Research On Key Technologies Of Spatial Target Feature Extraction And Identification

Deep space exploration is the forefront of the world’s scientific and technological development.It plays a prominent role in revealing the origin of the earth,the origin of mankind and leading the development of science and technology.In 2007,the Chang’e I around the moon became the starting point for China’s march into deep space exploration.The successful return of Chang’e V from the moon in 2020 is a milestone of China’s deep space exploration.In 2021,the "Tian Wen-1" Mars landing detector successfully landed on Mars.In the future,it will also carry out deep space exploration missions targeting asteroids and comets,and continue to carry out on orbit service projects such as space approaching,berthing and maintenance.In the past,space flight relied on ground decision-making,measurement and control to complete key or high-risk space operation tasks.Due to the long flight distance,long delay time of signal command transmission and high real-time requirements,ground decision-making and measurement and control methods can not meet the needs of deep space exploration and on orbit service tasks.In recent years,the research results of machine vision,pattern recognition,and artificial intelligence have been successfully applied in some industrial fields and commercial fields.However,due to the uncertainty of space flight environment and characteristics of space missions,there is a considerable distance between the theories of machine vision,pattern recognition and artificial intelligence and the applications in the spacecraft GNC(Guidance Navigation and Control)field.The research focuses on several key issues in the extraction and identification of spatial target features,and carries out experimental verification using the ground tests data and the flight mission data.The main innovative work includes:1.Aiming at the problem of high cost and low efficiency of target sample mobile phone in space environment,a training sample generation method based on spacecraft dynamics simulation is proposed.First,the research establishes the spacecraft dynamics model,and adds the spacecraft attachment disturbances,including the tank liquid sloshing,the sensor measurement error and other random disturbance items,and outputs the relative 6-DOF data between the spacecraft and the target.Second,the simulation image sequences and their corresponding spatial operation deviations were generated in combination with the internal parameters of the imaging device and the target structure model.Then the sequences are treated as the training samples of the classifier.This method improves the efficiency of training sample collection,and can obtain a large number of samples with diversity and representativeness for classifier training.2.Aiming at the problem of large image feature matching error caused by complex space lighting conditions and various reflection characteristics of spacecraft convenience materials,a space target feature extraction and inter frame matching method based on SIFT clustering is proposed.First,the research proposes a feature clustering method based on SIFT feature distance similarity criteria.It selects features that can match exactly from a large number of SIFT features,and then re-clusters the filtered features.Second,according to the distance between cluster centers,selectively recover some features from the filtered features.This method not only guarantees the number of SIFT features extracted from the image,but also guarantees the accuracy of these features.3.Aiming at the problem of low target recognition rate due to the strong similarity of nose tip structure of aerospace surface,a target recognition method based on knowledge base construction with hierarchical semantic rule reasoning is proposed.First,the research firstly divided the image into several levels.The features are extracted in each level area,and the targets are initially identified,Then a rule-based knowledge base which describes the relationship between the target domain and the image hierarchies is constructed by training decision trees.Then,this method decomposes the target classification problem into two steps.The first is the low-level annotation of the target.The second is the high-level reasoning.The proposed method increases the generalization ability of the classifier.Evaluated on the standard test data set,the experimental results verify the good performance of the method.The results of ground-based semi-physical simulation test also verify the effectiveness of this method for space operation safety judgment and space target recognition.4.In order to verify the effectiveness of the methods proposed in the thesis,the safety judgment test verification process of lunar orbit autonomous rendezvous and docking is designed.After describing the test environment and is introduced,the composition and structure of the test system,the effectiveness of the proposed method is verified by the rendezvous and docking safety judgment experiment.Finally,the Chang’5 flight data are used evaluate the effectiveness of the proposed methods.The methods proposed and validated in the research are an exploration of machine vision,pattern recognition and artificial intelligence theory and technology in the field of aerospace engineering.The theories and methods,experience and lessons,data resources and experimental results in the research are the technical accumulation of moon exploration,on-orbit service,deep space exploration and other tasks in China.