在足球赛事中,对运动员和球实施目标识别,能够为自动化的跟踪与拍摄系统提供必要的算法支撑。针对传统技术在识别球员和足球方面准确度不足的情况,本文介绍了一种改进的检测方法,该方法结合了ECA注意力机制和BiFPN (即加权双向特征金字...在足球赛事中,对运动员和球实施目标识别,能够为自动化的跟踪与拍摄系统提供必要的算法支撑。针对传统技术在识别球员和足球方面准确度不足的情况,本文介绍了一种改进的检测方法,该方法结合了ECA注意力机制和BiFPN (即加权双向特征金字塔网络),简记为ECABiF-Y5n。此模型优化了ECASPPF组件,以增强局部特征表达,并改善小物体的识别效率;同时利用BiFPN来整合多层级的特征信息,生成更为有效的特征描述。实验在SoccerNet_v3_H250数据集上进行,结果表明,相较于原作者采用的YOLOv8n,整体类别的精度提升了4.4%,召回率增加了2.3%,mAP50增长了3.7%,而mAP0-95则提高了0.2%。特别是对于较难识别的“球”类别,精度提高了6.1%,召回率上升了5.9%,mAP50增进了7.6%,mAP0-95也有了2.8%的进步。这些实验对比结果证明,ECABiF-Y5n在提升检测准确性方面表现出色,特别增强了对足球赛事中小型目标的辨识能力。In football matches, implementing object recognition for athletes and the ball can provide essential algorithmic support for automated tracking and filming systems. To address the issue of insufficient accuracy in traditional techniques when identifying players and the football, this paper introduces an improved detection method that combines ECA attention mechanisms with BiFPN (weighted bidirectional feature pyramid network), abbreviated as ECABiF-Y5n. This model optimizes the ECASPPF component to enhance local feature representation and improve the efficiency of small object recognition;it also employs BiFPN to integrate multi-level feature information, generating more effective feature descriptions. Experiments were conducted on the SoccerNet_v3_H250 dataset, showing that compared to YOLOv8n used by the original authors, the overall category precision increased by 4.4%, recall rose by 2.3%, mAP50 grew by 3.7%, and mAP0-95 improved by 0.2%. Specifically, for the “ball” category, which is more challenging to detect, precision improved by 6.1%, recall increased by 5.9%, mAP50 by 7.6%, and mAP0-95 by 2.8%. These comparative experimental results demonstrate that ECABiF-Y5n excels in enhancing detection accuracy, particularly strengthening the identification capability for small objects in football games.展开更多
文摘在足球赛事中,对运动员和球实施目标识别,能够为自动化的跟踪与拍摄系统提供必要的算法支撑。针对传统技术在识别球员和足球方面准确度不足的情况,本文介绍了一种改进的检测方法,该方法结合了ECA注意力机制和BiFPN (即加权双向特征金字塔网络),简记为ECABiF-Y5n。此模型优化了ECASPPF组件,以增强局部特征表达,并改善小物体的识别效率;同时利用BiFPN来整合多层级的特征信息,生成更为有效的特征描述。实验在SoccerNet_v3_H250数据集上进行,结果表明,相较于原作者采用的YOLOv8n,整体类别的精度提升了4.4%,召回率增加了2.3%,mAP50增长了3.7%,而mAP0-95则提高了0.2%。特别是对于较难识别的“球”类别,精度提高了6.1%,召回率上升了5.9%,mAP50增进了7.6%,mAP0-95也有了2.8%的进步。这些实验对比结果证明,ECABiF-Y5n在提升检测准确性方面表现出色,特别增强了对足球赛事中小型目标的辨识能力。In football matches, implementing object recognition for athletes and the ball can provide essential algorithmic support for automated tracking and filming systems. To address the issue of insufficient accuracy in traditional techniques when identifying players and the football, this paper introduces an improved detection method that combines ECA attention mechanisms with BiFPN (weighted bidirectional feature pyramid network), abbreviated as ECABiF-Y5n. This model optimizes the ECASPPF component to enhance local feature representation and improve the efficiency of small object recognition;it also employs BiFPN to integrate multi-level feature information, generating more effective feature descriptions. Experiments were conducted on the SoccerNet_v3_H250 dataset, showing that compared to YOLOv8n used by the original authors, the overall category precision increased by 4.4%, recall rose by 2.3%, mAP50 grew by 3.7%, and mAP0-95 improved by 0.2%. Specifically, for the “ball” category, which is more challenging to detect, precision improved by 6.1%, recall increased by 5.9%, mAP50 by 7.6%, and mAP0-95 by 2.8%. These comparative experimental results demonstrate that ECABiF-Y5n excels in enhancing detection accuracy, particularly strengthening the identification capability for small objects in football games.
