Non-Dimensional Star Identification for Un-Calibrated Star Cameras

Transcription

1 Non-Dimensional Star Identification for Un-Calibrated Star Cameras Malak A. Samaan, Daniele Mortari, and John L. Junkins Department of Aerospace Engineering, Texas A&M University, College Station, TX Abstract Star identification is the most critical and important process for attitude estimation, given data from any star sensor. The main purpose of the Star-ID is to identify the observed/measured stars with the corresponding cataloged stars. The precision of the observed star directions highly depend on the calibrated accuracy of the star camera parameters, mainly the focal length f, and the optical axis offsets (x 0, y 0 ). When these parameters are not accurate or when the camera is not well calibrated, the proposed Non-Dimensional Star-ID method becomes very suitable, because it does not require accurate knowledge of these parameters. The Non-Dimensional Star-ID method represents a unique tool to identify the stars of un-calibrated or inaccurate parameters cameras. The basic idea derives the identification process from some observed parameters (the focal plane angles) which are, to first order, independent from both the focal length and the optical axis offsets. The adoption of the k- vector range search technique, makes this method very fast. Moreover, it is easy to implement, accurate, and the probability of failing star-id is less than 0.1% for typical star tracker design parameters. 1. INTRODUCTION For successful Star-ID, some camera parameters such as effective focal length, and optical axis offsets, should be determined before any space mission. These parameters can be estimated in the laboratory or using a recently developed ground calibration algorithm (Ref. [1]). However, due to some in flight distortion of the focal plane, or Derived from paper AAS presented at the 13th Annual AAS/AIAA Space Flight Mechanics Meeting, 9-13 February 2003, Ponce, Puerto Rico. Post-Doc Research Associate, Spacecraft Technology Center, Texas A&M University, College Station, TX , Tel. (979) , samaan@tamu.edu Associate Professor, Department of Aerospace Engineering, Texas A&M University, College Station, Texas Tel. (979) , FAX (979) , mortari@tamu.edu George J. Eppright Chair Professor, Director of the Center for Mechanics and Control, Department of Aerospace Engineering, Texas A&M University, College Station, TX , Tel: (979) , Fax: (979) , junkins@tamu.edu 1

2 some cyclic thermal deformations, these parameters may not remain accurate over the mission lifetime. Therefore, a method that can accomplish the Star-ID process for poorly calibrated cameras, or when previous accurate camera parameters are not accurate for a camera undergoing environmental change, is highly required. Such a method is introduced in this paper: The Non-Dimensional Star-ID method. The Non-Dimensional Star-ID method is used herein to identify stars without a priori attitude information. Several algorithms have been developed before to solve the problem of the Lost-In-Space star identification as in Ref. [2] and Ref. [3], but all these algorithms depend on an accurate values of the camera parameters, while in this paper we present a novel method that can solve the same problem with poorly calibrated or un-calibrated star cameras. In particular, it is demonstrated that the relation between the angles of the cataloged triangles (α i ) and the corresponding angles of focal plane measured triangles (β i ) are independent, to the first order, of the camera basic parameters (x 0, y 0, f). Their differences will reflect primarily centroiding errors and, secondary, higher order distortions. For the typical case, usual manufacturing tolerance will result in these higher order distortions being sufficiently small that the Non-Dimensional Star-ID method is very reliable. To verify our results, both night-sky tests and Monte Carlo simulations will be used to validate the proposed Non-Dimensional Star-ID method. Figure 1 shows two star images for different focal lengths, and optical axis offsets. It is possible to demonstrate that, on a first approximation, the angles of focal plane triangles are independent of the focal length, the rotation, and the optical axis offsets. The fact that variations in (x 0, y 0, f) simply shift and scale the triangles (to first order), means that using in-plane angles for Star-ID can likely be done with very poor estimates of (x 0, y 0, f). The Non-Dimensional Star-ID algorithm can also be used as the first step for the focal plane calibration to find camera basic parameters without any need for manual Star-ID. 2. FOCAL PLANE AND INERTIAL ANGLES Assuming that the star tracker can be modelled as an ideal pin-hole camera (see Figure 2), the body vector (b i ) of any star measurement is related to the corresponding inertial unit-vector ˆr i through b i = (x i x 0 ) (y i y 0 ) f = m i ˆb i = m i A ˆr i (1) 2

