Visual Inspection And Security Prediction For Tracker Space Operation

This project originates from the aerospace project of the center of pattern recognition and intelligent system of Harbin Institute of Technology.It involves the problem of spatial rendezvous and docking pose measurement.The paper mainly studies the generation of simulation samples and methods of monocular vision detection method and safety prediction in the rendezvous proximity phase.The paper also verifies the feasibility and tests the performance of the above method on the simulation samples.The conditions already in place are the target spacecraft 3D model and the data generated by the control system.The conditions are used to generate simulation samples in three ways.Fristly,software simulation samples are generated by coding with a 3D image library.Secondly,“3D printing model” simulation samples are generated by using the 3D printing physical target model and a surveillance camera.Thirdly,semi-physical simulation samples are generated by physical target model and the nine-degree-of-freedom turntable's accurately controlling of surveillance camera.The large scale of simulation samples makes it possible to train the deep neural network.First,the common feature method is used to complete the visual measurement task.The pixel coordinates of n salient features are detected from the target spacecraft image,then the relative pose can be calculated by the solution of the PnP problem.A feature detection process based on template matching is proposed in this paper.By detecting the ORB features of a group of templates and the image to be detected,the method selects the best matching pair of points to obtain a homography matrix,thereby calculating n feature points' coordinates of the image to be detected.A new PnP solution is proposed in the paper to calculate pose,which combining EPnP algorithm and LHM orthogonal iterative process.The paper also uses a faster feature tracking algorithm to improve the speed of the operation and deal with the failure of feature detection.Observing and analyzing the three-dimensional distribution characteristics of the target data,this paper builds a two-dimensional Gaussian model to evaluate safety of rendezvous and docking.Considering the wide application of deep neural networks in the field of computer vision,the paper studies the pose estimation and security prediction methods based on deep neural networks.In this paper,the pose estimation neural network is modified on the basis of PoseNet network structure,adding the spatial pyramid pooling layer and improving the loss function.The LSTM network is also used to predict the current pose according to the position of the first k frame.The paper trains a simple fully connected network and combining it with the pose estimation network to get an end-to-end network structure from image to security for security evaluation.In order to complete the space rendezvous and visual inspection and security forecasting tasks better,a rendezvous and docking process monitoring and inversion system is developed in the paper.The two methods above proposed in this paper are used to complete pose measurement and safety prediction.The OpenGL library is used to reverse the rendezvous and docking process with 3D animation.The manual intervention module is added to improve the reliability of the system.