Scoring smarter: deep learning-based basketball scoring detection in real time

Abstract

Basketball scoring detection is challenged by scene complexity, varying camera angles, and occlusions. This paper presents a real-time basketball scoring system that combines you only look once version 8 (YOLOv8) for hoop detection, a vision transformer (ViT) for spatiotemporal motion modeling, and deep simple online and realtime tracking (DeepSORT) for object tracking, in order to overcome the accuracy loss that classical systems suffer during dynamic gameplay. The system was evaluated on a custom dataset of 4,000 annotated images and five full-length broadcast videos containing 44 verified scoring events. The proposed YOLOv8+ViT+DeepSORT model achieved 96.84% average precision at 50% intersection over union (AP50), 95.2% precision, 93.8% recall, 94.5% F1-score, and 94.55% accuracy, while sustaining 85 frames per second (FPS). Comparison against YOLOv3 with frame differencing (baseline), you only look once version 8 nano (YOLOv8n) with bidirectional feature pyramid network (BiFPN) and global attention module (GAM) attention mechanisms, and BiFPN, GAM, and SimC2f-YOLO (BGS-YOLO) confirms the best overall balance across all indicators. ViT captures the temporal dynamics of motion while DeepSORT preserves object identity across frames. The framework is therefore well suited to sports analytics, automated highlight generation, and referee assistance systems where accuracy and real-time performance are essential.