基于点云的空间非合作目标结构识别问题研究

Research on Structure Recognition of Spatial Non-cooperative Objects Based on Point Cloud Data

在空间非合作目标捕获任务中,从传感器数据中识别出目标表面的可抓取结构是一个有待解决的问题。以卫星点云数据集作为对象,对4种基于神经网络算法(PointNet、PointNet++、SPLATNet和SO-Net)在卫星结构分割识别任务中的性能进行了比较分析。为了能够更好地测试算法性能,基于NASA在线数据库构建了训练测试数据集,并给出一种点云数据的快速构建方法。使用该方法,可以实现成批量地生成点云数据。仿真测试结果显示:PointNet++在卫星完整点云数据集和非完整点云数据上的分割准确率都是最高,并且分割效果也优于其他算法。

How to recognize the graspable structure of non-cooperative target surface from the sensor data is a problem to be solved in spatial non-cooperative capture tasks. In this paper,the performance of four neural networkbased algorithms,i.e.,PointNet,PointNet++,SPLATNet,and SO-Net,in satellite structure segmentation and recognition tasks is compared and analyzed with a satellite point cloud data set. In order to better test the algorithm performance,a training test data set is built based on the NASA online data set,and a fast point cloud data building method is proposed. With the proposed method,point cloud data can be generated in batches. The simulation results show that among the four neural network-based algorithms,PointNet++ has the highest segmentation accuracy on both satellite complete and incomplete point cloud data sets,and has the best segmentation effect.