A Real-time Recognition Algorithm for Basketball Referee Gestures Based on Lightweight Attention Mechanism

In basketball games, accurate and efficient hand gesture recognition is critical for referee decision-making and automated systems that assist in rule enforcement, making the need for lightweight algorithms capable of operating in real-time under resource constraints increasingly important. To meet the high precision and low resource consumption requirements of the hand gesture recognition system in basketball games, this paper proposes a lightweight hand gesture recognition algorithm based on the attention mechanism. The algorithm is based on the YOLOv5s model and introduces the Involution operator, Coordinate Attention (CA), lightweight content-aware upsampling operator, and Ghost-Net with SiLU activation function-based model pruning to effectively reduce the computational complexity and model parameters while maintaining the recognition performance. The experimental results show that the proposed algorithm significantly reduces the computation, parameters, and model size on the self-built basketball referee hand gesture dataset, and can meet the actual application requirements in terms of detection accuracy and frame rate.