gAIt: Deep Learning Based Evaluation of Injurious Running Biomechanics Using Pose Estimation

Biomechanical errors in running form have increasingly been linked to the development of musculoskeletal injuries, precluding runners from the physical and mental benefits of long-distance running. Existing works on biomechanical anal-ysis of running form utilize expensive equipment and focus on understanding injury causes instead of predicting injury risk. In this paper, we present a novel dataset (available upon request) and three deep learning models for video analysis of running biomechanics using two-dimensional pose estimation. Our approach predicts scores corresponding to the severity of in-jurious biomechanical errors, enabling runners to improve their form based on these predictions. We collect training examples through web-scraping and supplemental recordings, to which we apply BlazePose, a pre-trained pose estimation model, to obtain spatiotemporal data series. We explore three deep learning models for time series analysis, of which a Residual Neural Network (which we term gAIt) performs best, and all models achieve greater accuracy than baseline K-Nearest Neighbors with Dynamic Time Warping. Our gAIt model and future work with this dataset may provide recreational runners with an accessible and accurate tool to improve their running form in order to stay injury-free.