3 where A is the attitude matrix, m i = (x i x 0 ) 2 + (y i y 0 ) 2 + f 2, f is the focal length, and (x 0, y 0 ) are the optical axis offsets. Equation (1) allows us to write b ij = b j b i = b ik = b k b i = x i x j y i y j 0 x i x k y i y k 0 = A(m jˆr j m iˆr i ) = A(m kˆr k m iˆr i ) The angle α is simply evaluated from the reference frame vectors by (2) cos α = ( r ij) T ( r ik ) r ij r ik (3) Also, assuming the usual pin-hole camera model, the corresponding focal plane angle is evaluated as cos β = ( b ij) T ( b ik ) b ij b ik = (x j x i )(x k x i ) + (y j y i )(y k y i ) (xj x i ) 2 + (y j y i ) 2 (x k x i ) 2 + (y k y i ) 2 (4) Now, the vector modulus may be expressed as m i = f(1 + δ i ) δ i = m i f f where i = j, k. Now, for a 10 pin-hole camera (max(φ i ) = 5, cos φ i = and f = m i cos φ i ), we have the upper limit δ i = (cos φ i ) 1 1 < Therefore, Eq. (2) becomes { bij = f A [ (1 + δ j ) ˆr j (1 + δ i ) ˆr i ] = f A [ ˆr j ˆr i + δ j ˆr j δ i ˆr i ] (6) b ik = f A [ (1 + δ k ) ˆr k (1 + δ i ) ˆr i ] = f A [ ˆr k ˆr i + δ k ˆr k δ i ˆr i ] We can notice the last two terms in Eq. (6) are very small < Substituting Eq. (6) into Eq. (4), and using A T A = I, we obtain (5) where, cos β = (r j r i ) T (r k r i ) + N + ε D 1 D 2 (7) N = δ j rj T (r k r i ) δ i ri T (r k r i ) + δ k rk T (r j r i ) δ i ri T (r j r i ) D 1 = (r j r i ) T (r j r i ) + 2(δ j r j δ i r i ) T (r j r i ) D 2 = (r k r i ) T (r k r i ) + 2(δ k r k δ i r i ) T (r k r i ) ε = δ j δ k rj T r k δ j δ i ri T r j δ i δ k ri T r k + δi 2 (8) For negligible (δ i, δ j, δ k ), Eq. (7) will be identical to Eq. (3), and then β = α. Notice that cos β depends on (x 0, y 0, f) only through the small (δ i, δ j, δ k ) terms. 3

4 Since the δ i terms are all (at worst) of the order , small errors in (x 0, y 0, f) just perturb small terms, making sensitivity with respect to (x 0, y 0, f) errors very low. For example, the δ i variation due to f error, can be written as δ i = δ i f +h.o.t. = f m i f 1 f m i f f 2 f +h.o.t. = 1 cos φ i m i f f f +h.o.t. The first order approximation of equation (9) for the δ i variation due to f error is equal to zero, but the higher order approximation create small errors of δ i variation due to f. For a typical example if f = 60 mm, and f = 30 mm (which is very large error in focal length!), we obtain δ i < For a more reasonable 1 mm error in f, we find δ i < 10 6, which is smaller than typical centroiding errors of modern star trackers. So, the worst case variation in δ i is at least one order of magnitude smaller than δ i. This indicates that cos β has low sensitivity to any reasonable errors in f. This error in the cos β is typically lower than the tolerance (associated with camera accuracy) for our search. Even when we ignore δ i corrections, we can achieve reliable Star-ID by considering redundant measured stars, as shown in the next section. (9) 3. THE NON-DIMENSIONAL STAR-ID The angles of a given catalog star triangle are evaluated as follows cos α 1 = rt ijr ik r ik r ik, cos α 2 = rt jkr ij r jk r ij, cos α 3 = rt ikr jk r ik r jk where r ij = r j r i, r ik = r k r i, and r kj = r j r k. The indices i, j, and k are chosen such that the minimum angle of the triangle (α 3 ) is always at vertex k. The indices of the cataloged vectors and the maximum and minimum angles of the focal plane triangles are stored in a matrix. Table 1 shows a portion of this star triangle data matrix sorted in ascending order of minimum angle (α 3 ). The reference star catalog is obtained, in this paper, from Smithsonian Astrophysical Observatory (SAO) data in epoch 2000 reference frame. The size of this matrix depends on the magnitude threshold. For instance, for M th = 5.0 and 8 8 FOV, the number of admissible triangles is 55,309, while for M th = 5.5 and the same FOV size, the number of admissible triangles becomes 338,369. Figure 3 shows a plot of the smallest angle (α 3 ) versus the focal plane triangle index (k). In particular, no prior knowledge of (x 0, y 0, f) has been assumed, and the δ i and δi 2 corrections in Eq. (10), have been ignored. We adopt a worst case tolerance for angle matching of rad corresponding to the maximum δ i possible over FOV. Let us divide the smallest angle (α 3 ) into very small and uniform intervals of step size (δα 3 ), and let us interpolate the value of (k) at any value of (0 < α 3 < 60 ). So, 4 (10)

