物体表面三维虚拟图像点云数据提取仿真

Simulation of Point Cloud Data Extraction from 3d Virtual Image of Object Surface

由于三维虚拟现实物体表面的点云多边形网格是非结构化的,想要将物体表面图像特征直接输入至注意力机制中难度非常大,且提取点云目标数据时普遍存在物体被遮挡问题,难以获取高质量的点云数据。为此提出三维虚拟现实物体表面图像点云数据提取方法。标定摄像机,获取点云数据对应的坐标。通过kd-tree方法获取点云数据的空间拓扑关系,采用双边滤波融合算法消除小尺度的噪声,完成点云数据的去噪处理。采用ResNet神经网络和3D卷积神经网络分别提取点云数据特征和三维虚拟现实物体表面图像特征,将提取的特征输入注意力机制中,完成点云数据的增强处理,通过上述过程提取高质量的点云数据。仿真结果表明,提出方法应用下可提取出低噪声、高完整度的三维虚拟现实物体表面点云数据,且点云属于的漏提取率低于2%,准确率更高。

At present, the point cloud polygon mesh on the surface of 3D virtual reality object is unstructured, so it is difficult to directly input the image features of object surface into the attention mechanism and obtain high-quality point cloud data. Therefore, this article presented a method of extracting point cloud data of 3D virtual reality image on object surface. At first, we calibrated the camera and then obtained the coordinates corresponding to the point cloud data. After that, we used kd-tree method to obtain the spatial topological relationship of point cloud data. In the meanwhile, we used the bilateral filter fusion algorithm to eliminate the small-scale noise and thus to remove noise from the point cloud data. On this basis, we used ResNet neural network and 3D convolution neural network to extract features of point cloud data and 3D virtual reality object surface image respectively. Moreover, we input these features into the attention mechanism, thus completing the enhancement for point cloud data. Finally, high-quality point cloud data was extracted. Simulation results prove that the proposed method can extract the surface point cloud data of 3D virtual reality object with low noise and high integrity. Meanwhile, the leakage extraction rate of point cloud is less than 2%.