CN107883947A - Star sensor method for recognising star map based on convolutional neural networks - Google Patents

Star sensor method for recognising star map based on convolutional neural networks Download PDF

Info

Publication number
CN107883947A
CN107883947A CN201711458120.6A CN201711458120A CN107883947A CN 107883947 A CN107883947 A CN 107883947A CN 201711458120 A CN201711458120 A CN 201711458120A CN 107883947 A CN107883947 A CN 107883947A
Authority
CN
China
Prior art keywords
star
constellation
neural networks
convolutional neural
numbering
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201711458120.6A
Other languages
Chinese (zh)
Other versions
CN107883947B (en
Inventor
吴峰
朱锡芳
徐也
相入喜
于秋阳
缪志康
吴涛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Changzhou Institute of Technology
Original Assignee
Changzhou Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Changzhou Institute of Technology filed Critical Changzhou Institute of Technology
Priority to CN201711458120.6A priority Critical patent/CN107883947B/en
Publication of CN107883947A publication Critical patent/CN107883947A/en
Application granted granted Critical
Publication of CN107883947B publication Critical patent/CN107883947B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01CMEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
    • G01C21/00Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
    • G01C21/02Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 by astronomical means
    • G01C21/025Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 by astronomical means with the use of startrackers
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01CMEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
    • G01C21/00Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
    • G01C21/20Instruments for performing navigational calculations

Abstract

The invention discloses a kind of star sensor method for recognising star map based on convolutional neural networks, comprise the following steps:Star filtration treatment is done to original star catalogue and establishes guide star catalog, counts the constellation belonging to whole day ball nautical star, and constellation is numbered, Sample Storehouse is made up of the numbering of emulation star chart and the most constellations of corresponding star number;Former star chart is replaced with sparse matrix, Sample Storehouse star chart input convolutional neural networks are carried out into training;Shooting gained star chart passes through obtaining star location, and inputs convolutional neural networks after being converted into sparse matrix, carries out thick posture importance in star map recognition, obtains general orientation;With local day area star Pattern Recognition Algorithm, the fixed star in visual field is identified.Thick posture whole day ball importance in star map recognition is realized using the convolutional neural networks trained, without search guide star catalog, local day area importance in star map recognition, it is only necessary to search for fraction database;Convolutional neural networks have the ability of autonomous extraction artwork feature, and applied to importance in star map recognition, antinoise and anti-pseudolite performance are strong.

