Radar Imaging with Compressed Sensing Yang Lu April 2014 Imperial College London Outline • Introduction to Synthetic Aperture Radar (SAR) • Background of Compressed Sensing • Reconstruct Radar Image by CS methods Introduction to SAR Important elements of SAR 1. Range Resolution and Azimuth Resolution 2. Chirp signal and Matched Filter Range Resolution and Azimuth Resolution of SAR http://www.radartutorial.eu/ Range Resolution 1 • Pulse signal (constant frequency signal) Range Resolution 2 • The resolution related with pulse width ๐๐ = ๐๐ ×๐ 2 ๐๐ : pulse width c : speed of pulse 2 : that is a round trip ( slant range resolution) Range Resolution 3 • If the incident angle is ๐๐ • Then the ground range resolution will be ๐๐บ๐ = ๐๐ ๐๐ ๐๐ ๐๐บ๐ ๐๐ ๐ ๐๐๐๐ = ๐๐ ×๐ 2๐ ๐๐๐๐ Azimuth Resolution 1 • Assume two points with same range Can’t distinguish A from B if they are in the radar ๐๐ beam at the same time Azimuth Resolution 2 • The azimuth resolution defined by ๐๐ : ๐ ×๐ ๐๐ด = ๐ × ๐๐ = ๐ฟ ๐ ๐๐ = ๐ฟ R: the slant range ๐: the wavelength L: the length of antenna LFM Signal • Linear Frequency Modulated Signal (Chirp Signal) ๐ก ๐ ๐ก = ๐๐๐๐ก ๐๐ฅ๐ ๐๐๐พ๐ก 2 ๐ Where ๐๐๐๐ก ๐ฅ = 1 ๐๐๐ ๐ฅ ≤ 1 2 0 ๐๐กโ๐๐๐ค๐๐ ๐ t: a time variable (fast time) T: duration of the signal K:is the chirp rate So the bandwidth of the signal is: ๐ต๐ = ๐พ ๐ Matched Filter • The output of the a Matched Filter is: ๐ ๐๐ข๐ก ๐ก = ๐ ๐ ๐ก โจโ ๐ก +∞ = −∞ ๐ ๐ ๐ข โ ๐ก − ๐ข du ๐ ๐ ๐ก : the received signal and โจ means convolution โ ๐ก = ๐∗ (−๐ก) ๐(๐ก) : the duplicated signal of the original signal and ∗ means complex conjugate Matched Filter (example) • If ๐ ๐ก = ๐๐๐๐ก ๐ก ๐ ๐๐ฅ๐ ๐๐๐พ๐ก 2 After ๐ก0 delay, we receive the signal ๐ก − ๐ก0 ๐ ๐ ๐ก = ๐๐๐๐ก ๐๐ฅ๐ ๐๐๐พ(๐ก − ๐ก0 )2 ๐ The reference signal will be −๐ก ∗ โ ๐ก = ๐ −๐ก = ๐๐๐๐ก ๐๐ฅ๐ −๐๐๐พ(−๐ก)2 ๐ ๐ก = ๐๐๐๐ก ๐๐ฅ๐ −๐๐๐พ๐ก 2 ๐ Matched Filter (example) • Output signal of the matched filter ๐ ๐๐ข๐ก ๐ก ≈ ๐๐ ๐๐๐(๐พ๐(๐ก − ๐ก0 )) 1 −๐ก ๐พ๐ 0 1 +๐ก ๐พ๐ 0 ๐ก0 So 3dB width of the main lobe= 0.886 ๐พ๐ ≈ 1 ๐พ๐ Range resolution improved • The range resolution improved Now we can distinguish B from C Range resolution improved Original ground range resolution: ๐๐บ๐ = ๐๐ ๐ ๐๐๐๐ = ๐๐ ×๐ 2๐ ๐๐๐๐ Now replace ๐๐ with 3dB main lobe width= 1 ๐พ๐ Finally, the improved ground range resolution will be : ๐๐ ๐ ๐ ๐๐บ๐ = = = ๐ ๐๐๐๐ 2๐ต๐๐ ๐๐๐๐ 2 ๐พ ๐๐ ๐๐๐๐ Phase difference 2๐ × 2๐ ๐= ๐ ๐: phase difference between the transmitted and the received signal 2๐: the distance (round trip) ๐: the wavelength of the transmitted