Developing an AI-based assistant system for Paris Olympics live broadcasts: swimming and fencing

Abstract

Recent advances in computer vision and real-time graphics have enabled 
significant improvements in augmented live sports broadcasts. However, 
achieving the millisecond-level latency required for Olympic events remains 
challenging. In this paper, we introduce an AI-powered broadcast assistant 
successfully deployed during the Paris 2024 Olympic swimming and fencing 
competitions. Leveraging a custom-built dataset comprising 520k annotated 
frames, we developed two specialized deep-learning pipelines: (i) fine grained recognition and enhancement of signal lights in fencing, and (ii) high-angle scene understanding and lane recognition for swimming events. In fencing broadcasts, our assistant effectively magnifies small indicator lamps signalling valid touches. For swimming competitions, the system automatically detects overhead camera angles, segments individual lanes, and overlays national flags in real time. We present comprehensive overview of the system architecture, dataset creation methods, model architectures, and on-site deployment strategies. Live operational trials demonstrated significant improvements in broadcast efficiency, supported by an integrated human-in-the-loop approach to maintain editorial control. Finally, we investigate the applicability of our real-time scene classification and automatic CG cut-out methods to a broader range of sports, paving the way for more advanced AI-assisted broadcast systems.