Abstract
Accurate and timely detection of basketballs is crucial for ensuring fairness in games, enhancing the precision of data analysis, optimizing tactical planning for coaches, and improving the spectator experience. However, current basketball detection technologies face challenges such as variations in target scale, scene complexity, and changing camera angles, which limit automated systems' accuracy and real-time performance. To address these issues, this study introduces a novel real-time basketball detection model, BGS-YOLO, incorporating several key innovations. First, the model integrates a BiFPN (Bidirectional Feature Pyramid Network) that enhances detection accuracy by efficiently merging feature maps across different resolutions, allowing for more effective feature extraction from basketball targets. Second, the Global Attention Mechanism (GAM) dynamically adjusts the model's focus, optimizing feature attention in complex or partially occluded scenes, boosting recall in occluded scenarios by 3.2%, thereby improving localization precision. Finally, SimAM-C2f increases the model's robustness in high-interference environments by calculating similarity features between the target and the background, reducing false positives by 15%, ensuring more reliable detection. Experimental results show that BGS-YOLO surpasses existing models across key metrics such as precision, recall, F1 score, and mean average precision (mAP), achieving a mAP of 93.2%. All improvements were statistically significant (p < 0.001). These advancements significantly enhance the accuracy and robustness of basketball detection, offering valuable technical support for intelligent sports analytics.