Description

Star sensor method for recognising star map based on convolutional neural networks
Technical field
The invention belongs to celestial navigation technical field, is related to a kind of method for recognising star map for star sensor.
Background technology
Star Pattern Recognition Algorithm is one of core technology of star sensor, and nearest decades are both at home and abroad for the whole day of spacecraft Autonomous importance in star map recognition has carried out substantial amounts of research, it is proposed that many algorithms, mainly includes:Angular distance algorithm, triangle algorithm, grid Method, binary tree method and algorithm based on pyramid model etc..
Li Xinlu, Yang Jinhua of photoelectric project institute of Changchun University of Science and Technology et al., in order to overcome in triangle recognizer The Redundancy Match problem that triangle is brought because intrinsic dimensionality is relatively low, establish declination band angular distance feature database and choose optimization triangle about Beam condition, importance in star map recognition is carried out using triangulation method is improved.Its Simulation results shows:Emulation experiment obtains statistics and is identified as Power is 99.57%, and the average time of single identification is about 4.28ms.Harbin Teachers' Univ. Xing one is all, for based on unusual Coverage rate relatively low problem in whole day area that the star Pattern Recognition Algorithm that value is decomposed is likely to occur, caused by the optical axis is discontinuous, is carried A kind of improved star Pattern Recognition Algorithm based on singular value decomposition is gone out.In laboratory environments, with traditional triangle algorithm Contrasted in memory capacity, average operating time and discrimination etc., it is found that innovatory algorithm is calculated better than traditional triangle Method.The performance of the antinoise of algorithm above and the interference of anti-pseudolite also needs further to improve.
Li Lihong proposed the star Pattern Recognition Algorithm based on gene genetic in 2000.Its main thought:First with according to quick The vector initialising population of angular distance that each observation star obtains in sensor observation visual field, selects control parameter;Then according to individual institute Comprising angular distance information search Guide star database and calculate ideal adaptation angle value;Finally judge whether that the match is successful.The algorithm With good robustness and real-time, and the capacity of required Guide star database is small.But this method will be carried out every time Optimize, its accuracy and speed can be influenceed by Optimal Parameters.
Graduate school of National University of Defense Technology Nie is lucky to devise two kinds of star Pattern Recognition Algorithms, the i.e. knowledge based on primary matching Other algorithm and the star Pattern Recognition Algorithm based on triangle character vector with characteristic value.Based on the recognizer of primary matching, pass through Increase the mode of Turn angle match, improve the reliability of algorithm, effectively increase the accuracy identified every time.It is special based on triangle Sign vector and the star Pattern Recognition Algorithm of characteristic value, using triangle character vector sum characteristic value as identification feature, by multiple The mode of comparison improves the anti-interference of recognizer, improves the recognition success rate of single width star chart, but this method needs The navigational star table of Large Copacity, the speed of service are also undesirable.
The content of the invention
The purpose of the present invention is:In view of the shortcomings of the prior art, star image simulation and convolutional neural networks are combined, proposed A kind of star sensor method for recognising star map based on convolutional neural networks, saves guide star catalog search time, and raising is identified as Power, strengthen algorithm robustness.
The technical scheme is that:
Star sensor method for recognising star map based on convolutional neural networks, comprises the following steps:
Step 1:Establish Sample Storehouse;Star filtration treatment is done to original star catalogue and establishes guide star catalog, using constellation clustering side Method, the constellation belonging to whole day ball nautical star is counted, and constellation is numbered, Sample Storehouse is i.e. most by emulation star chart and corresponding star number The numbering composition of constellation;
Step 2:Establish simultaneously training convolutional neural networks;Wherein input is star chart, is exported as the most star of star number in star chart Seat numbering, former star chart is replaced with sparse matrix, and Sample Storehouse star chart input convolutional neural networks are carried out into training;
Step 3:Carry out importance in star map recognition;Shooting gained star chart passes through obtaining star location, and inputs volume after being converted into sparse matrix Product neutral net, carries out thick posture importance in star map recognition, obtains general orientation;With local day area star Pattern Recognition Algorithm, visual field is identified Interior fixed star.
Further, it is described to establish Sample Storehouse:First, according to the limiting magnitude of star sensor, star filtering is made to original star catalogue Processing, delete the fixed star that double star, variable and magnitude are higher than limiting magnitude;Then, using constellation clustering method, whole day ball is led Boat star cluster arrives different constellations, and constellation is numbered;Finally, whole day ball is traveled through, for the sensing of each optical axis and posture, generation Star chart is emulated, and counts the numbering of the most constellations of star number in the visual field;Sample Storehouse is i.e. by star chart and the most stars of corresponding star number The numbering composition of seat.
