Towards Autonomous Sports Production: A Multimodal Pipeline for Multi-Angle Basketball Highlight Generation

Sammanfattning

In multi-view sports broadcasting, extracting high-quality highlights and selecting the optimal camera angle from concurrent video streams is traditionally a labor-intensive process with significant computational overhead. This thesis explores the feasibility of automating the production of multi-view sports highlights using Multimodal Large Language Models (MLLMs). To manage the high computational demands of multi-camera environments, we investigate a hierarchical data-reduction strategy that progressively filters temporal and spatial redundancies. The research focuses on evaluating the effectiveness of vision-language alignment for precise action localization, comparing traditional video understanding architectures against emergent foundation models. Furthermore, we assess the capability of MLLMs to perform high-level decision-making in optimal view selection. By conducting a performance analysis on real-world multi-view datasets, this study validates the potential for multimodal architectures to achieve human-like directing logic with significantly reduced manual intervention and computational cost.