航天器部件的精细分割和稳定跟踪

Fine segmentation and stable tracking of spacecraft components

为了实现在轨服务过程中对于没有靶标的部件进行操作,需要精细地分割出相关部件,并对其在时序上进行稳定地跟踪。对于部件的精细分割,本文首先基于航天器部件实例分割数据集对实例分割网络Mask RCNN进行了训练,然后在其掩膜分割分支上添加一个优化模块对部件分割结果进行优化。对于部件跟踪,本文首先在Quit_trihard损失的基础上提出分层加权五元组损失,然后利用该损失在航天器部件重识别数据集上对有关重识别网络进行训练,最后将得到的重识别网路嵌入Deep OC SORT跟踪算法以实现对航天器部件的稳定跟踪。实验结果表明:经过掩膜优化后,在部件实例分割测试集上相关实例分割算法的分割精度可提升至84.90 mAP;使用改进后的损失进行部件重识别,在部件重识别测试集上的识别成功率提高至76.86%,同时相关跟踪算法在部件跟踪测试集上的跟踪成功率升至89.38%。因此,本文提出的方法基本可以满足航天器部件的精细分割和稳定跟踪。

In order to realize the operation of components without cooperation markers in on-orbit services,it is necessary to segment the area of the relevant components finely and then track them stably.For the refinement segmentation of components,firstly,the instance segmentation network,Mask RCNN,is trained on the spacecraft component instance segmentation dataset,and secondly a mask refinement module is added to its mask segmentation branch to optimize the component segmentation results.As to component tracking,a hierarchical weighted quintuple loss based on the Quit_trihard loss is proposed to train a re-identification network on the component re-identification dataset,and then the re-identification network trained before is embedded into the Deep OC SORT tracking algorithm for stable component tracking.The experimental results show that after mask optimization,the component segmentation accuracy of the relevant instance segmentation algorithm on the component segmentation test set can be improved to 84.90 mAP;by using the improved loss,the identification success rate on the component re-identification test set is improved to 76.86%,and the tracking success rate of the correlation tracking algorithm on the component tracking test set is improved to 89.38%.Therefore,the method proposed in this paper can basically satisfy the fine segmentation and stable tracking of spacecraft components.