Markerless Motion Capture of Hands and Fingers in High-Speed Throwing Task and Its Accuracy Verification

Abstract

In the motion capture systems of humans, reflective markers attached to the body have widely been used to track them by optical cameras. However, when the speed of motion increases, as the camera's angle of view and brightness are limited and the markers often fall off, especially for detailed body parts such as fingers in full-body movements, other parts of the body (palms) have been investigated instead. In this study, we attempted to acquire finger movements in high-speed throwing motion without attaching markers using automatic image recognition technology based on deep learning (DeepLabCut), and verified its accuracy comparing with conventional methods. As a result, the absolute distance between 3D coordinates obtained from the two motion capture systems was an average of 15.5 mm ~ 29.4 mm depending on tracked points and the correlation coefficients between them ranged 0.932 ~ 0.999. Therefore, the shapes of the time-series profiles of 3D coordinates obtained from the two motion capture systems were similar. These results suggest that motion measurement using markerless motion capture is possible in environments where it has been difficult to use conventional motion capture systems.