Star sensor star map recognition method based on K vector lookup method and CNN (Convolutional Neural Network)

Abstract

The invention discloses a star sensor star map recognition method based on a K vector lookup method and a CNN network. The method comprises the following steps: S1, converting a star sensor star map three-dimensional unit vector into a feature with the maximum one-dimensional standard deviation by using the CNN network; s2, creating an index table and a K vector lookup table which are sequenced by features, and performing K vector lookup; and S3, completing star point pixel distance matching based on the K vector lookup table. The star map recognition speed in the all-day recognition mode and the star tracking mode is increased, and the dynamic performance of the star sensor is greatly improved.

Description

Technical Field

[0001] The invention relates to the technical field of star sensors, and in particular to a star sensor star map recognition method based on a K vector search method and a CNN network. Background Art

[0002] Star sensors are optical measurement devices that output the attitude information of spacecraft in an inertial coordinate system based on the position of stars in that inertial coordinate system. They are currently widely used in space missions because of their high precision and high reliability. Star sensors generally work statically, so the dynamic performance of domestically developed star sensors is not high, which limits their application occasions and they are usually suitable for the stage of stable flight of the carrier. With the rapid development of today's aerospace technology, the rotation, movement, and shaking of the carriers on which star sensors are installed are becoming more and more frequent, especially high-dynamic carriers such as agile satellites have put forward higher requirements on the dynamic performance of star sensors. Therefore, it is particularly important to study high dynamic performance as a key indicator of star sensors.

[0003] The dynamic performance of star sensors is mainly affected by hardware systems and software algorithms. When the hardware system has been determined or has large limitations, there is an increasing demand to improve the dynamic performance of star sensors by studying software algorithms. The software algorithm of the star sensor determines that its main working modes are the all-day recognition mode and the star tracking mode. After obtaining the initial attitude through all-day star map recognition, the star sensor enters the star tracking mode, and the star sensor will subsequently remain in the star tracking mode. Therefore, the software algorithms of the two working modes, the all-day recognition mode and the star tracking mode, directly affect the dynamic performance of the star sensor. Summary of the invention

[0004] The invention aims to provide a star sensor star map recognition method based on a K-vector search method and a CNN network, thereby increasing the star map recognition speed in a full-day recognition mode and a star tracking mode and greatly improving the dynamic performance of a star sensor.

[0005] In order to achieve the above objectives, the present invention is implemented by the following technical solutions:

[0006] A star sensor star map recognition method based on K vector search method and CNN network is characterized in that it includes the following steps:

[0007] S1, star sensor star Figure 3 The dimensional unit vector is converted into the one-dimensional feature with the largest standard deviation using the CNN network;

[0008] S2, create an index table and a K-vector lookup table sorted by features, and perform K-vector lookup;

[0009] S3, completing star point pixel distance matching based on the K vector lookup table.

[0010] Furthermore, the step S1 is specifically as follows: obtaining a feature vector according to the CNN network, prioritizing the navigation star points according to the feature vector matching, and sorting the detection stars and the navigation stars. This function is to convert the three-dimensional unit vector into a feature with the largest one-dimensional standard deviation.

[0011] Furthermore, the step S2 is used for fast searching of the navigation star catalog.

[0012] Furthermore, the step S3 is specifically as follows: the matching confidence between the two can be determined according to the number of similar eigenvalues ​​found by the K vector, and the matching pair for star map recognition is selected according to the confidence.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] The star sensor dynamic performance method provided by the present invention can improve the dynamic performance of the star sensor. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The present invention is a flow chart of a star sensor star map recognition method based on a K-vector search method and a CNN network;

[0016] Figure 2 A flow chart of converting a three-dimensional unit vector of a star point in a CNN network star map of the present invention into a feature with a maximum one-dimensional standard deviation;

[0017] Figure 3 A flow chart for realizing star tracking by the star sensor of the present invention;

[0018] Figure 4 It is a flow chart of the star point pixel distance matching method of the star sensor of the present invention;

[0019] Figure 5 The following is a graph showing the actual performance of a star sensor after using the prior art and the method of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] The method for improving the dynamic performance of star sensors of the present invention proposes an algorithm for improving the dynamic performance of star sensors based on K vectors and CNN networks, which increases the speed of star map recognition in all-day recognition mode and star tracking mode, and greatly improves the dynamic performance of star sensors. The method uses CNN networks to convert three-dimensional unit vectors into features with the largest one-dimensional standard deviation, and then creates an index table and a K vector lookup table sorted by features, which speeds up the speed of searching for feature points by traversing indexes, and finally uses a pixel distance matching method to complete star map recognition. For the specific method flow, please refer to Figure 1 , the method of the present invention comprises the following steps:

[0022] S101, using CNN network to convert three-dimensional unit vector into one-dimensional feature with maximum standard deviation;

[0023] The specific principles are as follows:

[0024] The detector uses the Faster R-CNN algorithm. Faster R-CNN uses a shared convolutional network to form a regional proposal network RPN (regional proposal network), and uses RPN to directly predict the proposal box. Compared with selective search SS (selective search), RPN predicts fewer proposal boxes with higher quality. In addition, most of the RPN predictions are completed in the GPU, and the convolutional network and Fast R-CNN are partially shared, so the target detection speed is greatly increased.