Further, it is described to establish convolutional neural networks:Convolutional neural networks include 5 convolutional layers, 5 pond layers, 2 entirely Articulamentum, it is activation primitive from ReLu (Rectified Linear Units);Input as one with triple sparse matrix The star chart of expression, export as the most constellation numbering of star number in current star chart;Fixed star is calculated in the extraction of the star sensor astrology Coordinate position of the star image in body coordinate system is converted to the triple sparse matrix that number of lines and columns are all 1024, and each asterism accounts for According to 1 pixel, former star chart is replaced with this.
Further, the training convolutional neural networks:Star chart in randomly ordered Sample Storehouse, one by one star chart carried with the astrology Take algorithm to calculate position coordinates of the asterism in body coordinate system, be converted into the triple sparse matrix that number of lines and columns are all 1024 Afterwards, convolutional neural networks are input to, using the numbering of the corresponding most constellations of star number as output, carry out network training;The god trained It is used for thick posture importance in star map recognition through network.
Further, the development importance in star map recognition:Obtained star chart will be shot first, asterism is calculated with astrology extraction algorithm Position coordinates in body coordinate system, after being converted into the triple sparse matrix that number of lines and columns are all 1024, it is input to convolution Neutral net, export the numbering of the most constellations of star number;According to guide star catalog, the inertia for obtaining a star in the numbering constellation is sat Mark system coordinate, the general orientation of current field is calculated;Then, with local day area method for recognising star map, star chart is identified In fixed star.
Further, stating constellation clustering method specific steps includes:
1) setup parameter;It is 1/8th of the angle of visual field to take cluster angle θ, and threshold value t is θ one thousandth, defines a change Cluster is measured, and is initialized as 0, if nautical star is shared M, one is defined and includes the array cnum of M element, and initialize For 0, array cnum elements are corresponding with every nautical star, and for the constellation numbering belonging to recording, element value is 0 to represent corresponding fixed star Not yet pass through clustering processing;
2) right ascension α=0 ° and declination δ=- 90 ° pointed to optical axis are starting point;
3) (α is pointed to for current optical axisii), the fixed star in visual field is counted, provided with N;Define one and include N number of element Array flag [N], all elements are initialized as 0;The constellation numbering of current field fixed star is extracted from array cnum, and is stored in In flag;
4) the element sum that array flag intermediate values are 0 is counted, is set to Nvis;If Nvis=0, by the institute of current field There is star cluster to arrive respective constellation, perform step 9);If Nvis is more than 0, i.e., also star is not processed, then selection is wherein any one Star S0The position at place is original position, if it is (α in inertial coodinate system coordinate00), direction cosines vector V0For
5) statistics and S0Angular distance be less than θ all stars, provided with k star, wherein including fixed star S0, their right ascension and red Latitude is (αj, δj), wherein j ∈ [1, k];It is V that can calculate their direction cosines vector with reference to formula (1)j, they and S0Angle Away from according to formula (2) calculating;
6) if the constellation numbering of this k star is all 0, cluster increases by 1, and cluster is assigned in array flag Corresponding element, perform step 7);Otherwise, the minimum value of this k star non-zero constellation numbering is counted, is set to fm;In current field Inside distributed in the fixed star of constellation numbering, if in the presence of and k star seat number equal fixed star, then, their constellation is compiled Number fm is entered as again, this k star seat numbering also all be fm, is arrived all neighbour's fixed stars clusters being connected with this same Constellation, and they unified constellation numbering;
7) according to formula (3), this k star center S direction cosines vector is calculated, is set to Vnew
Wherein
If 8) original position S0It is less than t with new center S angular distance, then completes the constellation clustering of current field, Perform step 4);Otherwise by VnewIt is assigned to V0, S0It is moved at the S of position, return to step 5);
9) array flag each element values are assigned to corresponding element in array cnum, optical axis points to next orientation, return to step 3), until traversal terminates.
Further, if the focal length of shooting star chart, the line direction angle of visual field, the column direction angle of visual field are respectively f, wa、wbIf the astrology Coordinate in body coordinate system is (x, y ,-f), then after converting it to sparse matrix, the row, column number at place is respectively
Wherein round functions are to round up.
Compared with prior art, the present invention has the characteristics of following:
(1) guide star catalog search time is short.Thick posture whole day soccer star figure is realized using the convolutional neural networks trained Identification, without search guide star catalog.Local day area importance in star map recognition, it is only necessary to search for fraction database.
(2) strong robustness.Convolutional neural networks have the ability of autonomous extraction artwork feature, applied to importance in star map recognition, resist Noise and anti-pseudolite performance are strong.
Brief description of the drawings
Fig. 1 is method for recognising star map flow chart of the present invention.
Fig. 2 is constellation clustering flow chart.
Fig. 3 is Distribution of guide stars of the embodiment one when optical axis points to coordinate (120 °, 20 °) place on celestial sphere.
Fig. 4 is embodiment one when optical axis points to coordinate (120 °, 20 °) place on celestial sphere, and around optical axis rotate 30 ° when navigation Star is distributed.
Embodiment
The present invention is described in further detail below in conjunction with the accompanying drawings.
A kind of star sensor method for recognising star map based on convolutional neural networks of the present invention is as shown in figure 1, including following step Suddenly:
