Anatomical-Marker-Driven 3D Markerless Human Motion Capture.

Area of Science:

• Biomechanics
• Computer Vision
• Machine Learning

Background:

• Marker-based motion capture (mocap) is precise but time-consuming.
• Markerless mocap using deep learning is promising but sensitive to annotation errors.
• Improving 2D keypoint annotation is crucial for markerless mocap accuracy.

Purpose of the Study:

• To develop a precise 2D keypoint annotation method for markerless mocap.
• To create a high-quality annotated dataset (RRIS40) for training deep learning models.
• To validate the accuracy of the proposed markerless mocap system against marker-based systems.

Main Methods:

• Utilized a marker-based mocap system for synchronized, calibrated RGB camera setup.
• Created the RRIS40 dataset with surface anatomical landmarks annotated from marker-based mocap data.
• Trained a deep neural network to estimate 2D anatomical landmarks and employed ray-distance triangulation for 3D positions.

Main Results:

• Achieved a mean Euclidean error of 13.23 mm in 3D marker position, comparable to marker placement precision.
• The proposed method demonstrated superior performance compared to OpenCap's augmentation of 3D anatomical landmarks.
• The RRIS40 test set, containing data from 10 subjects, is publicly available.

Conclusions:

• The novel method enhances the precision of markerless mocap by improving 2D keypoint annotation.
• This approach offers a viable and accurate alternative to traditional marker-based mocap in biomechanics.
• Facilitates wider adoption of markerless mocap in scientific research through improved accuracy and accessibility.