Research On Visual Navigation Algorithms Of Lander Based On 3D Model Of Celestial Body

As one of the key technologies of the lander,the performance of navigation is related to the process of subsequent missions.The development of advanced navigation approaches is the focus of research in small body landing probes.At the landing stage of the small body exploration mission,this dissertation studies the visual navigation methods of the lander in both visual odometer and computer vision fields on the condition of the known 3D model of small body.The main contents of this dissertation are as follows:Firstly,at the perspective of traditional visual odometer research,a visual navigation approach based on image itensity alignment is proposed with the reference of current main visual navigation methods.In the descent process of the lander,the navigation approach estimates position and attitude of the lander only at some states via extracting and matching features of the target image taken by the navigation camera and the reference image generated by 3D model rendering,and obtaining the depth information of features.Here,these states are difined as the key frames.Between the adjacent key frames,the features of the presequence key frames are reprojected onto the target image,and the local motion trajectory of the lander is estimated via the itensity alignment of these reprojections with the reference image.It provides a higher navigation performance.This navigation approach uses the 3D model to render the reference image in real-time,which solves the problem that the images of small body change greatly in the mesoscale of navigation camera due to the landing vehicle descent,and it is difficult to match the features.The intensity alignment,from the global point of view,considers the internal constraint of the gray image,overcoming the problem that traditional feature matching is easy to fall into local minimum.It improves the performance of visual navigation.Secondly,at the perspective of computer vision research,a visual navigation approach based on control points recognition is proposed by referring the current deep learning technology.The approach detects the known control points and establishes correspondence of the control points and their 2D image proections via deep network and then uses a RANSAC-based Perspective-n-Point(Pn P)algorithm to compute position and attitude of the lander under the coordinate of the small body.This navigation method,from the global point of view,uses the vectors that the inlier pixel pointing to the control points to indirectly represent the position of control points in2 D image,which reduces the dependence of features in the general navigation methods,and effectively solves the problems that the features are not significant or the known landmarks fall outside the camera’s view field.And,when the initial position and attitude of lander are not precious enough,which traditional methods based on features matched have a difficult to correct,this approach could also estimate a more accurate pose of lander at first,providing a reference basis for the subsequent manipulation.It is remarkable.Finally,on the base of research that the visual navigation based on control points recognition,a kind of 6 dof vision navigation approach based on deep convolution network is proposed,which combines the two-stage paradigm in control points recognition visual navigation method into one stage,directly estimating postion and attitude of lander via deep convolution network.Compared the two-stage visual navigation algorithm,this approach avoids the cumulative error in two-stage paradigm and improves the performance of the navigation system.In addition,this dissertation also uses the idea of itensity alignment to further refine the pose of lander estimated by this method and two-stage visual navigation method,getting a higher accuracy and precision and providing better navigation performace.