(1) Sample Storehouse is established.First, according to the limiting magnitude of star sensor, make star filtration treatment to original star catalogue, delete Double star, variable and magnitude are higher than the fixed star of limiting magnitude.Then, using constellation clustering method as shown in Figure 2, by whole day ball Nautical star cluster arrives different constellations, and constellation is numbered.Finally, whole day ball is traveled through, it is raw for the sensing of each optical axis and posture Into emulation star chart, and count the numbering of the most constellations of star number in the visual field.Sample Storehouse is most by star chart and corresponding star number The numbering composition of constellation.
(2) convolutional neural networks are established.Convolutional neural networks include 5 convolutional layers, 5 pond layers, 2 full articulamentums, It is activation primitive from ReLu (Rectified Linear Units).Input as a star represented with triple sparse matrix Figure, export as the most constellation numbering of star number in current star chart.It is directly defeated because the star chart resolution ratio of star sensor shooting is higher When entering to convolutional neural networks, operational efficiency is low.Sat therefore, the extraction of the star sensor astrology is calculated into fixed star star image in body Coordinate position in mark system is converted to the triple sparse matrix that number of lines and columns are all 1024, and each asterism occupies 1 pixel, with This replaces former star chart.
(3) training convolutional neural networks.Star chart in randomly ordered Sample Storehouse, one by one star chart use astrology extraction algorithm meter Calculate position coordinates of the asterism in body coordinate system, after being converted into the triple sparse matrix that number of lines and columns are all 1024, input To convolutional neural networks, using the numbering of the corresponding most constellations of star number as output, carry out network training.The neutral net trained is used In thick posture importance in star map recognition.
(4) importance in star map recognition is carried out.Obtained star chart will be shot first, calculate asterism with astrology extraction algorithm sits in body Position coordinates in mark system, after being converted into the triple sparse matrix that number of lines and columns are all 1024, is input to convolutional neural networks, Export the numbering of the most constellations of star number.According to guide star catalog, the inertial coodinate system coordinate of a star in the numbering constellation is obtained, The general orientation of current field is calculated.Then, with local day area method for recognising star map, the fixed star in star chart is identified.
Star sensor method for recognising star map specific implementation process of the present invention based on convolutional neural networks is as follows.
The first step, establish Sample Storehouse.(1) according to the limiting magnitude of star sensor, star filtration treatment is made to original star catalogue, deleted Except double star, variable and magnitude are higher than the fixed star of limiting magnitude, using remaining star as nautical star.(2) constellation clustering method is used, Whole day ball nautical star cluster is arrived into different constellations, and constellation is numbered.(3) whole day ball is traveled through, is pointed to for each optical axis, choosing With 10 kinds of postures, generation emulates star chart, and counts the numbering of the most constellations of star number in visual field.Sample Storehouse is i.e. by star chart and correspondingly Constellation numbering composition.Wherein constellation clustering method concretely comprises the following steps
(1) setup parameter.It is 1/8th of the angle of visual field to take cluster angle θ, and threshold value t is θ one thousandth.Define a change Cluster is measured, and is initialized as 0.If nautical star is shared M, defines one and include the array cnum of M element, and initialize For 0.Array cnum elements are corresponding with every nautical star, numbered for the constellation belonging to recording, and element value is the corresponding fixed star of 0 expression Not yet pass through clustering processing.
(2) right ascension α=0 ° and declination δ=- 90 ° pointed to optical axis are starting point.
(3) (α is pointed to for current optical axisii), the fixed star in visual field is counted, provided with N.Define one and include N number of member The array flag [N] of element, all elements are initialized as 0.The constellation numbering of current field fixed star is extracted from array cnum, and is deposited Enter in flag.
(4) the element sum that array flag intermediate values are 0 is counted, is set to Nvis.If Nvis=0, by current field All star clusters arrive respective constellation, perform step (9).If Nvis is more than 0, i.e., also star is not processed, then selection is wherein appointed Anticipate a star S0The position at place is original position, if it is (α in inertial coodinate system coordinate00), direction cosines vector V0For
(5) statistics and S0Angular distance be less than θ all stars, provided with k star, wherein including fixed star S0, their right ascension and red Latitude is (αj, δj), wherein j ∈ [1, k].It is V that can calculate their direction cosines vector with reference to formula (1)j, they and S0Angle Away from according to formula (2) calculating.
(6) if the constellation numbering of this k star is all 0, cluster increases by 1, and cluster is assigned into array flag In corresponding element, perform step (7).Otherwise, the minimum value of this k star non-zero constellation numbering is counted, is set to fm.Working as forward sight Distributed in the fixed star of constellation numbering, if in the presence of and the equal fixed star of k star seat numbering, then, their constellation Numbering is entered as fm again, and this k star seat numbering is also all fm, is clustered all neighbour's fixed stars being connected to same with this Individual constellation, and they unified constellation numbering.
(7) according to formula (3), this k star center S direction cosines vector is calculated, is set to Vnew
