The accuracy of markerless motion capture combined with computer vision techniques for measuring running kinematics

Abstract

Background: Markerless motion capture based on low-cost 2-D video analysis in combination with computer vision techniques has the potential to provide accurate analysis of running technique in both a research and clinical setting. However, the accuracy of markerless motion capture for assessing running kinematics compared to a gold-standard approach remains largely unexplored. Objective: Here, we investigate the accuracy of custom-trained (DeepLabCut) and existing (OpenPose) computer vision techniques for assessing sagittal-plane hip, knee, and ankle running kinematics at speeds of 2.78 and 3.33 m s−1 as compared to gold-standard marker-based motion capture. Methods: Differences between the markerless and marker-based approaches were assessed using statistical parameter mapping and expressed as root mean squared errors (RMSEs). Results: After temporal alignment and offset removal, both DeepLabCut and OpenPose showed no significant differences with the marker-based approach at 2.78 m s−1, but some significant differences remained at 3.33 m s−1. At 2.78 m s−1, RMSEs were 5.07, 7.91, and 5.60, and 5.92, 7.81, and 5.66 degrees for the hip, knee, and ankle for DeepLabCut and OpenPose, respectively. At 3.33 m s−1, RMSEs were 7.40, 10.9, 8.01, and 4.95, 7.45, and 5.76 for the hip, knee, and ankle for DeepLabCut and OpenPose, respectively. Conclusion: The differences between OpenPose and the marker-based method were in line with or smaller than reported between other kinematic analysis methods and marker-based methods, while these differences were larger for DeepLabCut. Since the accuracy differed between individuals, OpenPose may be most useful to facilitate large-scale in-field data collection and investigation of group effects rather than individual-level analyses.