Abstract
Soccer has reached a high technical, physical, and tactical level, making data use increasingly common among major clubs. Advances in extracting and processing players’ data have encouraged more teams to explore this information. However, acquiring specialized data remains a challenge due to its complexity and high costs, which limits analysis and research. Publicly available data often include basic information like match results and player lineups, while commercial data, generated manually by individuals watching games, lack consistency. Positional data, crucial for advanced analysis, are typically obtained through expensive wearable GPS devices, limiting access to major soccer clubs. This research aims to propose a low-cost computer vision framework, named Advanced Player Identification and Positioning System, that is useful for extracting player positional data from television broadcast images. This approach enables the creation of advanced datasets, from which essential information for soccer can be extracted, such as tactical formation, for example. The proposed framework was divided into four steps: detection, tracking, identification, and positioning. Experiments were conducted across all stages of the APIPS system using the SoccerNet-GSR dataset, with its manual annotations serving as ground truth. The results indicate that player identification can be improved by the proposed temporal tracking strategy. Furthermore, except for a few outlier cases, the final player positioning error was below 5 meters in 91% of the evaluated instances.