Wherein
(8) if original position S0It is less than t with new center S angular distance, then completes the constellation clustering of current field, Perform step (4).Otherwise by VnewIt is assigned to V0, S0It is moved at the S of position, return to step (5).
(9) array flag each element values are assigned to corresponding element in array cnum, optical axis points to next orientation, return to step (3), until traversal terminates.
Second step, convolutional neural networks include 5 convolutional layers, 5 pond layers, 2 full articulamentums, the convolution face of convolutional layer Have 6,12,18,12,6 respectively, except the convolution kernel size of the second convolutional layer is 7 × 7, remaining is all 5 × 5, pond layer sliding window Mouth size is all 2 × 2, is activation primitive from ReLu (Rectified Linear Units).Input as one with triple The star chart that sparse matrix represents, export as the most constellation numbering of star number in current star chart.Due to the star chart of star sensor shooting Resolution ratio is higher, and when being directly inputted to convolutional neural networks, operational efficiency is low.Therefore, the extraction of the star sensor astrology is calculated The triple sparse matrix that number of lines and columns are all 1024 is converted to coordinate position of the fixed star star image in body coordinate system, each Asterism occupies 1 pixel, and former star chart is replaced with this.If shoot the focal length of star chart, the line direction angle of visual field, column direction angle of visual field difference For f, wa、wbIf coordinate of the astrology in body coordinate system is (x, y ,-f), then after converting it to sparse matrix, place Row, column number is respectively
Wherein round functions are to round up.
3rd step, training convolutional neural networks.Star chart in randomly ordered Sample Storehouse, star chart is with astrology extraction calculation one by one Method calculates position coordinates of the asterism in body coordinate system, after being converted into the triple sparse matrix that number of lines and columns are all 1024, Convolutional neural networks are input to, using the numbering of the corresponding most constellations of star number as output, carry out network training.Train neutral net For thick posture importance in star map recognition.
4th step, carry out importance in star map recognition.Obtained star chart will be shot first, asterism is calculated at this with astrology extraction algorithm Position coordinates in body coordinate system, after being converted into the triple sparse matrix that number of lines and columns are all 1024, it is input to convolutional Neural Network, export the numbering of the most constellations of star number.According to guide star catalog, the inertial coodinate system of a star in the numbering constellation is obtained Coordinate, the general orientation of current field is calculated.Then, with local day area method for recognising star map, identify in star chart Fixed star.
Below in conjunction with the accompanying drawings and embodiment the invention will be further described.
Embodiment one
It is original star catalogue to choose SAO star catalogues, and limiting magnitude is 5.2 grades, the angle of visual field is 20 ° × 20 °, star sensor detector Resolution ratio is 1024 × 1024, focal length 43.56mm.After doing star filtration treatment to original star catalogue, 1607 stars are obtained, choosing is done Nautical star.When optical axis is oriented to the position of coordinate on celestial sphere (120 °, 20 °), Distribution of guide stars in visual field as shown in figure 3, 16 stars are shared, details are as shown in table 1.
Travel through constellation clustering by whole day ball, their constellation numbering is as shown in table 1, wherein serial number 9,10,12,13, 14th, 15, the 16 constellation star numbers that totally 7 stars form are most, numbering 410.
When optical axis point to it is constant, visual field after optical axis rotates 30 °, obtained star chart as shown in figure 4, share 15 stars, it Position in visual field it is as shown in table 2.For part star in Fig. 3 outside visual field, the star of serial number 17 and 18 is newly to go out Existing.Now the constellation star number that totally 5 stars form of serial number 9,10,12,14,16 is most, numbering 410.
Fixed star data in the visual field of table 1
For star chart after the astrology extract, the coordinate value in body coordinate system is converted to triple sparse matrix, astrology phase The X and the line number of Y-direction and row number answered is as shown in table 1,2.Represent the triple sparse matrix and star number most constellations of star chart Numbering is the input and output of convolutional neural networks.The parameter of convolutional neural networks, including 5 convolutional layers, 5 ponds are set Layer, 2 full articulamentums, the convolution face of convolutional layer have 6,12,18,12,6 respectively, except the convolution kernel size of the second convolutional layer is 7 × 7, remaining is all 5 × 5, and pond layer sliding window size is all 2 × 2, from ReLu (Rectified Linear Units) For activation primitive.All convolution nuclear elements and weights initialisation are a random number.Convolutional neural networks after training terminates are used for Thick posture importance in star map recognition, the most constellation numbering of output star number.
Fixed star data after the visual field of table 2 rotates 30 degree
If inputting Fig. 3,410 are exported.Because the constellation that numbering is 410 includes the star of serial number 12, its inertial system it is red Through being respectively 2.11 and 0.49 with declination.Then using known local day area method of identification, it is only necessary in above right ascension and declination institute It is corresponding nearby to carry out local day area importance in star map recognition in small range day area, identify the fixed star in visual field, it is not necessary to wide area search Star storehouse.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention.All essences in the present invention All any modification, equivalent and improvement made within refreshing and principle etc., should be included in the scope of the protection.

