投审稿平台
投稿指南
下载专区
地  址:北京市海淀区中关村科学院
南路6号中国科学院计算所342号 [地图]
《计算机辅助设计与图形学学报》编辑部
邮政编码:100190
电  话:010-62562491
     010-62600342
订阅信息
ISSN   1003-9775
CN    11-2925/TP
邮发代号:82-456
单  价:80.00元
全年订价:960.00元
订阅电话:010-64017032
在线期刊

基于自适应关键点热图的遮挡篮球运动员检测算法

任媛1), 雒江涛2)*, 梁旭鹏3)
1) (重庆邮电大学通信与信息工程学院 重庆 400065) 2) (重庆邮电大学电子信息与网络工程研究院 重庆 400065)3) (重庆邮电大学体育学院 重庆 400065)
分类号: TP391.41 DOI: 10.3724/SP.J.1089.2021.18707
出版年,卷(期):页码: 2021 , 33 ( 9 ): 1450-1456 任媛
摘要: 运动员检测是篮球运动智能化分析的基础, 由于篮球视频存在场景复杂、目标运动快速、目标间遮挡严重的问题, 现有目标检测技术不能实现对密集遮挡运动员的精确检测. 为此, 提出一种基于自适应关键点热图的遮挡篮球运动员检测算法. 首先通过预先构建的全卷积编码-解码网络进行运动员特征提取, 利用高斯核函数在特征图上渲染关键点热图, 热图的渲染采用自适应策略, 高斯核半径随着目标宽和高的变化而变化, 能够加快网络收敛; 然后在热图中提取运动员中心点, 回归得到运动员宽高、位置等信息, 省去了基于锚框检测中复杂耗时的后处理过程, 更利于在遮挡条件下区分2个运动员. 在篮球运动数据集BasketballPlayer上进行实验的结果表明, 在复杂篮球视频场景下, 该算法能有效地解决密集遮挡运动员之间漏检、误检和检测精度不高的问题, 处理速度可达到26帧/s.
关键词: 运动员检测; 自适应关键点热图; 可变高斯核; 密集遮挡
Algorithm for Detecting Occluded Basketball Players Based on Adaptive Keypoint
Ren Yuan1), Luo Jiangtao2)*, and Liang Xupeng3)
1) (School of Communications and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065) 2) (Electronic Information and Networking Research Institute, Chongqing University of Posts and Telecommunications, Chongqing 400065) 3) (School of Physical Education, Chongqing University of Posts and Telecommunications, Chongqing 400065)
abstract: Player detection is the basis of intelligent analysis of basketball events. Due to the complexity of basketball sports, such as fast player movement and serious occlusion of players, existing object detection techniques cannot afford to achieve accurate detection. To address this issue, an object detection algorithm of basketball player is proposed based on an adaptive keypoint heatmap. First, fully-convolutional encoder-decoder networks are built to extract feature of players. Then, the keypoint heatmap is rendered in feature map through an adaptive variable Gaussian kernel radius. The rendering of the keypoint heatmap adopts an adaptive strategy, in which the Gaussian kernel radius is changed with the width and height of the object, so that the network convergence can be accelerated. With the center point of the player extracted from heatmap, the player’s size, position and other information are retrieved through regression. Since the player is detected based on the center point, eliminating the complicated and time-consuming post-processing procedure in anchor-based detections, it is more conducive to distinguish two different players under occlusion conditions. The effectiveness of the proposed approach is validated on the BasketballPlayer dataset, and the experimental results show that lots of missed detection, misdetection, and low detection accuracy among densely occluded players are significantly improved at the processing rate of 26 frames per second, compared with existing algorithms.
keyword: player detection; adaptive keypoint heatmap; variable Gaussian kernel; dense occlusion
 
Copyright © 2004《计算机辅助设计与图形学学报》版权所有
电话:010-62600342 传真:010-62562491
E_mail:jcad@ict.ac.cn