Markerless 2d kinematic analysis of deepwater running using deep learning

We used a markerless Deep Learning-based method to perform 2D kinematic analysis of deepwater running. Twenty-one volunteers performed deepwater running while a GoPro camera recorded sagittal plane lower limb motion. A deep neural network was trained to predict the locations of lower limb landmarks. Just 300-500 labelled images were sufficient to train the network to be able to position joint markers with an accuracy close to that of a human labeler (mean difference ~1cm). This level of accuracy is probably sufficient for many 2D applications (e.g. sports biomechanics, coaching, rehabilitation), and could likely be further improved by modifying model parameters. Our approach represents a low-cost solution for kinematic analysis that could easily be modified for use in other movements and settings, including 3D motion analysis outside of the laboratory.