Claims (7)

1. the star sensor method for recognising star map based on convolutional neural networks, comprises the following steps:
Step 1:Establish Sample Storehouse;Star filtration treatment is done to original star catalogue and establishes guide star catalog, using constellation clustering method, system The constellation belonging to whole day ball nautical star is counted, and constellation is numbered, Sample Storehouse is i.e. by emulation star chart and the most constellations of corresponding star number Numbering composition;
Step 2:Establish simultaneously training convolutional neural networks;Wherein input is star chart, exports and is compiled for the most constellation of star number in star chart Number, former star chart is replaced with sparse matrix, Sample Storehouse star chart input convolutional neural networks are carried out into training;
Step 3:Carry out importance in star map recognition;Shooting gained star chart passes through obtaining star location, and inputs convolution god after being converted into sparse matrix Through network, carry out thick posture importance in star map recognition, obtain general orientation;With local day area star Pattern Recognition Algorithm, identify in visual field Fixed star.
2. the star sensor method for recognising star map according to claim 1 based on convolutional neural networks, it is characterised in that:Institute State and establish Sample Storehouse:First, according to the limiting magnitude of star sensor, make star filtration treatment to original star catalogue, delete double star, become Star and magnitude are higher than the fixed star of limiting magnitude;Then, using constellation clustering method, whole day ball nautical star cluster is arrived different Constellation, and constellation is numbered;Finally, whole day ball is traveled through, for the sensing of each optical axis and posture, generation emulation star chart, and is counted The numbering of the most constellations of star number in the visual field;Sample Storehouse is made up of the numbering of star chart and the most constellations of corresponding star number.
3. the star sensor method for recognising star map according to claim 1 based on convolutional neural networks, it is characterised in that:Institute State and establish convolutional neural networks:Convolutional neural networks include 5 convolutional layers, 5 pond layers, 2 full articulamentums, from ReLu (Rectified Linear Units) is activation primitive;Input as a star chart represented with triple sparse matrix, output For the most constellation numbering of star number in current star chart;Fixed star star image is calculated in body coordinate system in the extraction of the star sensor astrology In coordinate position be converted to the triple sparse matrix that number of lines and columns are all 1024, each asterism occupies 1 pixel, with this generation For former star chart.
4. the star sensor method for recognising star map according to claim 1 based on convolutional neural networks, it is characterised in that:Institute State training convolutional neural networks:Star chart in randomly ordered Sample Storehouse, one by one star chart with astrology extraction algorithm calculate asterism exist Position coordinates in body coordinate system, after being converted into the triple sparse matrix that number of lines and columns are all 1024, it is input to convolution god Through network, using the numbering of the corresponding most constellations of star number as output, carry out network training;The neutral net trained is used for thick posture Importance in star map recognition.
5. the star sensor method for recognising star map according to claim 1 based on convolutional neural networks, it is characterised in that:Institute State development importance in star map recognition:Obtained star chart will be shot first, asterism is calculated in body coordinate system with astrology extraction algorithm Position coordinates, after being converted into the triple sparse matrix that number of lines and columns are all 1024, convolutional neural networks are input to, export star number The numbering of most constellations;According to guide star catalog, the inertial coodinate system coordinate of a star in the numbering constellation is obtained, is calculated The general orientation of current field;Then, with local day area method for recognising star map, the fixed star in star chart is identified.
6. the star sensor method for recognising star map according to claim 1 based on convolutional neural networks, it is characterised in that:Institute Stating constellation clustering method specific steps includes:
