TY - BOOK
T1 - Deriving Sport Scientific Insights from Broadcast Footage
T2 - Advanced Computer Vision for Human Motion Analysis
AU - Baumgartner, Tobias
N1 - Kumulative Dissertation
PY - 2024
Y1 - 2024
N2 - This dissertation explores the advanced domain of human pose estimation (HPE) within sports science, harnessing the latest in computer vision and deep learning to shed light on human motion and biomechanics through sports video analysis. It addresses the critical need for precise and domain-specific 3D HPE methods to accurately capture the complex nuances of human movement, essential for high-performance sports analysis. By leveraging extensive archives of athletic performance footage, this research pioneers a novel approach to biomechanical study, transforming broadcast video into a rich source of sports scientific insight.The work presented aims to break the circular dependency plaguing the field: the need for large, validated datasets for method development versus their scarcity. It employs a multifaceted methodology that spans laboratory studies for kinetic data collection, developing innovative 3D HPE techniques, and creating manual annotation tools for curating large-scale datasets. Focused on running, this comprehensive approach seeks to advance our understanding of human motion within sports, tackling the challenges of accurate 3D pose estimation head-on.The research further delves into the variability of human athletic performance, recognizing both the diversity among athletes and the significant impact of minor biomechanical changes on an individual's performance and injury risk. To address these complexities, a tiered study approach is adopted:1. Running Laboratory Study with 29 Subjects: A detailed analysis within a controlled setting, capturing precise biomechanical data to explore intra-athlete nuances and their implications on a commercial product for estimating running power.2. Video annotation study of 119 Athletes: A broader investigation using manually annotated competitive event footage of the NFL combine to understand diverse movement strategies within the sprint start.3. Fully Automated Study with Tens of Thousands of Images: A large-scale analysis leveraging synthetically generated data to improve computer vision methods for identifying precision of measurement in long-distance running.Collectively, these studies offer a nuanced understanding of human movement, balancing the variability among athletes with the detailed precision required for meaningful sports science research. This dissertation showcases the potential of leveraging large datasets for biomechanical analysis and emphasizes the importance of methodological precision and validation. Through this work, new methodologies and insights are brought to the forefront of human motion analysis, contributing significantly to the fields of sports science, training, and performance enhancement.
AB - This dissertation explores the advanced domain of human pose estimation (HPE) within sports science, harnessing the latest in computer vision and deep learning to shed light on human motion and biomechanics through sports video analysis. It addresses the critical need for precise and domain-specific 3D HPE methods to accurately capture the complex nuances of human movement, essential for high-performance sports analysis. By leveraging extensive archives of athletic performance footage, this research pioneers a novel approach to biomechanical study, transforming broadcast video into a rich source of sports scientific insight.The work presented aims to break the circular dependency plaguing the field: the need for large, validated datasets for method development versus their scarcity. It employs a multifaceted methodology that spans laboratory studies for kinetic data collection, developing innovative 3D HPE techniques, and creating manual annotation tools for curating large-scale datasets. Focused on running, this comprehensive approach seeks to advance our understanding of human motion within sports, tackling the challenges of accurate 3D pose estimation head-on.The research further delves into the variability of human athletic performance, recognizing both the diversity among athletes and the significant impact of minor biomechanical changes on an individual's performance and injury risk. To address these complexities, a tiered study approach is adopted:1. Running Laboratory Study with 29 Subjects: A detailed analysis within a controlled setting, capturing precise biomechanical data to explore intra-athlete nuances and their implications on a commercial product for estimating running power.2. Video annotation study of 119 Athletes: A broader investigation using manually annotated competitive event footage of the NFL combine to understand diverse movement strategies within the sprint start.3. Fully Automated Study with Tens of Thousands of Images: A large-scale analysis leveraging synthetically generated data to improve computer vision methods for identifying precision of measurement in long-distance running.Collectively, these studies offer a nuanced understanding of human movement, balancing the variability among athletes with the detailed precision required for meaningful sports science research. This dissertation showcases the potential of leveraging large datasets for biomechanical analysis and emphasizes the importance of methodological precision and validation. Through this work, new methodologies and insights are brought to the forefront of human motion analysis, contributing significantly to the fields of sports science, training, and performance enhancement.
M3 - Dissertations
BT - Deriving Sport Scientific Insights from Broadcast Footage
PB - Deutsche Sporthochschule Köln
CY - Köln
ER -