The application of Artificial Intelligence (AI) on player and ball tracking in sports video is of great significance in analyzing performance, evaluating tactics, and enriching broadcast for team sports. However, the large repertoire of models available and the complexity related to real-time constraints make model choice difficult. This paper presents a comparative evaluation of nine state-of-the-art detection and tracking models developed for real-time sports analytics. The strategies to be compared contain a mixture of detection architectures, from convolutions with attention, to novel tracking approaches like graph-based modeling, filtering approaches, and transformer-based predictors. The evaluation focuses on four key criteria: tracking accuracy, identity consistency, inference speed, and robustness to occlusion and rapid motion. Results show that models combining enhanced detection modules with context-aware tracking deliver superior performance. In particular, the attention-based detection model coupled with a Kalman filter achieves the highest tracking accuracy for both players and the ball, while graph-based and transformer-based methods offer an effective balance between speed and precision. In contrast, older model combinations perform poorly under realistic game conditions. Overall, the findings provide practical guidance for selecting robust and efficient real-time tracking systems suitable for applications in sports broadcasting, automated analysis, and augmented coaching tools.