Swimming Style Recognition for Video Sequences Using Deep Learning
Ahmad Hafrizalman Abdul Hamid, Zalinda Othman, Suhailayani Suhaimi · 2023
Today, deep learning is pivotal in automating computer vision tasks previously handled by humans. Convolutional Neural Networks (CNNs) excel in signal, image, and video processing. This study conducted experiments to create a CNN model classifying swimming video sequences (freestyle, breaststroke, butterfly stroke). Datasets were assembled from YouTube videos adhering to specific criteria for underwater views and pool-side monitoring. The study addressed challenges like camera viewpoint and noise. A 2DCNN architecture with two convolutional layers served as the baseline model, enhanced with a third convolutional layer. It achieved remarkable accuracy (99.29%), precision (99.38%), and recall (99.37%) using an input size of 64x64 for the swimClass0 dataset. This model was selected for validation against the UCF15 benchmark, achieving 97.11% accuracy, 97.11% precision, and 97.08% recall. The results highlight the superior performance of the proposed model compared to the comparative study.