gAIt is a completed research project on Deep Learning techniques for the evaluation of injurious running biomechanics using pose estimation from videos. BlazePose is used for pose estimation due to its specificity to the single-person case and balance of performance with efficiency.
Presently, a light ResNet with 384K parameters achieves the best performance, compared to a CNN, FCNN, and K-Nearest Neighbors with Dynamic Time Warping. This work introduces a new dataset (available upon request) for the development of these models as well as more nuanced models by others.
The paper for this research project won the Runner-Up Student Paper Award at the 13th IEEE International Conference on Pattern Recognition Systems.