基于深度学习的空间多目标识别方法研究

Research on Spatial Multi-objective Recognition Based on Deep Learning

随着智能感知技术、目标识别技术的快速发展,以卫星为主要代表的太空飞行器已成为各国太空攻防出奇制胜的重要军事资源。精确识别卫星的类型,并精确定位卫星的帆板、喷管、星敏感器等部件是实施太空攻防和在轨维护的重要前提及保障技术。利用基于深度学习的卷积神经网络YOLO模型对空间卫星及其部件进行识别,对两种卫星模型的三维模型及实物模型图片集进行训练,对近距离正视、远距离、遮挡、运动模糊等不同条件下的卫星及卫星部件进行识别,几种情况下卫星及卫星部件的识别准确率均达到了90%以上,对在轨服务、太空攻防对抗等领域有重要意义。

With the rapid development of intelligent perception technology and target recognition technology,spacecraft which is mainly represented by satellites,has become an important military resource for the extraordinary success of space attack and defense in various countries.Accurately identifying the type of satellite and pinpointing the components of the satellite’s windsurfing,nozzles,and star sensors are important prerequisites and safeguards for space attack and on-orbit maintenance.In this paper,the deep learning-based convolutional neural network YOLO model is used to identify the space satellite and its components,and three-dimensional model and physical model image set of the two satellite models are trained for close-range front view,long distance,occlusion,and motion blur.Satellites and satellite components are identified under different conditions.In some cases,the accuracy of satellite and satellite components is more than 90%,it is of great significance in the field of on-orbit services,space attack and defense confrontation.