[0025] 1) Region Proposal Network

[0026] The RPN structure takes an image (of any size) as input and outputs a set of rectangular target proposal boxes, each of which corresponds to the probability and location information of a target. RPN is a fully convolutional network, and its core idea is to use a convolutional network to directly generate candidate regions. Figure 2 As shown in the figure, the RPN network uses a 3×3 spatial pooling window to slide over the feature map of the shared convolutional layer. In order to cope with objects of different sizes, Faster R-CNN uses three different types of sliding windows with aspect ratios of 1:1, 2:1, and 1:2, and uses three scales to scale the sliding windows, a total of nine types of sliding windows. These windows are convolved to form 256-dimensional vectors, and finally the 300 windows with the highest scores are selected as the final recommended windows through classification.

[0027] 2) Vector data conversion process

[0028] The three-dimensional unit vector data of the star points in the star map is input into the convolutional neural network (CNN) for one-dimensional maximum standard deviation features, and the RPN is used to generate the proposed window; the proposed window is then mapped to the feature map of the last convolutional layer of the CNN; the RoI pooling layer is used to generate a fixed-size feature map for each region of interest (RoI); finally, the classification probability and bounding box regression are jointly trained using Softmax Loss and Smooth L1 Loss. During testing, the original candidate window is corrected with the bounding box regression value to generate the predicted window coordinates. In the entire model framework, the multi-task loss function of the Faster R-CNN model is defined as follows:

[0029]

[0030] Where: i is the index; p i is the predicted value of the target bounding box; is the true value of the target bounding box; t i is the predicted value of the category; is the true value of the category; N cls is the normalization parameter of the classification item; N reg is the normalization parameter of the regression term; λ is the balance weight; L cls is the classification loss, i.e., the logarithmic loss of the two categories (target and non-target).

[0031] S102, creating an index table sorted by features and a K-vector lookup table, and performing a K-vector lookup.

[0032] like Figure 3 As shown in the figure, the purpose of introducing K vector search is to realize fast search of navigation star catalog. Then create index table and K vector search table sorted by features. In actual application, when searching for a point, we first convert the one-dimensional standard deviation obtained by S101 step and locate the upper and lower bounds of the feature; then use K vector search table to find the corresponding index range of the feature in the navigation star catalog, which is a circular area around the blue dot. The selection of initial cluster center points of K clusters is crucial to the final result of clustering. In the algorithm, each iteration divides the data point into the cluster with the closest cluster center, and then recalculates the cluster center, and then iterates repeatedly until each data point is no longer re-divided.

[0033] S103, completing star point pixel distance matching based on the K vector lookup table.

[0034] like Figure 4As shown, before matching, the feature vector corresponding to each star point calculated by step S102 needs to be completed. Here, the pixel distance between the detected stars is used as the feature value of the detected star, and the pixel distance between the predicted star No. 1 and the predicted star No. 2 is used as the feature value of the predicted star. Calculate the pixel distance between two star points. For example, if there are N detected stars, each detected star has N-1 feature vectors, and sort them from small to large. After that, traverse the detected stars and predicted stars, and match the feature vectors of the two. If there are one or more similar values ​​in the feature vectors of a detected star and a predicted star, the two may match. The matching confidence of the two can be judged according to the number of similar feature values. Finally, all matching pairs are sorted by confidence, and matching pairs are selected according to the confidence.

[0035] Example 1

[0036] In order to verify the actual improvement of the star sensor algorithm on the dynamic performance of the picostar sensor, this paper developed the algorithm into a star sensor application software, and conducted field stargazing dynamic performance tests on the star sensor product to visually compare the dynamic performance differences between the algorithm software and the old algorithm software. The performance parameters of the star sensor selected for the test are shown in Table 1 below.

[0037] Table 1 The performance parameter of Pico-StarTracker

[0038]

[0039] Figure 5 The measured performance curve of the star sensor after using the prior art and the method of the present invention is shown. Figure 5 From the left and right columns, it can be seen that after the star sensor product is burned with the new method software proposed in this paper, the curve is more stable at the same angular velocity.

[0040] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0041] In summary, the present invention provides a star sensor star map recognition method based on the K-vector search method and the CNN network, which increases the star map recognition speed in the all-day recognition mode and the star tracking mode, and greatly improves the dynamic performance of the star sensor.

[0042] It should be noted that, in the embodiments of the present invention, the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the embodiments, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0043] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0044] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A star sensor star map recognition method based on K vector search method and CNN network, characterized in that: The steps include: S1, the three-dimensional unit vector of the star sensor star map is converted into the one-dimensional feature with the largest standard deviation using the CNN network; S2, create an index table and a K-vector lookup table sorted by features, and perform K-vector lookup; S3, completing star point pixel distance matching based on the K vector lookup table.

2. The star sensor star map recognition method based on K vector search method and CNN network as claimed in claim 1, characterized in that, The step S1 is specifically as follows: obtaining a feature vector according to the CNN network, prioritizing the navigation star points according to the feature vector matching, and sorting the detection stars and the navigation stars. This function is to convert the three-dimensional unit vector into a feature with the largest one-dimensional standard deviation.

3. The star sensor star map recognition method based on K vector search method and CNN network as claimed in claim 1, characterized in that, The step S2 is used for a quick search of the navigation star catalog.

4. The star sensor star map recognition method based on K vector search method and CNN network as claimed in claim 1, characterized in that, The step S3 is specifically as follows: the matching confidence between the two can be determined according to the number of similar eigenvalues ​​found by the K vector, and the matching pair for star map recognition is selected according to the confidence.

Images