signal http://www.radartutorial.eu/ SAR Azimuth Resolution • The phase change of the radar signal will be 4๐๐ (๐) ๐ ๐ = ๐ By Pythagorean theorem ๐ ๐ = ๐ 02 + (๐ฃ๐)2 ๐ฃ 2 ๐2 ≈ ๐ 0 + 2๐ 0 ๐: a time variable (slow time) ๐ฃ:the speed of plane Synthetic Aperture Radar Polarimetry (J.V. Zyl and Y. Kim) SAR Azimuth Resolution Substitute ๐ ๐ ≈ ๐ 0 + ๐ฃ 2 ๐2 2๐ 0 4๐๐ (๐) 4๐๐ 0 2๐๐ฃ 2 2 ๐ ๐ = ≈ + ๐ ๐ ๐ ๐ 0 ๐ The instantaneous frequency change of this 1 ๐๐ ๐ 2๐ฃ 2 signal is ๐ ๐ = = ๐ 2๐ ๐๐ ๐ 0 ๐ Which also can be considered as LFM signal And the total time ๐๐ก๐๐ก ≈ ๐๐ ๐ 0 ๐ฃ = ๐๐ 0 ๐ฟ๐ฃ SAR Azimuth Resolution The ๐ต๐๐ = ๐ ๐๐ก๐๐ก๐๐ = 2๐ฃ ๐ฟ The time resolution will be 1 ๐ต๐๐ = ๐ฟ 2๐ฃ So the azimuth resolution (in distance) will be 1 ๐ฟ ๐๐ด = ๐ฃ × = ๐ต๐๐ 2 2D signal of the target • One target have two equations-one is in the range direction (variable: fast time t) and another is in the azimuth direction (variable :slow time ๐) • If consider the signal on the two direction simultaneously, that will be a 2-dimensional signal with variable t and ๐. 2D signal of the target • ๐ ๐ก, η = ๐ด0 ๐๐ (๐ก −๐4๐๐0 2๐ (๐) − ) ๐ ๐ (๐) ๐ ๐ ×๐ ๐ด0 : a complex constant ๐ก ๐๐ t = rect( ) ๐ ๐(๐) 2 ๐๐ (๐) ≈ ๐ ๐๐๐ ๐๐ ๐๐ ๐ × ๐๐๐พ(๐ก− 2๐ (๐) 2 ) ๐ ๐๐ ๐(๐) ๐0 : the centre frequency (carrier frequency) Signal Energy 2D signal space • The received signals are stored in the signal space Digital Processing of Synthetic Aperture Radar Data: Algorithms and Implementation ( G.Cumming and H.Wong) SAR impulse response • If we ignore the constant ๐ด0 of ๐ ๐ก, η , we get the impulse response of SAR sensor: • โ๐๐๐ ๐ก, η = ๐๐ (๐ก − −๐4๐๐0 ๐ (๐) ๐ 2๐ (๐) ) ๐ ๐๐ ๐ × ๐๐๐พ(๐ก− 2๐ (๐) 2 ) ๐ ๐ ×๐ The received signal of the ground model can be built as the convolution of the ground reflectivity with the SAR impulse response (with additive white noise): ๐ ๐ก, ๐ = ๐ ๐ก, ๐ ⊗ โ๐๐๐ ๐ก, ๐ + ๐(๐ก, ๐) Radar Image • ๐ ๐ก, ๐ = ๐ ๐ก, ๐ ⊗ โ๐๐๐ ๐ก, ๐ + ๐(๐ก, ๐) Radar algorithms are trying to obtain the ground reflectivity function ๐ ๐ก, ๐ based on the received radar signal. • Traditional methods Range-Doppler Algorithm Chirp scaling algorithm Omega-K algorithms Background of Compressed Sensing Assume an N-dimensional signal ๐ฅ has a Ksparse representation (๐) in the basis Ψ ๐ฅ = Ψ๐ If we have a measurement matrix Φ (๐ × ๐) to measure and encode the linear projection of the signal we get measurements ๐ฆ = Φ๐ฅ = ΦΨ๐ If ๐ < ๐, there will be enormous possible solutions. And we want the sparsest one. Compressed Sensing • CS theory tells us that when the matrix