卫星关键部件智能识别与三维重建方法

打开文本图片集

中图分类号：TP391.41;V448.25 文献标志码：A doi：10.37188/CO.2025-0091 CSTR：32171.14.CO.2025-0091

Abstract： To achieve the recognition and 3D reconstruction of space target components in complex， low-texture environments for space situational awareness tasks，we propose an end-to-end intelligent perception framework for space targets based on deep learning. This framework enables inteligent recognition and highprecision 3D reconstruction of key space target components. First， based on the lightweight YOLOvl1s network，an attention mechanism is introduced to focus features， achieving precise localization and recognition of space targets and their key components while ensuring real-time performance.This facilitates the extraction of target regions for accurate 3D reconstruction. Subsequently， a novel 3D reconstruction algorithm named Sat-TransMVSNet， specifically designed for low-texture space targets， is proposed. This algorithm employs a multi-scale feature enhancement network for feature extraction and utilizes a novel cost volume regularization method to strengthen geometric constraints at space target edges. It incorporates a backgroundsuppression and foreground-enhancement module，combined with a dynamic depth sampling strategy，to accurately reconstruct space targets. Finally，the overallframework is tested using a self-built multi-angle space target dataset comprising various types.Experimental results indicate that the component recognition algorithm achieves an mAP50 of 0.95， and the comprehensive 3D reconstruction error is 0.2886mm . This demonstrates the framework's capability to meet the requirements for high-precision 3D reconstruction of space targets and intelligent recognition of key components.

Key words： 3D reconstruction; component recognition; deep learning; space targets

1引言

卫星部组件识别和三维重建是空间态势感知[]、碎片清除[2]、在轨服务[]等诸多空间任务的关键技术。（剩余17160字）