1) setup parameter;It is 1/8th of the angle of visual field to take cluster angle θ, and threshold value t is θ one thousandth, defines a variable Cluster, and 0 is initialized as, if nautical star is shared M, defines one and include the array cnum of M element, and be initialized as 0, array cnum elements are corresponding with every nautical star, and for recording the constellation numbering belonging to fixed star, element value is that 0 expression is corresponding permanent Star not yet passes through clustering processing;
2) right ascension α=0 ° and declination δ=- 90 ° pointed to optical axis are starting point;
3) (α is pointed to for current optical axisii), the fixed star in visual field is counted, provided with N;Define a number for including N number of element Group flag [N], all elements are initialized as 0;The constellation numbering of current field fixed star is extracted from array cnum, and is stored in flag In;
4) the element sum that array flag intermediate values are 0 is counted, is set to Nvis;If Nvis=0, by all stars of current field Respective constellation is clustered, performs step 9);If Nvis is more than 0, i.e., also star is not processed, then selection wherein any one star S0The position at place is original position, if it is (α in inertial coodinate system coordinate00), direction cosines vector V0For
<mrow> <msub> <mi>V</mi> <mn>0</mn> </msub> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <msub> <mi>cos&amp;alpha;</mi> <mn>0</mn> </msub> <msub> <mi>cos&amp;delta;</mi> <mn>0</mn> </msub> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>sin&amp;alpha;</mi> <mn>0</mn> </msub> <msub> <mi>cos&amp;delta;</mi> <mn>0</mn> </msub> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>sin&amp;delta;</mi> <mn>0</mn> </msub> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow>
5) statistics and S0Angular distance be less than θ all stars, provided with k star, wherein including fixed star S0, their right ascension and declination be (αj, δj), wherein j ∈ [1, k];It is V that can calculate their direction cosines vector with reference to formula (1)j, they and S0Angular distance press Calculated according to formula (2);
6) if the constellation numbering of this k star is all 0, cluster increases by 1, and cluster is assigned into pair in array flag Element is answered, performs step 7);Otherwise, the minimum value of this k star non-zero constellation numbering is counted, is set to fm;In current field Distribute in the fixed star of constellation numbering, if in the presence of and k star seat number equal fixed star, then, their constellation numbering weight Fm newly is entered as, this k star seat numbering is also all fm, and all neighbour's fixed star clusters being connected are arrived into same star with this Seat, and they unified constellation numbering;
7) according to formula (3), this k star center S direction cosines vector is calculated, is set to Vnew
<mrow> <msub> <mi>V</mi> <mrow> <mi>n</mi> <mi>e</mi> <mi>w</mi> </mrow> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <msqrt> <mrow> <msup> <mi>a</mi> <mn>2</mn> </msup> <mo>+</mo> <msup> <mi>b</mi> <mn>2</mn> </msup> <mo>+</mo> <msup> <mi>c</mi> <mn>2</mn> </msup> </mrow> </msqrt> </mfrac> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mi>a</mi> </mtd> </mtr> <mtr> <mtd> <mi>b</mi> </mtd> </mtr> <mtr> <mtd> <mi>c</mi> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>3</mn> <mo>)</mo> </mrow> </mrow>
Wherein
If 8) original position S0It is less than t with new center S angular distance, then completes the constellation clustering of current field, perform step It is rapid 4);Otherwise by VnewIt is assigned to V0, S0It is moved at the S of position, return to step 5);
9) array flag each element values are assigned to corresponding element in array cnum, optical axis points to next orientation, return to step 3), directly Terminate to traversal.
7. the star sensor method for recognising star map according to claim 3 based on convolutional neural networks, it is characterised in that:If The focal length of shooting star chart, the line direction angle of visual field, the column direction angle of visual field are respectively f, wa、wbIf seat of the astrology in body coordinate system It is designated as (x, y ,-f), then after converting it to sparse matrix, the row, column number at place is respectively
Wherein round functions are to round up.
CN201711458120.6A 2017-12-28 2017-12-28 Star sensor star map identification method based on convolutional neural network Active CN107883947B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711458120.6A CN107883947B (en) 2017-12-28 2017-12-28 Star sensor star map identification method based on convolutional neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711458120.6A CN107883947B (en) 2017-12-28 2017-12-28 Star sensor star map identification method based on convolutional neural network

