<p>Multi-object tracking (MOT) is crucial for multi-agent analysis activities such as tactics, player movements, and performance in team sports. However, frequent player occlusions and visual similarity caused by team uniforms pose challenges, often resulting in identity switches. Conventional tracking-by-detection methods still struggle to address these issues effectively. To address these challenges, we propose a novel pose-based virtual marker (VM) MOT method for team sports, named Sports-vmTracking. First, we constructed two open-source 3x3 basketball pose datasets using active learning to efficiently annotate informative samples. Then, we overlaid the VMs on video frames to identify players and extract their poses with unique IDs. Finally, these poses were converted into bounding boxes for comparison with automated MOT methods. Experimental results on our 3x3 basketball datasets demonstrate the high effectiveness of the proposed VM configuration, enabling automatic validation on the pre-annotated datasets. Our approach achieved a mean HOTA score of 72.3%, over ten points higher than other state-of-the-art methods without VM, resulting in zero ID switches on the indoor dataset and a HOTA score of 42.5% on the outdoor dataset, outperforming all baseline methods. In addition to improved robustness against occlusions and appearance similarity, the proposed framework reduces the time and cost of manual annotation, enabling efficient and scalable tracking in sports analytics.</p>