Shuttlecock Movement Detection and Evaluation Method Based on Improved YOLOv8 and Image Segmentation

Abstract

In view of the problems existing in badminton sports analysis such as strong subjective judgment, difficulty in quantification, and separation of correlation analysis between action standardization and tactical effectiveness, this study proposes an improved YOLOv8 using the LAFF module. The network combines the shuttlecock intelligent detection and evaluation method of DBAG-SN, and builds an intelligent evaluation model for multi-source information fusion on this basis. First, the LAFF module and shape-aware loss are introduced in the detection network to enhance the ability to capture small targets. DBAG-SN is designed and combined with bone key point detection and action recognition technology to achieve refined analysis of key technical action links such as racquet introduction, swing, hitting, and follow-up swing. At the same time, multi-view geometric constraints are used to reconstruct the 3D trajectory, and the relationship between technical actions and tactical effects is established through the attention fusion network. Experiments show that in the self-built shuttlecock metric dataset, the improved model has a maximum mAP of 92.4% when the threshold is 0.5, an average increase of 7.1% compared with the baseline model before improvement, and the positioning accuracy improvement under high IoU thresholds is more significant. In the image segmentation task, the average IoU of DBAG-SN is 85.2%, which is better than mainstream algorithms such as Mask Region-Convolutional Neural Network. In actual application verification, the 3D trajectory reconstruction RMSEpos of this intelligent fusion evaluation model is reduced to 18.7 mm, and the Pearson correlation coefficient between the comprehensive score generated and the expert score is as high as 0.91. Compared with the single-dimensional evaluation method, the dual-stream fusion mechanism that combines action standardization and tactical effectiveness enables the evaluation's comprehensive evaluation index F1-Score to reach 85.5%, which is 12.6% and 17.7% higher than single action evaluation and single tactical evaluation. The main contribution of this study is to propose a closed-loop evaluation framework from multimodal perception to causal decision-making of "technical tactics". Through the dual stream fusion mechanism of action and landing point, it effectively solves the problem of separation between action and tactics in training analysis, and can provide objective data support for scientific badminton training.