Frontal Plane Gait Analysis using Pose Estimation Models

Historically, gait analysis has been performed through visual observation or assisted by tools such as wearable sensors, motion capture systems, and force plates. However, the accuracy of visual observation can be affected by the subjective judgment and experience of clinicians, and instrumented gait analysis requires trained operators to achieve precise results. In recent years, deep learning technology has advanced, and researchers have turned their attention towards deep learning-based pose estimation models, which have shown promise in assessing walking and running gait with encouraging outcomes. However, although markerless gait analysis based on pose estimation has potential, it is usually confined to the sagittal plane and is not suitable for use in a clinical or home setting. This research proposes a new method for analyzing gait in the frontal plane using pose estimation models such as OpenPose, YOLOv7 Pose, and MediaPipe Pose and a single camera. The main objective of the study is to evaluate the reliability of these models for frontal plane gait analysis by comparing them to the widely accepted 3D Vicon motion capture system. Additionally, the study aims to determine the best pose estimation model for automated gait analysis in the frontal plane. The MediaPipe Pose model was found to be the best model for frontal gait assessment due to its moderate inference speed (17.58 fps) and strongest correlation of gait outcomes with the Vicon motion capture system (r: 0.924, ICC(2,1): 0.919).