摘要
场景信息驱动人类调整动作轨迹,对人体动作预测影响较大。当前研究仅捕获场景信息更新动作特征,忽略了场景与动作的互影响关系。为此,提出一种视觉感知人景互影响的人体动作预测方法。提取动作特征和场景特征,然后循环执行场景信息捕获单元和场景适应度增强单元。前者捕获影响动作的场景信息,后者利用该信息更新动作特征以增强场景适应性。完成循环后,得到场景适应型动作特征。基于该特征执行噪声逆扩散完成动作预测。在3个数据集上进行实验,结果表明本文方法的预测误差低于当前主流方法,验证了其有效性。本文方法将为真实场景中的人体动作预测提供更加可靠的解决方案。
Scene information drives humans to adjust motion trajectories and greatly influences human motion predic-tion.Current research only updates motion features with scene information and ignores their mutual influences.Hence,a human motion prediction method with visual perception of human-scene mutual influence is proposed in this paper.Mo-tion and scene features are extracted,and scene information capture and adaptability enhancement are iteratively ex-ecuted.The former captures scene information affecting human motions,whereas the latter updates motion features with the information to enhance their scene adaptability.After the iteration,the scene-adaptive action features are obtained.Noise inverse diffusion is performed based on the features to complete motion prediction.Experiments conducted on three datasets demonstrate that the proposed method has lower prediction error than the current methods,which verifies its effectiveness.The proposed method provides a more reliable solution for human motion prediction in real scenes.
作者
李沁
陈飞扬
彭晗
王勇
刘利枚
张伟
LI Qin;CHEN Feiyang;PENG Han;WANG Yong;LIU Limei;ZHANG Wei(School of Artificial Intelligence and Advanced Computing,Hunan Technology and Business University,Changsha 410205,China;Xiangjiang Laboratory,Changsha 410205,China;School of Automation,Central South University,Changsha 410083,China;ByteDance,Shenzhen 518063,China)
出处
《智能系统学报》
北大核心
2025年第4期1010-1023,共14页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(62202161)
湖南省教育厅科学研究项目(20A125,22A0460,23B0597,24B0584)
湘江实验室重大项目(23XJ01007,23XJ01009)
湖南省自然科学基金项目(2025-JJ60384)。
关键词
人体动作预测
场景信息
视觉感知
动作特征
场景特征
人景互影响
场景适应性
噪声逆扩散
human motion prediction
scene information
visual perception
motion features
scene features
human-scene mutual influence
scene adaptability
noise inverse diffusion