5 for (δα 3 = ) we can store the values of α 3 and the corresponding index value k. Figure 4 illustrates this idea. Even though Fig. 3 looks smooth, there is evident fine structure variations on the finer scale of Fig. 4. The uniform in (α 3 ) table, a portion of which plotted in Fig. 4 is introduced to permit ultra high speed, search-less interpolation of the index k as a function of measured values of α 3. Now, once we have the coordinates of any star image (x i, y i ) we can find the focal plane angles established by any three stars. The smallest angle of each triangle is used to find without a search the index range (from k l to k u ) using the range searching k-vector technique (as described in Ref. [7]) ( ) α3 α 3error k l = floor δα 3 and ( ) α3 + α 3error k u = ceil δα 3 using this range and by accessing the stored matrix for the cataloged indices we can find the range of the candidate triangles. By checking the largest angle α 1 we can find exactly which triangle in the catalogue corresponds to measured one. As mentioned before, the worst case variation in δ i is at least one order of magnitude smaller than δ i, So we use a tolerance of α 3error = 500µrad corresponding to the maximum error in using Eq. (10) to approximate Eq. (7). For this crude tolerance, we may occasionally fail if 3 or 4 stars are matched, but adding a 5 th star results in a virtually certain match. Table 1 contains entries: k, α 3 (k), α 1 (k), I 1 (k), I 2 (k), I 3 (k), where I i (k) are the three stars indices for the k-th cataloged triangle. The master catalog contains entries I, M v (I), λ(i), µ(i), where I is the Star-ID integer, M v (I) is the visual magnitude, λ(i) is the right ascension and µ(i) is the declination. If more than one triangle is found this triad of stars is skipped, and another triangle with different stars is formed until a unique matched triangle is found. If a unique triangle is found (most common case), the Star-ID is then confirmed to very high confidence using a 4-th and 5-th star to form new triangles sharing two stars with the first matched triangle. The rule is that all the inter-star angles of the polygon must match to within a given tolerance (6 angles for a 4-star, and 10 for a 5-star pattern). For the additional triangles the indices of the two common stars must also match. When this fails, then the initial triangle match is rejected. The complete successful match represents the most common case, since the frequency of mismatch of four stars is on the order of (for a 5.5 magnitude threshold, 8 8 FOV, and CCD format). (11) 4. NON-DIMENSIONAL STAR-ID ALGORITHM RESULTS The proposed method has been coded in MATLAB and it has been successfully validated by both night-sky tests and Monte-Carlo simulations. 5

6 4.1. Using Night-Sky Tests Using the StarNav I prototype camera, the Pegasus (Ref. [4]), some nightsky tests have been performed. Figure 5 shows an actual star image centered near λ = 92, µ = 31. Figure 6 shows the coordinates of the star image evaluated by the centroiding algorithm (Ref. [5]). After applying the Non-Dimensional Star-ID we will get the Star-ID results shown in Table 2. Figure 7 shows the results obtained by the least square algorithm (see Ref. [1]) in estimating the optical axis offsets and the focal length. This provides us with the calibrated focal length f c = mm, the calibrated y-axis offset x 0 = mm, and the calibrated x-axis offset y 0 = mm. Moreover, using the GIFTS prototype camera Star1000 ( pixels), additional night-sky tests, have been performed. Figure 8 shows a night-sky picture taken by the Star1000 camera. Figure 9 shows the nine star coordinates (of the picture taken by the Star1000 Camera) obtained by the centroiding algorithm (Ref. [5]). The Star-ID results, obtained with the Non-Dimensional Star-ID Algorithm, are shown in Table 3. These results validate the identification process Using Monte-Carlo Simulations MATLAB simulations have been performed to verify the above results random tests, for various random spacecraft attitudes, have been executed using a simulation code of star images with Gaussian random centroiding errors. The resulting histogram of the number of observed star occurrences, is shown in Fig. 10, while Fig. 11 provides the overall time required to perform the Non-Dimensional Star-ID (in seconds). The performed 1000 tests provide a 0% empirical frequency of Star-ID failure, while the actual probability is approximately over 99%. The only failure occurred with a too few valid imaged stars scenario to form a triangle, or when the observed stars are thereabout aligned (this case may occurs with probability less than 0.1%). These results highly depend on integration time and on the catalog construction. An alternative approach can be easily developed to take into account the small number of these special cases. For attitude estimation using rate gyro data, an occasional drop out due to sparse star fields, is well tolerated. The execution times given in Fig. 11 are obtained using a 600Mhz Pentium III PC. These times, obviously, will be reduced with a better code implementation, but nonetheless, they provide an optimistic basis. k-vector based Star-ID Algorithms, such as the most robust Pyramid Algorithm (see Ref. [3], [6]), are dramatically faster. However, they require a well calibrated camera, that is, an accurate knowledge of the focal length and of the optical axis offsets. It is clear that once the first Star-ID has been performed, the camera can be then calibrated using the method presented in Ref. [1], and, consequently, the Pyramid 6

