Single player tracking in multiple sports videos

• Main Authors: Petilla, Carlos Anthony B., Yap, Gary Daniel G., Yuson, Patrick Laurence L., Zheng, Nathaniel Y.
• Format: text
• Language: English
• Published: Animo Repository 2015
• Online Access: https://animorepository.dlsu.edu.ph/etd_bachelors/10962
• Institution: De La Salle University

Vision-based analysis of sports videos provides a consistent and concise performance statistics which are essential for developing athlete training programs. Existing applications such as Sport VU [8] and SAGIT [17], are vision-based systems for analyzing player performance which requires precise and rigid setup of equipment. This limits the application to large organizations. In order to provide the same service for minor league coaches, the objective of the study is to develop a tracking system capable of adapting multiple camera from any vantage point for tracking. The developed system tracks a selected player using Speeded-Up Robust Features (SURF) with Symmetric Nearest Neighbor Filtering, and Kalman filter for object tracking. Using the spatial overlap metric, the average tracking accuracy of the system is 40% with a tracking precision of 74% in basketball clips wherein tracked player is un-occluded. On the other hand, the tracking accuracy during occlusion significantly drops due to the implemented motion models even with the use of Kalman Filter as a motion estimator. With the use of multiple cameras, cross-correlation is applied on the translated 2D coordinates of the target to synchronize location information from each camera. The synchronized data are averaged together into a single representation to decrease the location error. Improvement after data averaging is dependent on the availability of reliable data. A reliable data with a tracking accuracy of 59.8% and tracking precision of 93.8% could improve location data by as much as 70%