利用点云处理的实时的类别级位姿估计

Real-Time Category-Level Pose Estimation Utili zing Point Cloud Processing

针对传统的视觉算法无法准确地测量物体3D质心和尺寸、无法检测数据集中未出现过的物体等问题,提出一种基于点云的类别级物体质心与位姿估计算法.首先通过2D分割算法对物体待检测区域进行分割,并通过BlendMask提高实时性;然后将分割区域生成点云,并结合先验形状输入到神经网络中重建物体模型;最后输出一个变形场以及对应矩阵,并通过Umeyama算法对物体进行位姿估计.实验结果表明,所提算法达到了较高的精度;在质心估计方面也优于包络盒中心作为质心的算法;实时速率达到了13帧/s.在GPUNVIDIA2060上进行仿真实验,证明了所提算法计算的质心优于对比算法,以及该算法的准确性及可靠性.

Aiming at the problems that traditional vision algorithms cannot accurately measure the 3D center of mass and size of an object,and cannot detect the objects that have not appeared in the dataset,we propose a category-level object center of mass and position estimation method based on point cloud.Firstly,a 2D segmentation algorithm is used to segment the region to be detected,and BlendMask is used to improve the real-time performance;then,a point cloud is generated from the segmented region and combined with the a priori shapes to reconstruct the object model in a neural network;finally,a deformation field and its corre-sponding matrix are outputted,and the position of the object is estimated by the Umeyama algorithm.The experimental results show that the proposed method achieves high accuracy,outperforms the cen-ter-of-envelope-box algorithm in center-of-mass estimation,and achieves a real-time rate of 13 FPS.Simu-lation experiments on a GPU NVIDIA 2060 demonstrate that the center-of-mass of the proposed method outperforms the comparative methods,and that the method is both accurate and reliable.