基于动态图注意力机制的秦俑点云鲁棒配准

Robust point cloud registration of terra-cotta warriors based on dynamic graph attention mechanism

针对目前的点云配准方法在处理秦俑等文物模型时不能很好地解决分辨率不匹配、点云部分重叠、噪声点较多等问题,提出一种基于动态图注意力机制的ResUNet配准模型。该模型将残差模块融入U-Net网络中,使用三维稀疏体素卷积计算点云特征,并引入一种新的归一化技术:批邻域归一化(Batch-Neighborhood Normalization,BNHN),来提高特征对于点密度变化的鲁棒性;为了进一步提高配准性能,该模型通过自注意力机制和交叉注意力机制聚合局部特征和上下文特征,最后结合随机抽样一致性算法来估计源点云与目标点云之间的变化矩阵,完成秦俑文物模型的鲁棒配准。为了验证本文方法的有效与鲁棒,使用四组数据集(3DMatch、3DLoMatch、分辨率不匹配的3DMatch数据集以及两组秦俑数据)对配准模型进行测试,实验结果表明,该算法在3DMatch数据集和3DLoMatch数据集上的配准召回率分别达到90.1%和61.0%;在分辨率不匹配的3DMatch数据集,相比与基于特征学习的配准算法,该算法在配准召回率上提升了5%~20%;在秦俑数据集上,相对旋转误差均小于0.071,相对平移误差均小于0.016,相较于同类算法减少了一个量级或几倍。因此,本文的模型能够提取三维点云的关键特征信息,并且对点密度和重叠度变化具有更高的鲁棒性。

The current point cloud registration methods cannot effectively address resolution mismatches,partial overlaps of point clouds,and numerous noise points when used for cultural relic models such as Terra-cotta Warriors. Hence,a ResUNet registration model based on the dynamic graph attention mechanism is proposed. The model integrates the residual module into the U-Net,performs three-dimensional(3D)sparse voxel convolution to calculate the features of point clouds,and applies a new normalization technology known as batch-neighborhood normalization to improve the robustness of features against point density changes. To improve the registration performance,the model aggregates local and context features via self-and cross-attention mechanisms. Finally,a random sampling consensus algorithm is used to estimate the change matrix between the source and target point clouds to complete the robust registration of the Terra-cotta Warriors model. To verify the effectiveness and robustness of the proposed method,four datasets(3DMatch,3DLoMatch,3DMatch with resolution mismatches,and two sets of terra-cotta warrior data)were used to test the registration model. Experimental results show that the registration recall was 90. 1%and 61. 0% in the 3DMatch and 3DLoMatch datasets, respectively. In the mismatched-resolution3DMatch dataset,compared with feature learning-based registration algorithms,our algorithm improved the registration recall by 5%–20%. In the terra-cotta warrior dataset,the relative rotation and translation errors were less than 0. 071 and 0. 016,respectively,which are several times to one order of magnitude lower than those of other algorithms. The model proposed herein can extract key feature information from a 3D point cloud and is more robust to variations in point density and overlapping compared with other models.