Abstract: Reconstructing the movement trajectories of basketball players during matches plays a significant role in analyzing game dynamics and optimizing team tactics. While multi-camera systems capture player movements from various perspectives, selecting the optimal observation viewpoint is crucial for accurate player trajectory reconstruction. Existing single-view methods face inherent limitations, such as restricted field of view, identity confusion, and occlusion-induced blind spots. To address these issues, This paper proposes an automatic optimal observation viewpoint selection method. Firstly, cross-camera re-identification is employed to verify the feasibility of consistently tracking the same player across views. Secondly, we improve the Multi-Instance Learning (MIL) framework by integrating a Loss Attention Mechanism (LAM) to enhance candidate viewpoint filtering through loss function optimization. Finally, a scoring evaluation mechanism is established to select the most suitable tracking viewpoint. Experimental results on the public APIDIS dataset demonstrate that our method outperforms mainstream single-view tracking techniques such as ByteTrack by achieving approximately 10% higher HOTA and 30% improvement in IDF1. Comparable results are obtained on an additional dataset, confirming the robustness of the approach. The proposed model not only mitigates ID confusion and occlusion challenges common in single-view systems but also provides a foundation for reconstructing global player trajectories. This advancement offers practical value for technical-tactical analysis in sports science.