航天器舱内结构语义三维重建

3D Semantic Reconstruction of Spacecraft Cabin Structures

对航天器舱内结构进行三维重建是辅助航天员混合现实训练的有效途径,为进一步提高重建结果的适应性,需要对目标结构附加语义信息。首次将语义三维重建技术引入航天员混合现实训练领域,针对航天器舱内结构复杂、排列错落等特点,采用卷积神经网络和即时定位与建图相结合的方式做出针对性设计。在现有算法的基础上,对二维语义分割进行简化并增强,采用加权迭代最近点算法进行目标重建,采用加权平均的方式进行语义更新,充分考虑图像间的联系,利用条件随机场进行后端优化。试验结果证明,算法无论在二维语义分割、SLAM还是最终的语义地图方面都取得了较高的精度;同时运算速度更快,计算机资源占用更小。

3D reconstruction of spacecraft cabin structures is an effective way to assist the mixed reality training for astronauts.To further improve the robustness of reconstruction,it is necessary to add semantic information to target structures.The semantic 3D reconstruction was introduced into the mixed reality training for astronauts.Considering the complicated structures and staggered arrangements in the spacecraft cabin,a method combining the convolutional neural networks and Simultaneous Localisation and Mapping(SLAM)was used.Based on the state-of-the-art system SemanticFusion,the semantic segmentation network was redesigned,the weighted iterative closest point for SLAM was used,and the weighted average was applied to the semantic update.For the connection between images,the conditional random field was adopted for map regularisation.The experiments proved that our method achieved high accuracy in 2D semantic segmentation,SLAM and the semantic map,while the running speed was faster and the use of memory was smaller.