Validity of a Convolutional Neural Network-Based, Markerless Pose Estimation System Compared to a Marker-Based 3D Motion Analysis System for Gait Assessment—A Pilot Study

Gait analysis is a valuable tool for a wide range of clinical applications. Until now, the standard for gait analysis has been marker-based 3D optical systems. Recently, markerless gait analysis systems that utilize pose estimation models based on Convolutional Neural Networks (CNNs) and computer vision have gained importance. In this pilot study, we validated the performance of a CNN-based, markerless pose estimation algorithm (Orthelligent^® VISION; OV) compared to a standard marker-based 3D motion capture system in 16 healthy adults. Standard gait metrics were analyzed by calculating concordance correlation coefficients (CCCs) and coefficients of variation. With regard to gait event detection, we found good overlaps for both systems. Compared to the marker-based motion analysis, OV achieved a strong to almost complete concordance regarding the sagittal measurement of cadence, gait variability, step time, stance time, step length, and double support (CCC ≥ 0.624), as well as regarding the frontal plane parameters of cadence, step time, stance time, and step width (CCC ≥ 0.805). For gait symmetry only, we found a moderate to weak correlation. These results support the CNN-based, markerless gait analysis system OV as an alternative to marker-based 3D motion capture systems for a broad variety of clinical applications.