Shaanxi Normal University, Xi'an, Shaanxi 710000, China.
Comput Intell Neurosci. 2022 Apr 1;2022:4727375. doi: 10.1155/2022/4727375. eCollection 2022.
With the explosive growth of the number of sports videos, the traditional sports video analysis method based on manual annotation has been difficult to meet the growing demand because of its high cost and many limitations. The traditional model is usually based on the target detection algorithm of manual features, and the detection of human posture features is not accurate. Compared with global image features such as line features, texture features and structure features, local image features have the characteristics of rich quantity in the image, low correlation between features, and will not affect the detection and matching of other features due to the disappearance of some features in the case of occlusion. Referring to the practice of Deep-ID network considering both local and global features, this paper adjusts the traditional neural network, and combines the improved neural network with the human joint model to form a human pose detection method based on graph neural network, and then applies the algorithm to multiperson human pose estimation. The results of several groups of comparative experiments show that the algorithm can better estimate the human posture in sports competition video, and has a good performance in solving multiperson pose estimation in sports game video.
随着体育视频数量的爆炸式增长，基于人工标注的传统体育视频分析方法由于成本高、局限性大，已经难以满足日益增长的需求。传统模型通常基于人工特征的目标检测算法，对人体姿态特征的检测不够准确。与线特征、纹理特征和结构特征等全局图像特征相比，局部图像特征在图像中具有数量丰富、特征之间相关性低的特点，并且在某些特征因遮挡而消失的情况下，不会影响其他特征的检测和匹配。参考 Deep-ID 网络同时考虑局部和全局特征的实践，本文调整了传统神经网络，并将改进后的神经网络与人关节模型相结合，形成基于图神经网络的人体姿态检测方法，然后将该算法应用于多人人体姿态估计。几组对比实验的结果表明，该算法可以更好地估计体育竞赛视频中的人体姿态，在解决体育比赛视频中的多人姿态估计问题方面具有良好的性能。