Publications (2)

Publication Number Publication Date
CN107883947A true CN107883947A (en) 2018-04-06
CN107883947B CN107883947B (en) 2020-12-22

Family

ID=61770440

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711458120.6A Active CN107883947B (en) 2017-12-28 2017-12-28 Star sensor star map identification method based on convolutional neural network

Country Status (1)

Country Link
CN (1) CN107883947B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109813303A (en) * 2019-03-08 2019-05-28 北京航空航天大学 A kind of method for recognising star map for not depending on calibrating parameters based on the ballot of angle mould formula cluster
CN111156988A (en) * 2019-12-31 2020-05-15 中国科学院紫金山天文台 Space debris astronomical positioning and photometry method based on automatic pointing error determination

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB112012A (en) * 1916-12-12 1918-12-19 Henri Louis Victor Desire Giot Improvements in Stars' Transit Calculators and Indicators for Mariners.
CN102840860A (en) * 2012-08-30 2012-12-26 北京航空航天大学 Hybrid particle swarm algorithm-based star graph recognition method
CN103363987A (en) * 2013-06-26 2013-10-23 哈尔滨工业大学 Star map identification method of multi-view-field star sensor
CN105547286A (en) * 2016-01-11 2016-05-04 常州工学院 Composite three-view-field star sensor star map simulation method
CN106971189A (en) * 2017-03-31 2017-07-21 西北工业大学 A kind of noisy method for recognising star map of low resolution

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB112012A (en) * 1916-12-12 1918-12-19 Henri Louis Victor Desire Giot Improvements in Stars' Transit Calculators and Indicators for Mariners.
CN102840860A (en) * 2012-08-30 2012-12-26 北京航空航天大学 Hybrid particle swarm algorithm-based star graph recognition method
CN103363987A (en) * 2013-06-26 2013-10-23 哈尔滨工业大学 Star map identification method of multi-view-field star sensor
CN105547286A (en) * 2016-01-11 2016-05-04 常州工学院 Composite three-view-field star sensor star map simulation method
CN106971189A (en) * 2017-03-31 2017-07-21 西北工业大学 A kind of noisy method for recognising star map of low resolution

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
李春艳: ""利用神经网络技术实现星敏感器的星图识别"", 《中国优秀博硕士学位论文全文数据库 (硕士) 工程科技Ⅱ辑》 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109813303A (en) * 2019-03-08 2019-05-28 北京航空航天大学 A kind of method for recognising star map for not depending on calibrating parameters based on the ballot of angle mould formula cluster
CN109813303B (en) * 2019-03-08 2020-10-09 北京航空航天大学 Star map identification method independent of calibration parameters based on angular pattern cluster voting
CN111156988A (en) * 2019-12-31 2020-05-15 中国科学院紫金山天文台 Space debris astronomical positioning and photometry method based on automatic pointing error determination
CN111156988B (en) * 2019-12-31 2021-03-16 中国科学院紫金山天文台 Space debris astronomical positioning and photometry method based on automatic pointing error determination

Also Published As

Publication number Publication date
CN107883947B (en) 2020-12-22

Similar Documents

Publication Publication Date Title
CN105975931A (en) Convolutional neural network face recognition method based on multi-scale pooling
CN106683091A (en) Target classification and attitude detection method based on depth convolution neural network
CN105223546B (en) Indoor orientation method based on received signal strength and reference point locations double focusing class
CN105956532A (en) Traffic scene classification method based on multi-scale convolution neural network
CN106980895A (en) Convolutional neural networks Forecasting Methodology based on rotary area
CN107403163B (en) A kind of laser SLAM closed loop automatic testing method based on deep learning
CN107883947A (en) Star sensor method for recognising star map based on convolutional neural networks
CN104090972A (en) Image feature extraction and similarity measurement method used for three-dimensional city model retrieval
CN109190684A (en) SAR image sample generating method based on sketch and structural generation confrontation network
CN107220611A (en) A kind of space-time feature extracting method based on deep neural network
CN104463856B (en) The ground extracting method of the outdoor scene three dimensional point cloud based on normal vector ball
CN108647738A (en) Global Scale remote sensing image water body intelligent extract method based on multi index option
CN107292339A (en) The unmanned plane low altitude remote sensing image high score Geomorphological Classification method of feature based fusion
CN106897538B (en) Geomagnetic chart direction suitability computational methods based on convolutional neural networks
CN107273938A (en) Multi-source Remote Sensing Images terrain classification method based on binary channels convolution ladder net
CN104154929B (en) Optimal selection method of navigational stars of star map simulator based on star density
CN103578093B (en) Method for registering images, device and augmented reality system
CN105447274A (en) Method of performing coastal wetland drawing for medium-resolution remote sensing image by utilizing object-oriented classification technology
CN107066941A (en) A kind of face identification method and system
CN110163187A (en) Remote road traffic sign detection recognition methods based on F-RCNN
CN110163836A (en) Based on deep learning for the excavator detection method under the inspection of high-altitude
CN106529484A (en) Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning
CN104318215B (en) A kind of cross-view face identification method based on domain robust convolution feature learning
CN107644230A (en) A kind of spatial relationship modeling method of remote sensing images object
CN105547286B (en) A kind of compound three visual fields star sensor star map simulation method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant