<p>Currently, the calibration of basketball game data is shifting from manual observation to intelligent recognition. However, existing methods for detecting small and fast-moving targets still suffer from low accuracy. To address this, this study proposes an improved YOLOv7 detection model based on a Hierarchical Attention Fusion (HAF) module, and conducts experiments on both object detection and basketball goal recognition. Experimental results demonstrate that in object detection tasks, the proposed model reduces memory usage by 15% and increases single-image processing speed by 7.7% compared to the baseline. Its accuracy and mean Average Precision (mAP) reach 88.2% and 89%, respectively, representing average improvements of 27.31% and 28.24% over other detection methods. On general datasets such as COCO2017 and VOC2012, the model further improves mAP by 19.7%, increases recall by 7%, while reducing the number of parameters by an average of 57.5%. Moreover, in basketball goal recognition, the dual-camera configuration improves the recognition accuracy by 13.2% compared to a single-camera, with the false detection rate and missed detection rate decreasing by 9.1% and 14.1%, respectively. The results indicate that the proposed basketball goal recognition model not only enhances object detection precision but also effectively improves the recognition accuracy in goal scenes, thereby promoting the intelligent development of basketball game outcome determination.</p>