基于改进沙漏的攀岩运动关键点检测算法

Rock climbing keypoint detection algorithm based on improved hourglass

针对关键点检测中目标尺度多变以及不同特征适应性等难题,为进一步提升现有的姿态估计方法在实现姿态估计任务时的性能,验证单阶段和多阶段姿态估计方法各自的有效性,提出一种基于改进沙漏的攀岩运动关键点检测算法。首先设计一个多路池化残差结构,改善由于沙漏网络多次上下采样带来的信息损失和上下文信息提取不足的局限性,提升浅层特征在关键点检测中的表现;其次在沙漏网络中引入沙漏注意力结构,通过利用特征映射将输入信息划分为不同大小的特征块序列,在特征编码和特征解码两个过程中,充分挖掘图像有效信息,使得在特征匹配过程中不仅考虑本身的拟合程度,更考虑到关节位置之间的关联信息。实验表明,提出的算法在公开数据集MPII、COCO和针对攀岩运动的数据集上表现良好,且算法泛化能力较好,能够应用于多种运动场景中的人体关键点检测任务。

In view of the variable target scales and adaptability for different features in keypoint detection,a rock climbing keypoint detection algorithm based on improved hourglass is proposed in order to further improve the performance of the existing attitude estimation methods during the process of achieving attitude estimation tasks and verify the effectiveness of single-stage and multi-stage attitude estimation methods.A multi-channel pooling residual structure is designed to eliminate the information loss caused by multiple up-samplings and down-samplings of the hourglass network and the limitations of insufficient context information extraction,and improve the performance of shallow features in keypoint detection.An hourglass attention structure is introduced into the hourglass network.The input information is divided into feature block sequences of different sizes by feature maps.The effective information of the image is fully exploited in the two processes of feature encoding and feature decoding,so that not only the fitting degree of itself is considered,but also the correlation information between joint positions is considered in the process of feature matching.The experiments show that the proposed algorithm performs well on the public data sets MPII and COCO and the data sets for rock climbing,and the algorithm is of good generalization ability,so it can be applied to the tasks of human keypoint detection in a variety of sports scenes.