首页|联合滑动窗口和加权评分的多人比赛球员重识别算法

联合滑动窗口和加权评分的多人比赛球员重识别算法

扫码查看
针对多人比赛球员重识别特征提取时硬性水平划分造成图像特征不连续,以及关节点识别错误引入噪声特征等问题,提出一种联合滑动窗口和加权评分的球员重识别算法。首先修改ResNet-50主干网络,并利用空间关系感知注意力获取噪声更少的特征图;然后采用滑动窗口划分策略,使每一水平条带都包含其相邻区域的部分特征,以保持图像特征的连续性;再利用姿态信息提取局部关节点特征,赋予有效特征更高权重并通过多层感知层进行评分,以优化关节点分数;最后采用联合损失函数策略约束网络的学习。在Occluded-Duke,Market-1501和自建数据集上的实验结果表明,在3个数据集上,所提算法的Rank-1指标分别达到60。90%,94。70%和80。35%。
Algorithm for Players Re-Identification in Multiplayer Sports Game Joint Sliding Window and Weighted Scoring
To solve the problems of discontinuity of image features caused by hard horizontal division during the extraction of player re-recognition features in multiplayer matches,and noise features introduced by joint points recognition errors,a joint sliding window and weighted scoring player re-recognition algorithm is proposed.Firstly,the ResNet-50 backbone network is modified and spatial relationships are used to perceive attention to obtain feature maps with less noise.Secondly,the sliding window division strategy is adopted so that each horizontal band contains part of the features of its adjacent area to maintain the continuity of image features.Then,the pose information is used to extract local joint point features,give the effective features higher weight,and score them through multiple perception layers to optimize the joint point score.Finally,the joint loss function strategy is used to constrain the learning of the network.Experiments were carried out on Occluded-Duke,Market-1501 and self-built datasets,and the results showed that the Rank-1 index of the proposed algorithm reached 60.90%,94.70%and 80.35%on the three datasets,respectively.

multiplayer sports gamesliding windowweighted scoringpose informationre-identification

覃海波、雒江涛、许国良

展开 >

重庆邮电大学通信与信息工程学院 重庆 400065

多人比赛 滑动窗口 加权评分 姿态信息 重识别

2024

计算机辅助设计与图形学学报
中国计算机学会

计算机辅助设计与图形学学报

CSTPCD北大核心
影响因子:0.892
ISSN:1003-9775
年,卷(期):2024.36(11)