视差不连续区域和重复纹理区域的误匹配率高一直是影响双目立体匹配测量精度的主要问题,为此,本文提出一种基于多特征融合的立体匹配算法。首先,在代价计算阶段,通过高斯加权法赋予邻域像素点的权值,从而优化绝对差之和(Sum of Absolute...视差不连续区域和重复纹理区域的误匹配率高一直是影响双目立体匹配测量精度的主要问题,为此,本文提出一种基于多特征融合的立体匹配算法。首先,在代价计算阶段,通过高斯加权法赋予邻域像素点的权值,从而优化绝对差之和(Sum of Absolute Differences,SAD)算法的计算精度。接着,基于Census变换改进二进制链码方式,将邻域内像素的平均灰度值与梯度图像的灰度均值相融合,进而建立左右图像对应点的判断依据并优化其编码长度。然后,构建基于十字交叉法与改进的引导滤波器相融合的聚合方法,从而实现视差值再分配,以降低误匹配率。最后,通过赢家通吃(Winner Take All,WTA)算法获取初始视差,并采用左右一致性检测方法及亚像素法提高匹配精度,从而获取最终的视差结果。实验结果表明,在Middlebury数据集的测试中,所提SAD-Census算法的平均非遮挡区域和全部区域的误匹配率为分别为2.67%和5.69%,测量200~900 mm距离的平均误差小于2%;而实际三维测量的最大误差为1.5%。实验结果检验了所提算法的有效性和可靠性。展开更多
在点云处理过程中,法向估计是非常重要的一步。现有的深度霍夫变换的法向估计网络通过对点云进行霍夫变换得到邻域特征,再将其输入至卷积神经网络中学习估计法向。但由于霍夫变换过程中存在一定信息损失导致最后所得法向不准确,效果不...在点云处理过程中,法向估计是非常重要的一步。现有的深度霍夫变换的法向估计网络通过对点云进行霍夫变换得到邻域特征,再将其输入至卷积神经网络中学习估计法向。但由于霍夫变换过程中存在一定信息损失导致最后所得法向不准确,效果不够理想。对此,本文先通过霍夫变换将法向空间与二维平面相对应,并将二维空间离散化获得所有潜在切平面,设计特征聚合将点特征转化为潜在切平面特征作为CNN输入来降低霍夫变换过程中信息的损失,从而提升卷积神经网络的输入,进而提升网络整体的法向估计质量。实验结果表明,由此产生的法向估计网络的整体性能有所提升,对于不同噪声尺度也更具鲁棒性。In the process of point cloud processing, normal estimation is a very important step. The existing normal estimation network of deep Hough transform obtains neighborhood features by performing Hough transform on the point cloud and then inputs them into a convolutional neural network to learn and estimate the normal. However, due to certain information loss in the Hough transform process, the finally obtained normal is inaccurate and the effect is not ideal. In response to this, this paper first corresponds the normal space to a two-dimensional plane through Hough transform, and discretizes the two-dimensional space to obtain all potential tangent planes. Feature aggregation is designed to transform point features into potential tangent plane features as the input of CNN to reduce the information loss in the Hough transform process, thereby enhancing the input of the convolutional neural network and further improving the overall normal estimation quality of the network. Experimental results show that the overall performance of the resulting normal estimation network is improved, and it is also more robust to different noise scales.展开更多
文摘视差不连续区域和重复纹理区域的误匹配率高一直是影响双目立体匹配测量精度的主要问题,为此,本文提出一种基于多特征融合的立体匹配算法。首先,在代价计算阶段,通过高斯加权法赋予邻域像素点的权值,从而优化绝对差之和(Sum of Absolute Differences,SAD)算法的计算精度。接着,基于Census变换改进二进制链码方式,将邻域内像素的平均灰度值与梯度图像的灰度均值相融合,进而建立左右图像对应点的判断依据并优化其编码长度。然后,构建基于十字交叉法与改进的引导滤波器相融合的聚合方法,从而实现视差值再分配,以降低误匹配率。最后,通过赢家通吃(Winner Take All,WTA)算法获取初始视差,并采用左右一致性检测方法及亚像素法提高匹配精度,从而获取最终的视差结果。实验结果表明,在Middlebury数据集的测试中,所提SAD-Census算法的平均非遮挡区域和全部区域的误匹配率为分别为2.67%和5.69%,测量200~900 mm距离的平均误差小于2%;而实际三维测量的最大误差为1.5%。实验结果检验了所提算法的有效性和可靠性。
文摘在点云处理过程中,法向估计是非常重要的一步。现有的深度霍夫变换的法向估计网络通过对点云进行霍夫变换得到邻域特征,再将其输入至卷积神经网络中学习估计法向。但由于霍夫变换过程中存在一定信息损失导致最后所得法向不准确,效果不够理想。对此,本文先通过霍夫变换将法向空间与二维平面相对应,并将二维空间离散化获得所有潜在切平面,设计特征聚合将点特征转化为潜在切平面特征作为CNN输入来降低霍夫变换过程中信息的损失,从而提升卷积神经网络的输入,进而提升网络整体的法向估计质量。实验结果表明,由此产生的法向估计网络的整体性能有所提升,对于不同噪声尺度也更具鲁棒性。In the process of point cloud processing, normal estimation is a very important step. The existing normal estimation network of deep Hough transform obtains neighborhood features by performing Hough transform on the point cloud and then inputs them into a convolutional neural network to learn and estimate the normal. However, due to certain information loss in the Hough transform process, the finally obtained normal is inaccurate and the effect is not ideal. In response to this, this paper first corresponds the normal space to a two-dimensional plane through Hough transform, and discretizes the two-dimensional space to obtain all potential tangent planes. Feature aggregation is designed to transform point features into potential tangent plane features as the input of CNN to reduce the information loss in the Hough transform process, thereby enhancing the input of the convolutional neural network and further improving the overall normal estimation quality of the network. Experimental results show that the overall performance of the resulting normal estimation network is improved, and it is also more robust to different noise scales.
