Detecting environmental barriers affecting older adult pedestrians via Gramian angular field-based CNN of smartphone sensor data.

Introduction: Promoting safe walking among older adults requires precise identification of environmental barriers that disrupt gait. Traditional adult- and survey-based walkability assessments are labor-intensive and often miss transient hazards, while prior wearable-sensor methods-threshold-based acceleration, Maximum Lyapunov Exponent (MaxLE, a gait-stability index quantifying the local divergence of gait dynamics), and information entropy-either lack individual sensitivity or depend on aggregated data. This study introduces a framework that converts smartphone IMU time-series into Gramian Angular Field (GAF) images for classification by a lightweight CNN. Methods: Twenty older adults completed walking trials along a 1.2 km urban route featuring common barriers (uneven sidewalks, curb drops, narrow alleys, driveway crossings). IMU data were filtered, segmented into 2-s windows, transformed into 200 × 200-pixel GAF images, and evaluated under leave-one-subject-out cross-validation. Results: Among three benchmarks-peak-acceleration threshold, MaxLE (82.3% accuracy, F1-score = 0.45), and multi-user entropy-the GAF-CNN achieved 90.8% accuracy, 93.0% sensitivity, and 88.1% specificity, significantly outperforming the baselines (75-85% accuracy). Spatial mapping confirmed close correspondence between detected anomalies and true barrier locations. Discussion: These findings demonstrate that image-based deep learning provides a practical and interpretable solution for real-time, personalized detection of environmental barriers, offering a scalable tool for data-driven walkability enhancement in age-friendly urban design.