摘要
自主着舰引导距离跨度大,使得机载单目视觉引导获取的图像序列中舰船目标尺度变化大,已有位姿测量方法难以实现覆盖大距离跨度的高精度单目视觉位姿测量。基于已有基于稀疏关键点集合的单目视觉位姿测量方法,从提升关键点检测精度出发,分析目标尺寸、网络输入尺寸对关键点检测精度的影响规律。在兼顾精度和效率的前提下,提出一种新的基于多部件的单目视觉位姿测量方法,采用稀疏关键点集合对部件进行简化表示,在舰船目标整体部件的粗略位姿估计的基础上,引入路径聚合特征金字塔网络和分层编码模块,实现局部部件关键点高精度检测,进一步综合各部件高精度关键点检测结果,通过求解Perspective-n-Points(PnP)问题,实现覆盖着舰引导大距离跨度范围内的鲁棒、高精度位姿测量。仿真实验及缩比实物实验的结果表明,文中方法实现了着舰引导大距离跨度范围内鲁棒、高精度单目位姿测量,性能优于已有方法,在嵌入式平台上的平均单帧推理时间约为40 ms。
The autonomous landing guidance involves a large distance span,resulting in significant scale variations of the ship target in the image sequences obtained through monocular vision guidance.Existing pose measurement methods struggle to achieve high-precision monocular vision pose measurement across such a wide distance range.For current monocular vision pose measurement methods based on sparse keypoint sets,this paper focuses on improving the accuracy of keypoint detection,and analyzes the impact of target size and network input size on keypoint detection accuracy.Furthermore,this paper proposes a novel monocular vision pose measurement method based on multiple components,balancing both accuracy and efficiency.By using sparse keypoint sets to represent components in a simplified manner,and building on a coarse pose estimation of the overall ship target components,this method introduces a path aggregation feature pyramid network and a hierarchical encoding module to achieve high-precision detection of local component keypoints.Subsequently,by integrating the high-precision keypoint detection results of all components and solving the Perspective-n-Points(PnP)problem,the method achieves robust and high-precision pose measurement across the large distance span required for landing guidance.Simulation experiments and scaled physical experiments demonstrate that the proposed method achieves robust and high-precision monocular pose measurement across the large distance span for landing guidance,outperforming existing methods,with an average single-frame inference time of approximately 40 ms on embedded platforms.
作者
陈霖
顾曦文
陈知颖
张倬
孙晓亮
CHEN Lin;GU Xiwen;CHEN Zhiying;ZHANG Zhuo;SUN Xiaoliang(College of Aerospace Science and Engineering,National University of Defense Technology,Changsha 410073,China;Hunan Province Key Laboratory of Image Measurement and Vision Navigation,National University of Defense Technology,Changsha 410073,China;91351 Troops,Xingcheng 125106,China)
出处
《航空学报》
北大核心
2025年第15期274-288,共15页
Acta Aeronautica et Astronautica Sinica
基金
国家自然科学基金(12272404)。
关键词
单目视觉
着舰引导
位姿测量
深度学习
关键点检测
monocular
landing guidance
pose measurement
deep learning
keypoint detection
