A better understanding of mobility and gait patterns can substantially improve fall risk assessment. Recent progress in this field has been driven by wearable inertial measurement units (IMUs), which record temporal aspects of movement (such as step time) in both laboratory or clinical environments and during daily activities. Such data contribute to fall risk evaluations grounded in habitual behavior. However, comprehensive interpretation of IMU-derived mobility measures in free-living conditions is currently restricted by a lack of contextual awareness. For example, while variations in stride length can be quantified, the specific reasons for these changes—whether stemming from personal factors or external surroundings—frequently remain undetermined. To address this gap and achieve deeper insight into habitual fall risk via IMU-based metrics, wearable video cameras are proposed as a complementary tool. Nevertheless, manually analyzing video material is time-consuming and labor-intensive, requiring extended viewing and annotation of surroundings while also raising important privacy concerns. Therefore, robust automated artificial intelligence (AI) techniques, trained on varied datasets, are essential for reliable environment classification. The present work introduces a new dataset compiled from online video sources. It presents a deep learning system that employs chained convolutional neural networks to automatically detect environment type (indoor versus outdoor) and terrain characteristics (for example, carpet or grass). This dataset includes 146,624 video-based images (environment: 79,251; floor visible: 28,347; terrain: 39,026). After training, the individual classifiers reached F1-scores of ≥ 0.84 on an independent, manually annotated validation set (environment: 0.98; floor visible indoor: 0.86; floor visible outdoor: 0.96; terrain indoor: 0.84; terrain outdoor: 0.95). When applied to fresh data, the isolated networks achieved accuracies between 51 and 100%, while the full pipeline attained 45–90%. This project is ongoing, with continued optimization of the AI components to increase classification precision and enable automated contextualization of mobility and gait data for fall risk evaluation. Additional work will focus on collecting primary recordings from participants in their natural living environments to improve the dataset’s diversity and representativeness.
