Radar-Based Deep Learning for Gait Smoothness Estimation: A Feasibility Study

This study investigates the feasibility of using radar-based deep learning models to estimate gait smoothness, a key biomechanical parameter associated with motor control and rehabilitation outcomes. Traditional gait assessment methods, such as optoelectronic motion capture and wearable sensors, are often costly, intrusive, or require controlled environments. Radar-based approaches offer a promising alternative by enabling contactless, continuous monitoring of human movement. In this study, sixty healthy participants walked on a treadmill at three different speeds (2, 4, and 6 km/h) while their three-dimensional body center-of-mass (BCoM) velocity was recorded using an optoelectronic system, which served as the ground truth for smoothness estimation. Simultaneously, micro-Doppler radar signals were acquired and processed into spectrograms representing movement dynamics. Nine convolutional neural networks were trained to predict BCoM smoothness from radar-derived decibel-intensity matrices, with a dedicated model for each combination of walking speed and movement component (anteroposterior, mediolateral, and craniocaudal). The models achieved mean absolute percentage errors below 10% in most conditions, except for the anteroposterior component at 6 km/h, where error rates reached 14.4%. These findings suggest that radar-based deep learning can effectively estimate gait smoothness, particularly at lower walking speeds, and holds potential for real-world applications in clinical gait assessment and rehabilitation monitoring. Despite promising results, challenges remain regarding model generalization, especially at higher speeds where gait variability increases. Future research should explore more advanced deep learning architectures, such as residual networks or attention-based models, and extend the approach to clinical populations with neurological or musculoskeletal disorders. If validated in diverse conditions, radar-based smoothness estimation could provide a novel, unobtrusive tool for assessing gait quality in both clinical and everyday settings.