Research On Vision Based Pose Measurement Methods For Space Uncooperative Objects Using Line Features

Relative pose estimation between the servicing spacecraft and the target spacecraft is a key precondition of space rendezvous,space debris removal as well as on-orbit maintenance.Due to the high accuracy and strong autonomy,the vision based navigation system using optical sensor has been a major approach of the relative pose measurement in the close range rendezvous and docking/berthing phase.The target spacecraft can be classified into two types according to whether the cooperative markers are fixed.One is cooperative object and the other is uncooperative objects.According to whether the information for the object structure is known,the uncooperative objects can also be categorized into two classes,uncooperative object with known model and unknown object.The vision based pose measurement methods for the cooperative objects have been deeply researched and successfully utilized.Compared to the cooperative objects,it is full of challenges to measure the pose parameters for uncooperative objects.Since edges are rich on the man-made low-texture targets and provide a good in-variance to illumination changes and some resilience to hash imaging conditions such as noise and blur,there are huge advantages to measure the objects pose using line features.In order to fulfill the requirement of space uncooperative rendezvous,this paper studies the problem of the pose estimation of space uncooperative objects based on line segments.The content of this paper contains four parts including line feature extraction and matching,camera calibration based on line segments,pose estimation for uncooperative objects based line models as well as structure reconstruction and pose estimation for unknown objects based line segment.Several new methods are proposed in this dissertation,which are as follows:1)A line matching method based on both local descriptor and global topological constraint of lines for low texture objects under complex illumilation is proposed.Firstly,the image line support region is constructed based on the grident and local saliency of the pixels,and then the final image lines are obtained by controlling the false detections according to the Helmholtz principle.After that,the MSLD(Mean Standard Deviation Line Descriptor)is redesigned to get an initial set of correspondences.Then the sidedness constraint is used to get rid of candidates.Finally more matches are achieved by iterative topological filter.Moreover,global angle constrains are implemented to remove wrong matches and make the algorithm more efficient.Compared to the traditional methods,the proposed method is much more robust to illumination changes,noise and blur as well as scale changes.2)A method based on contour model to recognize and optimize the homograhpy for the textureless object in the clutter scene is proposed to handle the problem of camera calibration without control features.Firstly,four image line segments conforming to the certain geometry constraint are selected and utilized as the minimal sample set in the RANSAC framework.The parameters with the minimal sum of the distance errors are picked out as the initial transformation.Then the contour model is sampled as a set of model points and the optimized homography is obtained by minimizing the errors between the sample points and their corresponding image points.Based on the proposed homography estimation method,a flexible camera calibration is designed based on the edge model of the calibration object.Compared to the traditional methods,the proposed method does not need to prepare the calibration object and is much more flexible.3)A pose estimation method based on the integral distance between lines is proposed.Traditional line-based pose estimation methods only utilize the information of the endpoints of the image line segment and ignore the correlation between them.To overcome such issue,an average integral distance function is designed.By exploiting the new distance function,several least-squares techniques are proposed to estimate the pose.Compared to the tranditional methods,the robustness to the error of line fragmentation and oriention has been enhanced.Moreover,the noise model describing the probabilistic relationship between the 3D model line and their finite image observations is derived from the process of least square line fitting.Based on the proposed noise model,the maximum-likelihood approach is exploited to estimate the pose parameters.4)The equivalence between the two types of visual tracking methods for rigid objects has been given.Firstly,the proposed Pn L-like methods and the RAPi D-like methods are compared in the probabilistic framework and it is proved that both of these two types of visual tracking methods are fundamentally equivalent and all of them are maximum-likelihood approaches to estimate the pose parameters given the error model for the noisy edge points.Moreover,from the physical approach of the objective functions for these two types of methods,it is proved that the distance functions between the image line segments and the projected model lines have the equivalent forms.Moreover,comparisons under different performance criteria including computational efficiency,accuracy,and robustness are also conducted.Meanwhile,we improve the performance of the RAPi D-like method-another type of visual tracking approach without extracting image lines by fitting the interpolated location of the corresponding edge pixel in the local region.5)A key structure reconstruction and pose estimation simultaneously for unknown objects in multiple views based on line segments is proposed.Firstly,the correspondences between image lines is built based the proposed line matching method.After that the parameters for camera poses and the 3D lines are obtained linearly.Then an objective function based on the new error model between lines is designed.Finally,the Lie group is exploited to express the 3D line in the minimal formulation and then the optimization problem is converted to the problem of Iteratively Reweighted Least Square.Compared to the traditional methods,the proposed methods can calculate the object positions and the line parameters linearly with only the initial values of the camera rotation.By constructing the objective function,the accuracies of the parameters for object poses and lines are enhanced.Our work is a useful attempt for vision based pose estimation for space uncooperative objects.Moreover,our study can be applied in other vision tasks,i.e.vision-assisted landing of aircraft,scene matching,robot manipulation,vision based navigation,object detection and tracking,etc.