Research On Vision Based Close Range Sensing Of Uncooperative Targets

With the rapid development of aerospace technology,all countries have accelerated their march into space and the space environment has become the venue for competition.Under this background,on-orbit servicing technology,which aims at on-orbit assembly and maintenance,replacement of parts,and refueling to our own spacecrafts,reconnaissance,interference,and attack to enemy spacecrafts,and monitoring and active cleaning to space debris,has aroused widespread concern in various countries.This paper performs research on vision based close range sensing of uncooperative targets,focusing on object detection,tracking,and reconstruction.The research results can not only be used in the close range operation scenarios of on-orbit services,but also can be extended to many fields such as small UAVs,ground intelligent vehicles and mobile robots.The main work and innovations are as follows:1.The detection of a moving target from a video or image sequence is a prerequest to the following recognition,tracking,measuring,reconstructing,etc.The uncertainty of the target motion,the time-variance of the appearance,the dynamics and complexity of the surrounding environment make object detection challenging.This paper presents a moving object detection method based on coarse-to-fine image segmentation.The algorithm consists of motion cue generation,superpixel-level coarse segmentation and pixel-level fine segmentation.The motion cues provide guiding information for the segmentation process,the coarse segmentation reduces the complexity of subsequent processing,and the fine segmentation improves the accuracy of segmentation.In the coarse segmentation stage,a series of criteria for superpixel initial labelling are formulated.Based on the color,texture and spatial location information of superpixels,a distance measuring their similarity is designed for clustering.A series of rules for fusing superpixel initial labelling and clustering results of the coarse segmentation are made.In the fine segmentation stage,an automatically generated Quadmap is introduced to replace the manually labeled Trimap to realize unsupervised segmentation,as well as to provide a better initialization for GrabCut.Experimental results show that the proposed method can effectively detect moving targets and greatly improve the detection efficiency.2.The tracking of the target can obtain the changing pose,shape,etc.of the target,so as to provide navigation information for further approaching the target.This paper presents a size-aware correlation filter based tracker to deal with the affects caused by the varing size of the target.Inheriting the existing methods to estimate the target location and scale,the algorithm absorbs the advantages of various methods in feature selection and fast calculation,and also takes the estimation of aspect ratio into consideration,so it can better handle the impact of the varing size.The experimental results on large datasets validate the effectiveness of the algorithm.3.Estimating the relative pose of the sensor is the basis of object reconstruction.To improve the accuracy and robustness,this paper presents a relative pose estimation method fusing 2D-3D information.The algorithm consists of three parts: preprocessing,point cloud correspondence construction,and error function building and solving.In the preprocessing stage,the SIFT is selected for feature extraction and description.A series of feature matching and selecting criteria are used to improve the matching accuracy,including bidirectional nearest neighbor matching,symmetric selecting,pixel distance selecting and feature distance selecting.2D-3D information correlation,point cloud generation and downsampling provide input for subsequent processing.In the phase of point cloud correspondence,the random k-d tree is used in searching the nearest neighbors to improve efficiency.In the error function construction and solution stage,2D term is introduced into the error function.And by parametering the variables to be solved using Lie algebra,the constrained least squares problem is transformed into an unconstrained least squares problem.The experimental results show that the proposed method can effectively improve the accuracy and robustness of the original ICP method.4.The use of point cloud to represent the target is far from meeting the needs of tasks such as further approaching or capturing,thus further reconstructing is needed to obtain an analytical description of the target shape.Constructing a point cloud using a Gaussian Mixture Model(GMM)is effective,but the large numbers of point contained may cause assumption of large amount of time.To solve this problem,this paper presents a method to accelerate GMM and uses it for reconstruction,consisting of the algorithm flow design and CPU-GPU heterogeneous implementation.In the algorithm flow design,a series of preprocessing including K-means ++,K-mean,parameter initialization provide a good initialization for the parameter estimation of GMM.In the algorithm implementation,CUDA is used to facilitate the respective advantages of CPU and GPU,where CPU is responsible for process control,and GPU is responsible for computing intensive parallelization tasks.The experimental results show that the improvement of algorithm flow can improve the robustness of GMM and reduce the overall time consumption to a certain extent,and the CPU-GPU heterogeneous algorithm further improves the computational speed.