Yolov7-pose识别运动员姿态时关键点异常的消除方法  

METHOD FOR ELIMINATING ABNORMAL KEY POINTS IN ATHLETE POSTURE RECOGNITION USING YOLOV7-POSE

作　　者：黄建浩 钟映春[1] 张钢[1] 赖志飞 杨铠康 张永恒 HUANG Jianhao;ZHONG Yingchun;ZHANG Gang;LAI Zhifei;YANG Kaikang;ZHANG Yongheng(Automation School,Guangdong University of Technology,Guangzhou,Guangdong 510006,China;Guangzhou Panyu Polytechnic,Guangzhou,Guangdong 511483,China;Yundongjia Science and Technology Ltd,Shenzhen,Guangdong 518115,China)

机构地区：[1]广东工业大学自动化学院,广东广州510006 [2]广州市番禺职业技术学院,广东广州511483 [3]深圳云动家体育科技有限公司,广东深圳518115

出　　处：《井冈山大学学报（自然科学版）》2024年第3期71-78,共8页Journal of Jinggangshan University (Natural Science)

基　　金：国家自然科学基金项目(61975248)。

摘　　要：本研究首先构建了深度聚合网络(deep layer aggregation,DLA),用于替换Yolov7-pose算法中原有的特征融合网络,增强了特征的表达;其次,将高效聚合网络(efficient layer aggregation Net, ELAN)和尺度感知注意力网络(scale-aware attention Net, SAAN)融合在一起,引入到Yolov7-pose算法中;然后,引入尺度因子到Yolov7-pose算法的损失函数中,并通过实验确定尺度因子最佳值。实验结果表明,经过本法改进的方法识别篮球比赛视频流中的多名运动员姿态时,不再出现关键点异常现象,姿态识别的精度达到95%以上,运行效率达到24 FPS以上,综合性能显著优于原始Yolov7-pose算法等算法。This paper proposes several enhancements to the original Yolov7-pose algorithm:firstly,constructing a deep layer aggregation network(DLA)to replace the existing feature fusion network and enhance feature expression;secondly,fusing ELAN(efficient layer aggregation net)and SAAN(scale-aware attention net)into the Yolov7-pose algorithm;thirdly,introducing a scale factor into the loss function of the Yolov7-pose algorithm and determining its optimal value through experiments.Experimental results demonstrate that these improvements eliminate key point anomalies in posture recognition,achieving an accuracy rate exceeding 95%and operating efficiency surpassing 24FPS.The comprehensive performance surpasses that of the original Yolov7-pose algorithm and other algorithms significantly.

关 键 词：篮球比赛 姿态检测 深层聚合网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]