基于改进OpenPose的类人机器人关键点快速检测方法  

A rapid key points detection method for humanoid robots based on improved OpenPose

作　　者：张泽争 王君[1] 董明利[1] 王蕾[1] 燕必希[1] ZHANG Zezheng;WANG Jun;DONG Mingli;WANG Lei;YAN Bixi(School of Instrument Science and Opto Electronics Engineering,Beijing Information Science&Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学仪器科学与光电工程学院,北京100192

出　　处：《激光杂志》2023年第10期65-70,共6页Laser Journal

基　　金：教育部“长江学者与创新团队发展计划”(No.IRT_16R07);北京信息科技大学2021年人才培养质量提高项目(No.5102110802)。

摘　　要：针对现有的关键点检测方法并不能完全适用于类人机器人系统,且检测速度慢等问题,提出一种基于改进OpenPose的类人机器人关键点快速检测方法。基于人体骨骼结构,设计适用于类人机器人的简化系统结构,降低模型复杂度;改进卷积神经网络结构,简化卷积核大小和卷积层数,提高检测速度;制作类人机器人数据集,并对数据集进行增强处理,使图片多样化,降低过拟合风险,提升检测精度。实验结果表明,采用改进的类人机器人关键点检测方法,检测速度是原有方法的3.77倍,精度提升了1.8个百分点。Aiming at the problems of the existing key points detection methods’incomplete applicability to humanoid robot systems and low detection speed,a rapid key points detection method for humanoid robots based on improved OpenPose is proposed.Firstly,a system structure applicable to humanoid robot is designed to reduce the model complexity based on the simplified human skeleton structure.Secondly,improve the structure of convolutional neural network to simplify the size of convolutional kernel,reduce the number of convolutional layers and increase the detection speed.Finally,the humanoid robot dataset is produced and enhanced to diversify the images,reduce the risk of overfitting and improve the detection accuracy.Experimental results show that using the improved humanoid robot key points detection method,the detection speed is 3.77 times higher than before,and the accuracy is improved by 1.8%.

关 键 词：类人机器人 关键点检测 卷积神经网络 OpenPose 

分 类 号：TN27[电子电信—物理电子学]