点集配准的目的是获取对应关系和估计模型点集到目标点集的变换。非刚性点集配准的求解难度大,且点集可能含有噪声、遮挡等失真使其求解更加复杂。概率点集配准方法因其对变形、噪声和遮挡具有鲁棒性,本文将点集配准视为概率密度估计问...点集配准的目的是获取对应关系和估计模型点集到目标点集的变换。非刚性点集配准的求解难度大,且点集可能含有噪声、遮挡等失真使其求解更加复杂。概率点集配准方法因其对变形、噪声和遮挡具有鲁棒性,本文将点集配准视为概率密度估计问题,通过极大似然估计,并用EM算法求解对应关系及变换。在再生核希尔伯特空间中指定了两点集之间的变换,并对核函数(即高斯分布)中的高斯滤波器的宽度在迭代过程中逐渐缩小。在合成数据的实验表明,本文方法在变形、噪声等各种类型的畸变下具有鲁棒性,与CPD算法比较,本文方法比它的配准误差更小。The purpose of point set alignment is to obtain correspondences and estimate the transformation from the model point set to the target point set. Non-rigid point set alignment is difficult to solve, and the point set may contain distortions such as noise and occlusion to complicate its solution. Probabilistic point set alignment methods are robust to distortions, noise and occlusion, and in this paper, point set alignment is considered as a probability density estimation problem, which is estimated by great likelihood and solved by EM algorithms for the correspondences and transformations. The transformation between the two point sets is specified in the regenerated kernel Hilbert space, and the width of the Gaussian filter in the kernel function (i.e., the Gaussian distribution) is gradually narrowed down during the iteration process. Experiments on synthesized data show that the method of this paper is robust under various types of distortions such as deformation and noise, and compared with the CPD algorithm, the method of this paper has less alignment error than it.展开更多
针对多目标识别过程中点云分类和分割精度不高的问题,提出了一种基于改进Transformer模型的点云分类与分割方法DRPT(Double randomness Point Transformer),该方法在Transformer模型卷积投影层创建新的点嵌入,利用局部邻域的动态处理在...针对多目标识别过程中点云分类和分割精度不高的问题,提出了一种基于改进Transformer模型的点云分类与分割方法DRPT(Double randomness Point Transformer),该方法在Transformer模型卷积投影层创建新的点嵌入,利用局部邻域的动态处理在数据特征向量中持续增加全局特征属性,从而提高多目标识别中点云分类和分割的精度。实验中采用了标准基准数据集(ModelNet40、ShapeNet部分分割和SemanticKITTI场景语义分割数据集)以验证模型的性能,实验结果表明:DRPT模型的pIoU值为85.9%,比其他模型平均高出3.5%,有效提高了多目标识别检测时点云分类与分割精度,是对智能网联技术发展的有效支撑。展开更多
文摘点集配准的目的是获取对应关系和估计模型点集到目标点集的变换。非刚性点集配准的求解难度大,且点集可能含有噪声、遮挡等失真使其求解更加复杂。概率点集配准方法因其对变形、噪声和遮挡具有鲁棒性,本文将点集配准视为概率密度估计问题,通过极大似然估计,并用EM算法求解对应关系及变换。在再生核希尔伯特空间中指定了两点集之间的变换,并对核函数(即高斯分布)中的高斯滤波器的宽度在迭代过程中逐渐缩小。在合成数据的实验表明,本文方法在变形、噪声等各种类型的畸变下具有鲁棒性,与CPD算法比较,本文方法比它的配准误差更小。The purpose of point set alignment is to obtain correspondences and estimate the transformation from the model point set to the target point set. Non-rigid point set alignment is difficult to solve, and the point set may contain distortions such as noise and occlusion to complicate its solution. Probabilistic point set alignment methods are robust to distortions, noise and occlusion, and in this paper, point set alignment is considered as a probability density estimation problem, which is estimated by great likelihood and solved by EM algorithms for the correspondences and transformations. The transformation between the two point sets is specified in the regenerated kernel Hilbert space, and the width of the Gaussian filter in the kernel function (i.e., the Gaussian distribution) is gradually narrowed down during the iteration process. Experiments on synthesized data show that the method of this paper is robust under various types of distortions such as deformation and noise, and compared with the CPD algorithm, the method of this paper has less alignment error than it.
文摘针对多目标识别过程中点云分类和分割精度不高的问题,提出了一种基于改进Transformer模型的点云分类与分割方法DRPT(Double randomness Point Transformer),该方法在Transformer模型卷积投影层创建新的点嵌入,利用局部邻域的动态处理在数据特征向量中持续增加全局特征属性,从而提高多目标识别中点云分类和分割的精度。实验中采用了标准基准数据集(ModelNet40、ShapeNet部分分割和SemanticKITTI场景语义分割数据集)以验证模型的性能,实验结果表明:DRPT模型的pIoU值为85.9%,比其他模型平均高出3.5%,有效提高了多目标识别检测时点云分类与分割精度,是对智能网联技术发展的有效支撑。
