Key point detection and matching method for non-cooperative targets based on deep learning
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 detec-tion.Consequently,the binocular matching of keypoints is achieved indirectly through the process of object recogni-tion.Finally,this approach has been implemented within an independently developed system platform and its effi-cacy 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 meas-urement tasks.
non-cooperative targetrelative pose measurementdeep learningkey point detectionbinocular stereo vision
万方数据
维普
