Aplicação e análise de algoritmos de rastreamento visual para detecção de bola em partidas de voleibol

Abstract

The advancement of technologies, such as Computer Vision, has significantly transformed the way we watch and analyze sports events, optimizing internal processes like decision-making by referees and strategy development by coaches. In the context of volleyball, a sport characterized by high dynamics and tactical complexity, the application of ball tracking techniques faces particular challenges due to the small size of the object, its high speed, and frequent occlusions caused by players. This study aims to evaluate and compare the performance of different visual tracking algorithms, such as CSRT, KCF, MOSSE, MIL, TLD, MedianFlow, and Boosting, in volleyball game videos, with the goal of identifying the most effective approaches for real-time ball detection and trajectory tracking. Additionally, a YOLO model is trained to investigate how the use of Deep Neural Networks can contribute to the precise and efficient detection of the ball, considering the specific challenges of volleyball. The evaluation of the algorithms is based on metrics such as Intersection over Union to measure tracking accuracy, Frames Per Second to assess real-time performance, and Mean Tracking Error to quantify deviations in the ball’s position estimation. Although the evaluated algorithms exhibit limitations regarding the tested dataset, these restrictions are related to the complexity of the professional match videos obtained from the YouTube channel Volleyball Playoffs, which are used to construct the image set and include variations in lighting, low resolution, and rapid ball movement. To complement the analysis and assess the performance of the algorithms under more controlled conditions, tests are conducted with videos recorded in a controlled environment, using colored balls in green, orange, and blue to facilitate the detection and tracking of the trajectory. The comparison between the results obtained with the dataset created from professional match videos and the tests conducted with the controlled environment video set demonstrates that controlled conditions contribute to better algorithm performance. These findings provide a preliminary foundation that can support future research aimed at developing integrated systems to assist refereeing and sports performance analysis