7 algorithm can be adopted to proceed with a robust and fast Star-ID process. 5. ANALYTICAL ESTIMATION OF THE FOCAL LENGTH Once the observed stars have been identified, it is possible to estimate the focal length by the following procedure. As already mentioned, this is an important task, needed to initiate faster Star-ID algorithms as, for instance, the Pyramid. The procedure to estimate f, is summarized as follows. (1) From Eq. (1), if we assume (x 0, y 0 ) are negligible with respect to f, then we can write the interstar angle of the measured stars as cos θ ij = b T i b j = x i x j + y i y j + f 2 x 2 i + y 2 i + f 2 x 2 j + y 2 j + f 2 (12) Note the difference between this equation, which calculates the interstar angle θ ij, and Eq. (4) that calculates the focal plane angle. (2) The measured cosines should ideally equal the cataloged cosines cos θ ij which are, in turn, well known from dot product of cataloged vectors. Thus, the following equation contains just one unknown: the focal length f. cos θ 2 ij [ (x 2 i + y 2 i + f 2 )(x 2 j + y 2 j + f 2 ) ] = (x i x j + y i y j + f 2 ) 2 (13) (3) By re-arranging Eq. (13) to the quadratic polynomial, then the roots of this polynomial represent the possible solutions for f f 2 = (a + d)d2 2c ± [ (a + b)d 2 2c ] 2 4(1 d 2 )(c 2 abd 2 ) 2(1 d 2 ) where a = x 2 i + y 2 i, b = x 2 j + y 2 j, c = x i x j + y i y j, and d = cos θ ij. (4) Let us check the sign of (c 2 abd 2 ). For small FOV (d 2 = 1) we have c 2 abd 2 = (x i x j + y i y j ) 2 (x 2 i + y 2 i )(x 2 j + y 2 j ) = = [x 2 i y 2 j + y 2 i x 2 j 2x i x j y i y j ] = [x i y j x j y i ] 2 0 (14) (5) Since the sign of [4(1 d 2 )(c 2 abd 2 )] is negative, then the value of the square root will be greater than [(a + d)d 2 2c]. By choosing the positive sign of the square root (because f 2 must be positive), then we obtain an algebraic solution for f. With this solution, the focal length with any star pair, can estimated, and an improved average value of it is obtained if more known stars are available. 7

8 The above analytical procedure has been tested using night-sky images, and simulated images. Both provide accurate results for the focal lengths. This procedure assumed that the optical axis offsets (x 0, y 0 ) are negligible. However, this analytical solution has been validated with large specified optical axis offsets, such as x 0 = 0.2 mm, and y 0 = 0.25 mm because we assume a small FOV in which the focal length is dominate the offsets. Table 4 shows the star coordinates (in mm) and the Star-ID results obtained with a random attitude and a known camera focal length of f = 55 mm. The focal length has been evaluated using Eq. (14) for each star pair. Figure 12 plots the results obtained for f, for each star pair. The mean value of the focal length is found to be mm and the standard deviation is This results validate the assumption of a focal length which is relatively insensitive to (x 0, y 0 ) errors. The approximate estimation of f obtained using Eq. (14), along with starting estimates of (x 0 0, y 0 0) are, therefore, good enough to ensure convergence of the nonlinear least squares algorithm (given in Ref. [1]), as a tool to obtain the final calibration estimates (x 0, y 0, f). 6. CONCLUSION This paper presents a novel Star-ID method for an un-calibrated star camera, called Non-Dimensional Star-ID. The proposed method is derived from the fact that the angles of focal plane triangles (formed by the star locations), are weakly dependent on the camera focal length, and on the optical axis offsets. The proposed method is particularly suitable for a poorly calibrated camera, and it can be well adopted as a back-up approach to solve the lost-in-space case of the Star-ID process. Finally, this approach is particularly suitable to make the ground calibration for night-sky experiments easier, as well as the in-flight calibrations. References [1] Samaan, M.A., Griffith, T., Singla, P., Junkins, J.L., Autonomous on-orbit Calibration Of Star Trackers, 2001 Core Technologies for Space Systems Conference, Colorado Springs, CO, Nov [2] Ketchum, E. A. and Tolson, R. H., Onboard Star Identification Without a Priori Attitude Information, Journal of Guidance, Control and Dynamics, Vol. 18, No. 2, March-April 1995, pp [3] Mortari, D., Junkins, J.L., and Samaan, M.A. Lost-In-Space Pyramid Algorithm for Robust Star Pattern Recognition, Paper AAS Guidance and Control Conference, Breckenridge, Colorado, 31 Jan. - 4 Feb

