Gait Anomaly Detection With Low-Cost and Low-Resolution Infrared Sensor Arrays

Detecting anomalies in human gait could be used as indicators of human fall risk or other underlying health or psychological issues. This would require collecting reliable gait data. However, collecting human abnormal gait data is very challenging compared with data gathered from normal daily activities mainly because the former are relatively scarce and may exhibit an unmanageable variability with unpredictable combinations of distorted gait patterns. Recently, it was proposed that privacy concerns due to potential misuses of recorded gait images can be alleviated by using the thermal images captured by the low-resolution and low-cost thermal sensor arrays (TSAs). Therefore, to resolve the privacy concerns and data scarcity simultaneously, this article proposes a gait anomaly detection (GAD), to be created as a one-class classification (OCC) model and implemented as a reconstruction-based autoencoder (AE), while using TSAs to capture the input data. The data scarcity is conveniently addressed since this GAD design needs only the plentiful "normal" gait of one person of interest (POI) to build its base model. AE's were deployed since they learn the intricacies of normal gait patterns, with anomaly threshold placed on the reconstruction errors of the training data. The high performance in detecting specific classes of POI's gait anomalies, achieving impressive mean values across five critical classification metrics- F1 -score (95.26%), accuracy (96.20%), precision (92.76%), recall (97.92%), and specificity (95.00%)-demonstrates the model's feasibility and practicality. The proposed framework can facilitate independent living among older adults as an individualized data-efficient, privacy safe, and low-cost approach to GAD.