Fig. 4: Detailed SCAI-gait workflow.
From: 3D pose estimation for scalable remote gait kinematics assessment
The given figure represents the workflow that was followed to benchmark different state-of-the-art 3D pose estimation neural network models to identify the most suitable 3D pose estimation neural network model for the case of SCI individuals. The best model (VideoPose3D) was then used to extract 3D keypoints (skeletons) from SCI as well as healthy individuals. Using the output 3D keypoints, the lower body keypoints were used first filtered and converted into Subsets using K-Means Clustering29 followed by multi-variate time series feature extraction using MLP Classifier using SHAP values as well as K-Means Classifier31 using clustering method. The results obtained were used to carry out digital phenotyping of SCI individuals.
