基于深度学习的非合作目标关键点检测及匹配方法

Key point detection and matching method for non-cooperative targets based on deep learning

针对非合作目标相对位姿测量任务中特征点检测及双目匹配环节易受环境干扰、鲁棒性弱的问题,提出一种更具实用价值的方法。首先,将具有代表性的某型号卫星模型视为非合作目标实验对象,并针对其结构特点开发了关键点标注软件,以生成数据集并用于深度卷积神经网络(DCNN)模型的训练;之后使用不同算法对DCNN模型输出的两类信息进行分析,完成关键点检测;最后通过对识别对象进行双目匹配,从而间接完成关键点双目匹配。将该方法应用到自主搭建的系统平台,并与传统算法进行对比,结果表明,该算法可在实际应用环境中完成非合作目标的关键点检测及其双目匹配,并具有较强的鲁棒性,为非合作目标相对位姿测量任务的关键环节提供了一种新思路。

To address the challenges in the non-cooperative target relative pose measurement tasks,where feature point detection and binocular matching are prone to environmental interference and exhibit limited robustness,a more practical approach is proposed.Firstly,the satellite model is selected as a non-cooperative target for experimental evaluations.A keypoint annotation software based on its structural characteristics is developed to generate a dataset for training the deep convolutional neural network(DCNN)model.Subsequently,the analysis of the two types of information produced by the DCNN model is conducted by utilizing various algorithmic methods for keypoint detection.Consequently,the binocular matching of keypoints is achieved indirectly through the process of object recognition.Finally,this approach has been implemented within an independently developed system platform and its efficacy has been compared with that of traditional algorithms.The results indicate that the algorithm can complete keypoint detection and binocular matching of non-cooperative targets in practical application environments,with strong robustness.It provides a new perspective for the critical steps in non-cooperative target relative pose measurement tasks.