9 [4] Mortari, D., Pollock, T.C., and Junkins, J.L. Towards the Most Accurate Attitude Determination System Using Star Trackers, Advances in the Astronautical Sciences, Vol. 99, Pt. II, pp [5] Shalom, E., Alexander, J.W., and Stanton, R.H., Acquistion and Tracking Algorithms for the ASTROS Star Tracker, Paper AAS , The Annual Rocky Mountain Guidance and Control Conference, Keystone, CO, [6] Samaan, M.A., Mortari, D., and Junkins, J.L., Recursive Mode Star Identification Algorithms, Paper AAS Space Flight Mechanics Meeting, Santa Barbara, CA, February, [7] Mortari, D. Search-Less Algorithm for Star Pattern Recognition, Journal of the Astronautical Sciences, Vol. 45, No. 2, April-June 1997, pp

10 LIST OF TABLES Table 1: A Portion Of The Star Triangle Data Table 2: Star-ID Results (degree) Table 3: Star-ID Results for the Star1000 Picture Table 4: Star Data For Random Attitude 10

11 Table 1: A Portion Of The Star Triangle Data k I 1 I 2 I 3 α 1 α Table 2: Star-ID Results Star-ID R.A. (deg) Dec. (deg) Table 3: Star-ID Results for the Star1000 Image Star-ID R.A. (deg) Dec. (deg) Table 4: Star Data For Random Attitude X Y ID

12 LIST OF FIGURES Figure 1: Star Image For Two Different Focal Lengths Figure 2: Co-Linearity Condition Figure 3: The Smallest Focal Plane Angle Versus Its Index Figure 4: α 3 Versus Its Index Figure 5: Night-Sky Image Using Pegasus Camera Figure 6: Centroiding Results for the Pegasus Picture Figure 7: Estimated Focal Length And Optical Axis Offsets Figure 8: Night-Sky Image Using the Star1000 Camera Figure 9: Centroiding Results for the Star1000 Picture Figure 10: Histogram of the Observed Number of Stars Figure 11: Test Execution Times (in seconds) Figure 12: Focal Length Estimation Results (For Each Star Pair) 12

13 4 The effect of the camera parameters Y (mm) f=55 mm f=80 mm X (mm) Figure 1: Star Image For Two Different Focal Lengths 13

14 α r k r i r j f Φ i x (x, y ) j (x, y ) j 0 0 m i β (x i, y i ) y (x k, y k ) Figure 2: Geometry for a Pin-Hole Camera (Co-Linearity Condition) 14

15 α 3 (deg) index (k) x 10 5 Figure 3: The Smallest Focal Plane Angle Versus Its Index α 3 (deg) δα k 1 k index (k) Figure 4: α 3 Versus Its Index 15

16 Figure 5: Night-Sky Image Using Pegasus Camera y (pixels) x (pixels) Figure 6: Centroiding Results for the Pegasus Picture 16

17 0 The Calibration errors for the boresight position and the focal length x o y o Focal Length iteration Figure 7: Estimated Focal Length And Optical Axis Offsets) Figure 8: Night-Sky Image Using the Star1000 Camera 17

18 Centroided stars Y (pixles) X (pixels) Figure 9: Centroiding Results for the Star1000 Picture Occurrence No. of stars Figure 10: Histogram of the Observed Number of Stars 18

19 Time (sec) Test number Figure 11: Test Execution Times (in seconds) focal length (mm) Star pair # Figure 12: Focal Length Estimation Results (For Each Star Pair) 19