Lightweight markerless identification of temporal gait outcomes with BlazePose

Traditionally, gait analysis is carried out in controlled settings, conducted by a trained professional which can limit reproducibility, validity, and inter-rater reliability. Recently, wearable inertial measuring units (IMUs) have been used to objectively measure gait characteristics. However, the need for manual attachment as well as data extraction and analysis, can create practical limitations, especially within low-resource environments. Lightweight artificial intelligence (AI) models for scalable and direct use on smartphones may also offer a more practical and routine gait assessment. Here, the proposed approach in this paper implements BlazePose, a lightweight pose estimation framework able to estimate 3D anatomical locations from a smartphone-based video stream. Anatomical locations inform a gait feature extraction layer to estimate step, stride, contact and swing times. The proposed approach was validated against a reference (proxy) standard (MobilityLab, APDM), performing robustly (ICC2,1 = 0.859-0.99, p=0.756-0.984). The proposed approach eliminates the need for specific hardware such as wearables, making it more cost-effective, accessible and scalable for gait assessment in low-resource settings.