Research On Real-time Relative Pose Estimation Of Satellite Formation Based On Monocular Vision

The instrumentation of relative positions and attitude in satellite formation is one of the key technologies of satellite formation.By obtaining the relative poses and positions of each satellite in the formation,the formation can be perceived and controlled.It is an important application to distributed virtual space scientific instruments by using micro-satellite formation at present and in the future.Considering that the space scientific detection device requires high accuracy of instrumentation,high-precision relative position and attitude measurement is an inevitable requirement of scientific satellite formation.Limited by the size and space of micro-satellites,the measuring device should be simple in structure and low in power consumption.With the continuous development of scientific exploration tasks to the field of deep space exploration,it has become a demand that cannot be ignored to have high autonomy of the measuring device.After comparing a variety of relative position and attitude measurement methods between satellites,considering the long distance,long running time and autonomous requirements of deep space exploration,the measurement method based on monocular vision will become the development direction of micro-satellite formation in the future.Compared with other measurement methods,monocular vision has the characteristics of simple structure,low power consumption,high autonomy and long-term continuous work,which is more suitable for application in this context.In the algorithm of single image,there have been many researches on the cooperation form based on target,but the research on feature-based pose calculation method is not rich in this field.Therefore,this paper focuses on the real-time relative pose estimation algorithm based on feature points.The main research work is as follows:(1)In the feature point extraction and description,in order to improve the realtime ability of image processing,this paper uses FAST feature extraction method as the satellite feature extraction method.At the same time,in order to make the proposed features scale invariant and direction invariant,Gaussian pyramid and feature principal direction calculation are added in feature extraction.In order to ensure good real-time performance without losing calculation accuracy,this paper introduces BEBLID descriptor as a feature description method,and realizes a realtime and high-precision feature extraction algorithm.The experimental results show that the interior point rate of the feature extraction and description algorithm in this paper reaches 89% on the self-made data set.Compared with the commonly used algorithms for feature extraction of satellite images at present,the accuracy is not reduced,and the calculation speed is improved,taking into account the calculation performance and calculation speed.(2)Aiming at the problem of instantaneous mutation in attitude calculation results,the stability of calculation results is poor.In this paper,quadtree screening algorithm is analyzed,and a feature screening algorithm based on stable strategy is proposed,which effectively solves the problem of instantaneous mutation of calculation results.(3)In order to solve the problem that the mismatch in feature matching affects the calculation accuracy,this paper compares Lowe’s feature matching filtering algorithm with GMS feature matching filtering algorithm,and puts forward that GMS feature matching filtering is used before global pose calculation,and feature points that do not conform to the motion law are screened out through the principle of motion consistency,which can effectively improve the input data accuracy of pose calculation and improve the pose calculation results.(4)Based on the algorithm of feature extraction,matching and screening proposed above,the real-time pose estimation of self-built data is studied.In this paper,firstly,the features of satellite images are extracted by FAST-BEBLID algorithm,then the features are extracted and screened by improved quadtree algorithm,the screened features and matching results are screened by GMS algorithm,and then the pose is calculated by EPn P algorithm.According to the calculated pose results,BA is optimized to the final pose.Finally,the calculation accuracy of this algorithm is better than 2cm,which is 30% higher than the comparison algorithm,and the attitude accuracy is less than 1.6,which is 0.1 higher than the ORB-SLAM2 algorithm.