融合多注意力机制的布料仿真方法  

Cloth simulation with multi-attention mechanism

作　　者：王婷 靳雁霞[1] 南科良 王松松 WANG Ting;JIN Yan-xia;NAN Ke-liang;WANG Song-song(School of Computer Science and Technology,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学计算机科学与技术学院,山西太原030051

出　　处：《计算机工程与设计》2025年第3期886-894,共9页Computer Engineering and Design

基　　金：山西省自然科学基金项目(202103021224218)。

摘　　要：针对布料仿真中模拟精度与计算效率的平衡问题,提出一种融合多注意力机制的布料仿真方法。使用多精度点云采样对原始点云进行下采样,得到能反映布料真实形状的真实点云;使用多源多尺度特征融合全面捕获并整合查询点的特征,结合神经网络精准预测符号距离函数(signed distance function,SDF)值与梯度,指导拉动查询点;使用自控制损失(self-control loss,SCLoss)动态纠正拉动后的查询点云,缩小与真实点云之间的差异,使用Marching Cubes算法获取布料模拟结果。实验结果表明,该方法比其它布料仿真方法具有更高的模拟精度与计算效率,能够保留准确详细的褶皱细节,是一种高效的布料仿真方法。A cloth simulation method integrating a multi-attention mechanism was proposed to address the balance between simulation accuracy and computational efficiency in cloth simulation.The original point cloud was down-sampled using multi-precision point cloud sampling to obtain a real point cloud that reflected the true shape of the cloth.The multi-source and multi-scale feature fusion was employed to comprehensively capture and integrate the features of query points,and a neural network was combined to accurately predict the signed distance function(SDF) values and gradients,guiding the pulling of query points.A self-control loss(SCLoss) was utilized to dynamically correct the pulled query point cloud,reducing the difference from the real point cloud.The Marching Cubes algorithm was applied to obtain the cloth simulation results.Experimental results indicate that this method exhibits higher simulation accuracy and computational efficiency compared to other cloth simulation methods,effectively preserving accurate and detailed wrinkle details,thereby demonstrating itself as an efficient cloth simulation approach.

关 键 词：布料仿真 多注意力机制 点云采样 特征融合 符号距离函数 自控制损失 深度学习 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]