文献类型：期刊文献

中文题名：基于X3D特征和语义融合的篮球运动员检测跟踪方法

英文题名：Basketball Player Detection and Tracking Method Based X3D Feature and Semantic Fusion

作者：韩乾乾[1];顾华宁[2];金湓[3];赵永强[4]

机构：[1]广东海洋大学体育与休闲学院,广东湛江524088;[2]成都理工大学管理科学学院,四川成都610059;[3]西北民族大学化工学院,甘肃兰州730030;[4]广西师范大学基建处,广西桂林54100

年份：2025

卷号：48

期号：6

起止页码：101

中文期刊名：南京师大学报(自然科学版)

外文期刊名：Journal of Nanjing Normal University(Natural Science Edition)

收录：北大核心2023、、北大核心

基金：国家自然科学基金资助项目(42402278);甘肃省科技计划资助项目(23YFGA0073);2023年度广东省教育科学规划课题(高等教育专项)资助项目(2023GXJK297).

语种：中文

中文关键词：篮球视频;检测跟踪;X3D网络;单应性稳像;特征融合

外文关键词：basketball video;detection and tracking;X3D network;homography stabilization;feature fusion

中文摘要：针对单机位篮球视频中运动员频繁遮挡、队服相似及相机抖动导致的跟踪难题,本文提出融合X3D时空特征与多模态语义的实时检测跟踪方法.首先,采用单应性稳像将直播流配准至参考平面,抑制背景漂移.其次,利用轻量化X3D网络在16帧片段上提取1024维时空描述子(5 GFLOPs算力约束),捕获各种篮球比赛的关键动作模式,同时满足了边缘部署的延迟要求;最后,设计注意力驱动的特征融合模块,自适应结合几何位移、外观直方图与X3D特征.在NBA-SYN和UCF-Sports-Basket公开数据集上的实验表明,该方法分别达到77.43%和79.30%的MOTA,以45.31 FPS的实时性能,显著优于现有方案,为有效硬件条件下的篮球视频分析提供可靠技术支撑.

外文摘要：In order to solve the tracking problems caused by frequent occlusion of players,similar uniforms and camera shake in single-camera basketball videos,this paper proposes a real-time detection and tracking method that integrates X3D spatiotemporal features and multimodal semantics.First,homography stabilization is used to align the live stream to the reference plane to suppress background drift.Secondly,a lightweight X3D network is used to extract 1024-dimensional spatiotemporal descriptors(5 GFLOPs computing power constraint)on 16-frame segments to capture the key action patterns of various basketball games while meeting the latency requirements of edge deployment;finally,an attention-driven feature fusion module is designed to adaptively combine geometric displacement,appearance histogram and X3D features.Experiments on the NBA-SYN and UCF-Sports-Basket public datasets show that this method achieves 77.43%and 79.30%MOTA respectively,with a real-time performance of 45.31 FPS,which is significantly better than the existing solutions,providing reliable technical support for basketball video analysis under effective hardware conditions.