wang, T. M., Lin, F. Y. S., & Lu, T. W. (2023). Recurrent Neural Network Methods for Extracting Dynamic Balance Variables during Gait from a Single Inertial Measurement Unit. Sensors, 23(22), 9040.

Projects

Building...

A. AI-Based Wearable Technology

The development of AI-based wearable systems arises from the increasing need for continuous and quantitative assessment of human gait and balance control in daily environments. Conventional infrared motion-capture and forceplate systems, although precise, are restricted to laboratory use and are unsuitable for long-term monitoring. Recent advances in inertial measurement units (IMUs) combined with deep learning have enabled the estimation of whole-body dynamics using body-attached IMU sensors. This approach bridges the gap between laboratory-grade biomechanical accuracy and real-world mobility evaluation.

The aim of this research is to develop an intelligent wearable framework that can monitor gait speed, dynamic balance, and age-related declines in balance, cognition, and muscle strength in older adults. By embedding artificial intelligence in wearable sensing, the system provides rapid screening for mobility deterioration and mild cognitive impairment (MCI) through prolonged, self-administered assessments in real-life settings.

This research employs temporal neural network architectures, including recurrent neural networks (RNNs) and Transformer-based models, to map IMU signals to biomechanical reference variables such as inclination angles (IA), which represent the relative motion between the center of mass (COM) and the center of pressure (COP), as well as their rates of change (RCIA) (Figure 1). Federated learning frameworks are also implemented to enable privacy-preserving model training across distributed clinical sites while supporting large-scale data utilization under medical data protection regulations (Figure 2). The proposed algorithms are trained and validated using synchronized IMU and motion-capture datasets collected during walking.

Studies have demonstrated the feasibility and accuracy of this approach. RNN-based methods achieved root-mean-square errors within 3–5% of ground-truth reference data when estimating IA and RCIA variables during gait. Further developed models identified group differences between young and older adults and successfully classified MCI and healthy controls during dual-task walking. Future work will focus on large-scale and longitudinal data collection in community settings and to enable community-based AI screening for the early detection of mobility and cognitive decline.

Figure 2. Illustration of the federated learning (FL) training framework applied across medical institutes and research centers. The central server coordinates the training of a global federated model by aggregating parameters from multiple local models independently trained at different hospitals and research institutions. Each participating site trains its local model on private clinical gait data without sharing raw information, ensuring compliance with data protection regulations. The FL algorithm introduces a proximal term into the local optimization objective to constrain each client’s updates to remain close to the global model, thereby improving stability and convergence under heterogeneous (non-IID) data conditions across sites. Through iterative local training and global aggregation, this framework enables collaborative model development among institutions while safeguarding patient privacy and maintaining high generalization performance.

B. Markerless Motion Capture and Force Measurement

In clinical gait analysis, traditional infrared motion-capture systems with reflective markers and multi-forceplate setups remain the gold standard for quantifying joint kinematics and ground reaction forces (GRFs). However, these systems are costly, labor-intensive, and impractical for routine clinical use. The need to attach reflective markers precisely on anatomical landmarks is particularly challenging in pediatric patients and individuals with cognitive impairment, who may be uncooperative or unable to tolerate long preparation procedures. Similarly, forceplate-based measurement protocols require participants to place each foot on and within a separate forceplate in consecutive steps, which often necessitates multiple practice trials. This setup can be especially difficult for individuals with physical weakness or severe neuromuscular or orthopedic disorders and may alter their natural gait pattern. These challenges highlight the clinical demand for markerless and physics-informed approaches that can estimate motion and force data more naturally and adaptively across diverse patient populations.

Recent advances in artificial intelligence (AI) and computer vision have enabled markerless motion capture systems to reconstruct three-dimensional human motion directly from video recordings with unprecedented accuracy. In this line of research, the aim is to develop an integrated AI-based framework that combines vision-driven motion reconstruction with physics-informed force estimation for comprehensive gait analysis. To achieve this goal, human pose estimation (HPE) models are applied to synchronized video sequences to identify two-dimensional joint locations, and the corresponding three-dimensional joint centres are reconstructed using the weighted direct linear transformation algorithm. A marker augmentation approach was implemented, in which a recurrent neural network based on long short-term memory (LSTM) architecture is developed to predict the locations of reflective markers from the reconstructed joint centres (Figure 3). A physics-informed deep learning models are also developed to estimate bilateral ground reaction forces (GRFs) and centres of pressure (COPs) from a single forceplate or from reconstructed motion features during gait (Figure 4). By integrating AI-driven kinematic reconstruction with intelligent force prediction, this framework enables accurate, efficient, and marker-free assessment of whole-body dynamics while maintaining mechanical coherence between motion and force domains. The continued advancement of AI technologies is expected to further enhance clinical applicability, supporting objective, accessible, and scalable gait assessment for pediatric, neurological, and rehabilitation populations.

Figure 3. Marker augmentation using a long short-term memory (LSTM) network. Three-dimensional (3D) skeletal joint points obtained from markerless motion capture are used as inputs to a recurrent neural network based on an LSTM architecture. The network learns the temporal and spatial relationships between joint centres and skin-mounted reflective markers and predicts the corresponding 3D anatomical landmark positions. This process, termed marker augmentation, allows the generation of virtual markers dynamically consistent with experimentally measured trajectories, thereby improving the anatomical correspondence and biomechanical validity of the reconstructed motion data.

Figure 4. Architecture of the physics-informed residual recurrent neural network (PI-ResRNN) model for decomposing bilateral ground reaction forces (GRF) and centre of pressure (COP) from a single forceplate or from whole-body motion. The model receives six time-series signals (three force and three moment components) as input and passes them through ten residual bidirectional long short-term memory (LSTM) layers, each containing 4096 cells. Rectified linear unit (ReLU) activations and residual subtraction connections are incorporated to enforce physical constraints. The output layer produces GRF and ground reaction moment (GRM) for the leading foot, while the GRF and GRM for the trailing foot and the COP for both feet are computed based on the laws of mechanical equilibrium. A custom physics-informed loss function is used during training to ensure consistency between the model predictions and measured data in terms of force, moment, and pressure.

Wearable and Intelligent Movement Technologies

We do

Journal Publication

Projects

A. AI-Based Wearable Technology

B. Markerless Motion Capture and Force Measurement

此網站使用 cookie。