Fig. 2.

Data pipeline: we labeled the location of 5 keypoints on each leg and foot in 164 frames drawn from a set of 82 videos and use the labeled frames to fine tune a pretrained ResNet50 deep network architecture through DeepLabCut. At inference time, we used the fine-tuned model to track the keypoints in the video. The keypoint sequences of each joint (different colors indicate different joints) were filtered, normalized by z-scoring, and concatenated along the third dimension to train a second convolutional neural network at predicting 5 gait temporal parameters. Ground truth values for gait parameters were derived from data collected using a reference clinical system (GAITRite). A leave-1-subject-out cross-validation was used to evaluate the model performance.