Abstract
Special sports environment of swimmers increases the difficulty of event monitoring and daily training. Computational vision makes it possible to solve these problems. The pose estimation of swimmers is a basic problem to be solved in tasks of relevant computer vision. The past methods not only require a complex implementation process, but also have a very unstable performance in the face of frequent morphological changes of swimmers, and most of them only fit the scenarios that include a single swimmer.
In this thesis, we implement a method for the pose estimation of swimmers based on deep learning, which fits scenarios containing multiple swimmers. We follow the top-down method and combine the HRNet with YOLOv5 to implement our model. Our method achieves ideal accuracy, the model is easy to be trained and deployed. In addition, we also propose a dataset with annotated key points for swimmers and a slew of datasets for swimmer detection. Our key point dataset is composed of the underwater view of swimmers. Compared with the side view, the torso of swimmers collected by the underwater view is much suitable for a broad spectrum of deep learning tasks.
