Basketball Object Detection in Complex Occlusion Scenarios Based on an Attention Gating Mechanism

Abstract

With the development of intelligent sports analysis and game video understanding, issues like small size, fast movement, and frequent occlusion of basketball targets in complex occluded scenarios make traditional detection models prone to missed detections and positioning errors. In this study, based on YOLOv8, we built a basketball target detection model using an attention-gated mechanism. We introduced a lightweight multi-scale feature extraction module in the backbone network to enhance the texture representation of small targets. In the neck network, we designed a channel–spatial joint attention-gated module to adaptively suppress interference from players' bodies, arms, and background areas. At the same time, by combining deformable receptive field aggregation with cross-scale gated fusion strategies, the model's ability to recognize occluded basketball edges, blurry contours, and partially visible areas is improved. Finally, we use a decoupled detection head and an occlusion-sensitive loss function to optimize classification and localization accuracy.