一种基于YOLOv5改进的动捕Maker点识别方法

An Improved YOLOv5-based Method for Identifying Maker Points in Motion Capture

针对动作捕捉系统应用时,复杂背景下球形目标Marker点识别稳定性低的问题,论文结合YOLOv5识别技术提出了一种半监督训练的Marker点识别方法:首先利用Teacher student model进行伪标签制作,并根据所得的阈值来挑选高置信度的伪标签;随后将伪标签数据当做训练集对模型进行微调来得到最终模型;接下来对YOLOv5模型中CSP模块中的residual模块进行替换,同时对模型进行剪枝,并利用动捕小目标识别特征删减YOLOv5算法中的模块,从而进一步实现模型轻量化。最后,使用球形Marker进行测试,实验结果验证了所提方法的有效性,证实了应用论文方法可以较为稳定地提取复杂场景中的Marker点,大大增强了动捕系统在复杂环境中应用的适应能力。

To address the problem of low stability of Marker point recognition of spherical targets in complex backgrounds in motion capture system applications,this paper proposes a semi-supervised training method for Marker point recognition combined with YOLOv5 recognition technology.Firstly,the Teacher student model is used for pseudo-label production,and pseudo-labels with high confidence are selected according to the resulting thresholds.Subsequently,the pseudo-label data is used as a training set to fine-tune the model to get the final model.Next,residual module in the CSP module of the YOLOv5 model is replaced,pruning is performed on the model,and the modules in the YOLOv5 algorithm are deleted using the motion-capture small target recognition feature to lighten the model.Finally,spherical Marker is used for testing,the experimental results verify the effectiveness of the pro⁃posed method and confirm that the application of the method in this paper can extract Marker points in complex scenes more stably,which greatly enhances the adaptability of the motion capture system for application in complex environments.