基于扩散概率模型的去噪预测网络预测人体运动  

Denoising Prediction Network Based on Diffusion Probability Model for Predicting Human Motion

作　　者：王婷玉 谢文军 王冬[1] 李琳 刘晓平[1,3] WANG Tingyu;XIE Wenjun;WANG Dong;LI Lin;LIU Xiaopin(School of Computer Science and Information Technology,Hefei University of Technology,Hefei 230601,China;School of Software,Hefei University of Technology,Hefei 230601,China;Anhui Province Key Laboratory of Industry Safety and Emergency Technology,Hefei University of Technology,Hefei 230601,China)

机构地区：[1]合肥工业大学计算机与信息学院,合肥230601 [2]合肥工业大学软件学院,合肥230601 [3]合肥工业大学工业安全与应急技术安徽省重点实验室,合肥230601

出　　处：《小型微型计算机系统》2025年第4期883-891,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金面上项目(62277014)资助;安徽省重点研究与开发计划项目(2022f04020006)资助;中央高校基本科研业务费专项资金项目(PA2023GDSK0047)资助.

摘　　要：近年来,深度学习方法在人体运动预测方面取得了良好的进展,目前单一阶段方法在预测的准确性和多样性上仍存在改进空间,而采用多阶段方式则导致难以端到端预测.为此,本文提出了一种基于扩散概率模型的去噪预测网络,旨在以端到端的方式去预测出准确多样的人体运动.其中提出了一个基于Transformer的模块,包含一个无softmax的修正线性自注意力,提升了人体运动预测的准确性.此外,本文还提出了一种基于DPM-Solver++的扩散模型采样策略,预测出更加连续和多样性的人体运动姿态序列,并将预测出同一质量人体运动姿态序列的采样时间缩减至一半以下.最后,将本文方法在两个标准数据集Human3.6M和HumanEva-I上进行充分评估.实验结果表明,本文方法优于最先进的单一阶段的方法.In recent years,deep learning methods have made good progress in human motion prediction.Currently,there is still room for improvement in the accuracy and diversity of prediction using single stage methods,while using multi-stage methods makes it difficult to predict end-to-end.Therefore,we propose a denoising prediction network based on diffusion probability model,aiming to predict accurate and diverse human movements in an end-to-end manner.A Transformer based module is proposed,which includes a rectified linear self-attention without softmax,improving the accuracy of human motion prediction.In addition,we also propose a diffusion model sampling strategy based on DPM-Solver++to predict more continuous and diverse human motion posture sequences,and reduce the sampling time for predicting human motion posture sequences of the same mass to less than half.Finally,the method proposed in this paper was fully evaluated on two standard datasets,Human3.6M and HumanEva-I.The experimental results show that the proposed method outperforms the most advanced single stage method.

关 键 词：人体运动预测 扩散概率模型 去噪预测网络 修正线性自注意力 DPM-Solver++ 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]