Out-of-Clinic Measurement of Time-Domain Gait Parameters From Low-Frame-Rate Videos

Abstract

Time-domain gait parameters provide important indicators for clinical assessments, but their measurement often requires specialized in-clinic devices. Video provides an accessible alternative for out-of-clinic gait monitoring, but its low frame rate limits its ability of precise gait-event timing. This paper presents a video-based method for measuring time-domain gait parameters by performing sub-frame gait-event localization from human pose sequences. To effectively model sub-frame gait transitions, we employ velocity augmentation for poses and a spatial-temporal encoder with a multi-head decoder. Evaluation on the UnderPressure dataset shows that the system maintains a high ten-ms-level accuracy across stride, step, stance, and swing time with 30 fps input, and is robust against reduced frame rates. These results suggest that our method overcomes the limitation of video frame rate in time-domain gait parameter measurements.