摘要
为了更好地传承传统文化皮影戏,提出了一套基于人体姿态估计的多维交互智能皮影系统,包括皮影机器人和虚拟皮影人。该系统应用了人体姿态估计网络,使用运动算法计算骨骼旋转角度,并设计了一款类人体机械结构应用到皮影上,可以与观众进行多维交互。为了优化皮影系统的流畅性,文章提出了针对皮影戏表演特点的基于透视投影约束的三维关键点优化方案和防遮挡策略。在输出上,增加了卡尔曼滤波,在Jetson AGX上将模型轻量化,并实现了TensorRT加速。经过这些模块处理后,皮影系统的识别准确度和动作流畅度都得到了不小于30%的提升,帧速率达到了33帧/秒。
In order to better inherit traditional culture shadow play,a multi-dimensional interactive intelligent shadow play system based on human pose estimation is proposed,including shadow play robot and virtual shadow play person.The system applies a human pose estimation network,uses a motion algorithm to calculate the rotation angle of the bone,and designs a human-like mechanical structure applied to the shadow play,which can interact with the audience in multiple dimensions.In order to optimize the fluency of the shadow play system,this paper proposes a three-dimensional key point optimization scheme and anti-occlusion strategy based on perspective projection constraint for the characteristics of shadow play performances.In the output,the Kalman filter is added,the model is lightweight on Jetson AGX,and TensorRT acceleration is realized.After processing by these modules,the recognition accuracy and action fluency of the shadow play system are improved by no less than 30%,and the frame rate reaches 33 FPS.
作者
赵健文
曾军英
秦传波
张仲衡
吴健涛
吴柳
倪梓隽
ZHAO Jianwen;ZENG Junying;QIN Chuanbo;ZHANG Zhongheng;WU Jiantao;WU Liu;NI Zijun(School of Electronics and Information Engineering,Wuyi University,Jiangmen 529020,China;Guangzhou Preschool Teachers College,Guangzhou 511300,China)
出处
《现代信息科技》
2025年第3期56-60,67,共6页
Modern Information Technology
基金
广东普通高校人工智能重点领域(2019KZDZX1017)
广东省科技创新战略专项资金(攀登计划专项资金)(pdjh2024a375)
江门市基础与理论科学研究类科技计划项目(江科[2021]87号(2021030101900004391))
五邑大学创新创业基金(2023CY07)。
关键词
深度学习
人体姿态估计
多维交互
皮影戏
Deep Learning
human pose estimation
multi-dimensional interaction
shadow play
