基于PointNet改进的无序工件点云配准算法

Improved Unordered Workpieces Point Cloud Alignment Algorithm Based on PointNet

在无序工件抓取场景中,待抓取的工件处于散乱、堆叠的状态,抓取难度较大,传统配准算法精度不高。针对工件存在堆叠和点云数据含有噪声的场景下,无序工件点云配准的准确性不高的问题,研究提出基于PointNet改进的三维点云配准算法对无序工件进行位姿估计。算法用于模板点云和目标点云的特征提取和匹配,并结合ICP算法求解无序工件的位姿参数,最后通过迭代方式提高点云配准精度。采用结构光相机作为点云数据采集设备,以多种不同形状的工业零件作为配准对象,将采集得到的场景点云数据进行点云配准。实验制作并使用无序工件数据集(WorkpiecesDataSet)训练并测试点云配准网络的性能。实验表明,提出的网络模型(i-SAM)在点云配准任务中具有较小的旋转和平移误差,点云模型数据的误差为(0.783,0.011),场景点云数据的配准误差为(1.269,0.016)。与主流算法相比,对含有噪声和不完整的点云具有较强的鲁棒性。

In the unordered workpiece grasping scenario,the workpieces to be grasped are in a scattered and stacked state,which is difficult to grasp,and the accuracy of the traditional alignment algorithm is not high.For scenarios where there is stacking of artifacts and point cloud data containing noise,the accuracy of point cloud alignment of disordered artifacts is not well,and the research proposes a PointNet-based improved 3D point cloud alignment algorithm to estimate the bit pose of unordered workpieces.The algorithm is used for feature extraction and matching of the source and target point clouds,and combined with the ICP algorithm to solve the unordered workpiece's positional parameters,and finally the point cloud alignment accuracy is improved by iterative methods.Using a structured light camera as a point cloud data acquisition device,we use a variety of industrial parts of different shapes as alignment objects and align the acquired field point cloud data with the point cloud.Experiments were conducted to produce and train and test the performance of the point cloud alignment network using an unordered WorkpiecesDataSet.The experiments show for the proposed network structure(i-SAM)that it has small rotation and translation errors in the point cloud alignment task with(0.783,0.011)for point cloud model data and(1.269,0.016)for field point cloud data.Compared with the mainstream algorithm,it is more robust to point clouds containing noise and incompleteness.