| Accurate attitude measurement is the basis of space mission,the star sensor utilizes stars as reference to estimate three-axis attitude information of the spacecraft with high accuracy.Compared with other attitude measurement equipment,it has the characteristics of strong autonomy,high accuracy,and high reliability,so it has been widely used in various high-accuracy autonomous navigation systems.In recent years,with the rise of deep-space exploration and small satellite booms,more stringent requirements have been put forward on the size,weight,power,and accuracy of star sensors.Therefore,more research work on the related technology of star sensor is of great significance.The development of hardware technology can effectively achieve the miniaturization and low power consumption of star sensors,but the improvement of other key performances mainly depends on star extraction algorithms and star identification algorithms.At present,researchers have done much work on these algorithms,and some of algorithms have been applied in actual star sensor engineering for space missions.However,the following problems still exist:（1）The accuracy of the subpixel center of mass（Co M）centroid estimation algorithm widely used in star sensors is limited by its systematic error,and the centroid accuracy of the Co M algorithm needs to be further improved.（2）The feature patterns in the star identification algorithms based on pattern recognition theory are not robust,their patterns construction and match process are easily influenced by the noises.Moreover,the identification rate of this type algorithm becomes lower when there are relatively few stars in the FOV.（3）The match features in the star identification algorithms based on subgraph isomorphism theory are low-dimensional,and the identification rate and run-time of the algorithms are unsatisfactory when there are large noises in the star images.Based on the existing research work,this dissertation mainly focuses on the subpixel centroid estimation algorithm and star identification algorithm to find corresponding solutions to improve performance of star sensors.In addition,these methods can also be applied to space debris detection,space dim and small target recognition and tracking,space telescope pointing correction,and geodetic astronomical measurement.The main content and innovative points are summarized as follows:1.Aimed at eliminating the systematic error in the subpixel centroid estimation algorithm,an analytical study about the behavior of the systematic error in the Co M algorithm is presented by means of frequency field analysis and numerical simulations,and a compensation method based on the extreme learning machine optimized by the bat algorithm model（BA-ELM）is proposed,in which the BA-ELM model is adopted to predict the systematic error of the actual star centroid position obtained by Co M algorithm,and then compensate the systematic error under the different Gaussian widths for the Co M method to improve the centroid accuracy.Experiment results indicate that our compensation method based on BA-ELM model can effectively eliminate the systematic error,and the 3-pixel compensated systematic errors of the BA-ELM are both in 10-7 pixel level when Gaussian width is 0.3 pixel and 0.671 pixel,whose compensation accuracy is much higher than that of the other compensation methods.Moreover,the centroid accuracy of the BA-ELM systematic error compensation algorithm is higher when compared with reference algorithms under a variety of noise level and different star positions.When the Signal Noise Ratio（SNR）is 3.05,the positioning accuracy of the algorithm is 0.0641 pixel,which is 51.64%and 94.07%higher than the 3-pixel and 5-pixel Co M method,8.27%and10.14%higher than the Co M with threshold method and the Gaussian fitting method,respectively.2.Aimed at solving the problem that the feature patterns in the star identification algorithms based on pattern recognition theory are easily influenced by noises,a star algorithm based on radial and dynamic cyclic patterns is presented in this dissertation.The proposed star identification algorithm adopts a multi-step strategy involving an initial matching step and a chain part extension step to recognize stars,in which the radial pattern and dynamic cyclic pattern are selected as feature patterns to obtain initial matching results of several image stars and then a chain part extension technique is employed to quickly search for the longest match chain as the final output identification result.In the algorithm,the distribution information of neighboring stars is comprehensively utilized to some extent by adding dynamic cyclic pattern to the feature patterns,and a maximum cumulate comparison method is applied in the dynamic cyclic pattern match process to increase the reliability of matching results.Experiment results indicate that the proposed algorithm is more robust than the compared algorithms under a variety of noise,including star position noise,star magnitude noise and false stars,the details are presented as follows:when the star position noise is 2.0 pixels,the identification rate of the algorithm is 97.45%,which is 6.07%higher than that of the optimized grid algorithm;when the star magnitude noise is 0.4 Mv,the identification rate of the algorithm is 96.90%,which is2.75%higher than that of the optimized grid algorithm;when the false star number is4,the identification rate of the algorithm is 95.30%,which is 6.70%higher than that of the optimized grid algorithm.In addition,when the FOV is 12°×12°,the sensor minimum sensitivity is 6.0 Mv,and the number of observation stars to be identified in FOV is set to be 15,the average number of stars successfully identified by our algorithm is 9.0,which is higher than the optimized grid algorithm（4.98）.3.Aimed at improving the performance of the subgraph-isomorphism-based star algorithms under the noise condition,a multi-layers voting star identification algorithm is proposed in this dissertation.The algorithm utilizes triangle unit as the basic element in the first voting process to get candidates of some detected brighter stars in the initial step,and then utilize angular distance feature between brighter stars in the second voting process to verify the obtained candidates and then determine the final output identification results of the algorithm.The strategy of combining high-dimensional triangle unit voting scheme with low-dimensional angular distance voting scheme can improve the reliability of the identification results under the noise condition,and ensure the real-time performance as much as possible.Moreover,for the high-dimensional triangle unit searching and matching problem,SVD dimension reduction method and double lookup table strategy are adopted to shorten the required time in this process.Experiment results indicate that the proposed algorithm achieves higher identification rate than the compared algorithms under a variety of noise,including star position noise,star magnitude noise and false stars,the details are presented as follows:when the star position noise is 2.5 pixels,the identification rate of the algorithm is 95.50%,which is 25.74%higher than that of the geometric voting;when the star magnitude noise is 0.7 Mv,the identification rate of the algorithm is93.90%,which is 9.57%higher than that of the geometric voting;when the false star number is 6,the identification rate of the algorithm is 93.25%,which is 26.68%higher than that of the geometric voting.When the FOV is 25°×25°,the sensor’s minimum sensitivity is 5.5 Mv,and the number of observation stars to be identified in FOV is set to be 10,the average number of stars successfully identified by our star identification is 8.64,which is higher than the geometric voting algorithm（6.92）.In addition,the average run time of the algorithm is 35.3 ms,it is largely shorter than the geometric voting algorithm（76.6 ms）.4.Based on the Odroid-xu4 ARM embedded platform,the star extraction,star identification and attitude determination algorithms of the star sensor are transplanted into this embedded system.We design the suitable algorithms for star extraction module,star identification module and attitude determination module in the embedded platform respectively,and write the corresponding module program with C++,and then run the compiled program in the embedded system.The experiment results show that the average running time between star extraction and attitude determination of the star sensor based on the Odroid-xu4 platform in the"lost in space"condition is 18.62 ms. |
PDF Full Text Request