Egocentric vision-based detection of surfaces: towards context-aware free-living digital biomarkers for gait and fall risk assessment

Abstract

Background Falls in older adults are a critical public health problem. As a means to assess fall risks, free-living digital biomarkers (FLDBs), including spatiotemporal gait measures, drawn from wearable inertial measurement unit (IMU) data have been investigated to identify those at high risk. Although gait-related FLDBs can be impacted by intrinsic (e.g., gait impairment) and/or environmental (e.g., walking surfaces) factors, their respective impacts have not been differentiated by the majority of free-living fall risk assessment methods. This may lead to the ambiguous interpretation of the subsequent FLDBs, and therefore, less precise intervention strategies to prevent falls. Methods With the aim of improving the interpretability of gait-related FLDBs and investigating the impact of environment on older adults’ gait, a vision-based framework was proposed to automatically detect the most common level walking surfaces. Using a belt-mounted camera and IMUs worn by fallers and non-fallers (mean age 73.6 yrs), a unique dataset (i.e., Multimodal Ambulatory Gait and Fall Risk Assessment in the Wild (MAGFRA-W)) was acquired. The frames and image patches attributed to nine participants’ gait were annotated: (a) outdoor terrains: pavement (asphalt, cement, outdoor bricks/tiles), gravel, grass/foliage, soil, snow/slush; and (b) indoor terrains: high-friction materials (e.g., carpet, laminated floor), wood, and tiles. A series of ConvNets were developed: EgoPlaceNet categorizes frames into indoor and outdoor; and EgoTerrainNet (with outdoor and indoor versions) detects the enclosed terrain type in patches. To improve the framework’s generalizability, an independent training dataset with 9,424 samples was curated from different databases including GTOS and MINC-2500, and used for pretrained models’ (e.g., MobileNetV2) fine-tuning. Results EgoPlaceNet detected outdoor and indoor scenes in MAGFRA-W with 97.36% and 95.59 (leave-one-subject-out) accuracies, respectively. EgoTerrainNet-Indoor and -Outdoor achieved high detection accuracies for pavement (87.63%), foliage (91.24%), gravel (95.12%), and high-friction materials (95.02%), which indicate the models’ high generalizabiliy. Conclusions Encouraging results suggest that the integration of wearable cameras and deep learning approaches can provide objective contextual information in an automated manner, towards context-aware FLDBs for gait and fall risk assessment in the wild.