The NCAA first recognized Beach Volleyball as a championship sport in 2016, and it has since become the fastest growing NCAA sport in history. Despite this recent growth, there is limited research, media exposure, and official online training resources dedicated to beach volleyball. Florida State’s beach volleyball team is one of only two programs to appear in every NCAA championship since 2016 and is the only program in the country to have won 4 straight conference titles (2016-2019). FSU beach volleyball is in a unique position for athletic research, as it is one of the top current training programs in a sport with a lot of potential for further development. We are two athletes on FSU’s beach volleyball team and students at the FAMU-FSU College of Engineering. Our combined passion for volleyball and technology has inspired us to examine the ways technology is used to analyze other sports, and how it might be applied on the sand.
The goal of the current project is two-fold. First, to train an object detection or pose estimation model to detect beach volleyball athletes moving in the sand in real time with high accuracy. Second, to write corresponding algorithms that analyze various player movements. Currently, there two primary types of models we plan to compare to accomplish these goals. The first would be a simple object detection algorithm that could quickly detect where players are on the court and chart movements. More specifically, picture the sports analyst tools that map projected movements of a football play. We would devise something similar that could examine clips and draw lines of motion. This would be easy to train and effective in detecting player locations but would provide limited information regarding play analysis and body movements.
Another option we hope to explore is a program that uses vectors and pose estimation to examine player movements in greater detail. This type of program would be more complex to implement, and its accuracy in this application is uncertain, but it could potentially be used to analyze film and predict specific plays based on a player’s position. For instance, determining whether a player’s shoulders are lined up to hit straight or across their body. This has the could expand the work’s statistical applications by examining movements from different perspectives.
This research is inspired by previous research on YOLOv3, an open-source machine learning software used to identify specific objects found in images and videos. YOLO stands for “You Only Look Once,” a nod to the model’s speed compared to others of its kind, as it only scans a frame once to check for any identifiable objects. Throughout its uses in a past UROP project and IDEA Grant, the YOLOv3 model has demonstrated remarkable accuracy in identifying people, boats, marine life, and other objects from drone footage. However, since beach volleyball video analysis occurs after training and doesn’t need to harness YOLO’s real time capabilities, we plan to research related object detection and pose estimation algorithms models that utilize similar training methods to determine the best fit for this project.
While YOLO may be the fastest algorithm, this project wouldn’t need to make real-time detections and instead would prioritize accuracy. We have high confidence in the ability of these models to identify beach volleyball players on the court, which leads to the primary question of how to develop algorithms that can utilize detections to map player movements in real time and offer suggestions for more effective defensive movements during a play. This is especially relevant to beach volleyball training, as moving in sand poses a significant challenge and effective movements make a significant impact on a team’s success. Ultimately, we hope that our work will contribute towards both the sport of beach volleyball and the athletic applications of machine learning.