Deep learning-based golf swing sequencing from videos
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3141896Utgivelsesdato
2024Metadata
Vis full innførselSamlinger
Sammendrag
This thesis delves into the effectiveness of advanced deep learning configurations for identifying the dynamic and static phases of golf swings, a fundamental skill in golf that directly influences performance outcomes. Traditional deep learning models often struggle with detecting static movements, which are subtle but crucial for comprehensive motion analysis in sports. This research gap underscores a significant need for enhanced model architectures incorporating advanced deep learning techniques designed specifically for the complexity of sports motion analytics.
To address this challenge, the study explores four innovative deep learning configurations: MobileNetV2 + LSTM, ResNet50 + LSTM, MobileNetV3 + LSTM, and a novel integration of MobileNetV3 with a Convolutional Block Attention Module (CBAM + LSTM). Each configuration is rigorously tested to evaluate its proficiency in capturing the golf swing's pronounced and subtle movement. The experiments are structured to systematically assess and compare each model's ability to accurately detect phases of the swing, focusing on integrating spatial and temporal data critical for dynamic and static phase recognition.
The results demonstrate that while traditional configurations like MobileNetV2 + LSTM provide a solid foundation for detecting dynamic movements, they fail to capture static phases accurately. However, integrating CBAM with MobileNetV3 significantly enhances model performance, particularly detecting static phases. This improvement highlights the transformative potential of attention mechanisms in refining the focus and sensitivity of neural networks, enabling them to excel where conventional architectures falter.This research has profound implications. It offers a deeper understanding of the application of neural network architectures in sports analytics and paves the way for future advancements in automated coaching tools. By enhancing phase detection accuracy, this work contributes to developing more sophisticated analytics tools to provide athletes and coaches with precise, real-time feedback essential for performance optimization.