Autonomous measurement and semantic segmentation of non-cooperative targets with deep convolutional neural networks
Hang Du,Haidong Hu,Dayi Wang,Xinru Xie,Shuanfeng Xu,Chunling Wei,Yingzi He,Hao Gao
DOI: https://doi.org/10.1007/s12652-021-03553-7
IF: 3.662
2022-01-04
Journal of Ambient Intelligence and Humanized Computing
Abstract:In spatial missions, it is important to estimate kinematic state and to identify the shape of non-cooperative targets. In order to improve the accuracy and generality of the existing algorithm, a real-time recognition and detection method for tumbling non-cooperative target is proposed. First, we design a key point detection network to identify non-cooperative targets and their feature points in the process of long-distance non-cooperative target recognition. At the same time, the detected image is combined with the PNP algorithm to obtain the target’s 6D attitude. Then, a BiSeNet based model is trained for real-time semantic segmentation of satellite components as the base satellite is pursued close to its target. The segmented image is selected with depth information and the relative position of the capturing point is transmitted to the manipulator. Finally, we complete physical experiments under different lighting conditions with a spinning non-cooperative target on a 6-DOF air-bearing table. The experimental results show that the satellite recognition accuracy and object segmentation accuracy are 99.48% and 98.11%, respectively. The position measurement error is less than 1 mm, which achieves more than 50% improvement over the conventional methods.
computer science, information systems,telecommunications, artificial intelligence