A=ΦΨ has the Restricted Isometry Property (RIP), then it is indeed possible to recover the K-sparse signal from a set of measurement ๐ = ๐(๐พ๐๐๐(๐/๐พ)) But RIP condition is hard to check. An alternative way is to measure the mutual coherence ๐๐ , ๐๐ ๐ = ๐ ๐ด = max ๐≠๐ ๐๐ 2 ๐๐ 2 ๐๐ denotes the ๐ ๐กโ column of matrix A Compressive Radar Imaging (R. Baraniuk and P.Steeghs) Compressed Sensing • We want ๐ to be small (incoherence) CS theory has shown that many random measurement matrices are universal in the sense that they are incoherent with any fixed basis Ψ with high probability Compressive Radar Imaging (R. Baraniuk and P.Steeghs) Reconstruct Radar Image by CS methods • When RIP/incoherency holds, the signal ๐ฅ can be recovered exactly from ๐ฆ by solving an ๐1 minimization problem as: ๐ = ๐๐๐๐๐๐ ๐ 1 ๐ . ๐ก. ๐ฆ = ΦΨ๐ Reconstruct Radar Image by CS methods • If the measurement matrix Φ is a causal, quasi-Toeplitz matrix , the results also show good performance. (Right shift distance= ๐ ๐ ) Causal, quasi-Toeplitz Matrix (Example) ๐ ๐ If M=4, N=8 then right shift distance D= =2 ๐1 0 0 0 0 0 0 0 ๐3 ๐2 ๐1 0 0 0 0 0 ๐5 ๐4 ๐3 ๐2 ๐1 0 0 0 ๐7 ๐6 ๐5 ๐4 ๐3 ๐2 ๐1 0 ๐๐ is the ๐ ๐กโ element of a pseudo-random sequence ๐ Causal, quasi-Toeplitz Matrix The measurements of the signal will be ๐ฆ = Φ๐ฅ ๐ ๐ฆ ๐ = ๐ ๐ท๐ − ๐ ๐ฅ(๐) ๐=1 CS-based Radar We already know ๐ ๐ก, ๐ = ๐ ๐ก, ๐ ⊗ โ๐๐๐ ๐ก, ๐ + ๐(๐ก, ๐) For simplicity, just consider 1D range imaging model and ignore the noise ๐ ๐ก = ๐ ๐ก ⊗ โ๐๐๐ ๐ก Under this condition, โ๐๐๐ ๐ก can be considered as the transmitted radar pulse ๐ ๐ก =๐ด ๐= 2๐ ๐ ๐(๐)๐ ๐ (๐ก − ๐)๐๐ is the time delay. A is the attenuation. CS-based Radar Assume, the target reflectivity function ๐(๐ก) is ksparse in some basis. The PN or Chirp signals transmitted as radar waveforms ๐ ๐ (t) form a dictionary that is incoherent with the time, frequency and timefrequency bases. CS-based Radar • Let the transmitted radar signal be the PN signal ๐ ๐ • The target reflectivity generated from N Nyquist-rate samples ๐ฅ(๐) n=1,…,N via ๐ ๐ก = ๐ฅ ๐ก โ where 0 ≤ ๐ก ≤ ๐โ • The PN signal generated from a N-length Bernoulli โ1 vector ๐ ๐ via ๐ ๐ ๐ก = ๐ ๐ก โ CS-based Radar The received signal will be ๐ ๐ก = ๐ด ๐(๐)๐ ๐ (๐ก − ๐)๐๐ And we sample it every ๐ทโ second ๐ฆ ๐ = ๐ ๐ก |๐ก=๐×๐ทโ ๐โ =๐ด =๐ด ๐ ๐ ๐ ๐ (๐๐ทโ − ๐)๐๐ 0 ๐โ ๐ 0 ๐ =๐ด ๐ โ ๐โ ๐ ๐ (๐๐ท − )๐๐ ๐(๐๐ท − ๐) ๐=1 ๐ =๐ด (๐−1)โ ๐ ๐๐ท − ๐ ๐ฅ(๐) ๐=1 ๐(๐)๐๐ CS-based Radar (Results) The target reflectivity function can be recovered by using an OMP greedy algorithm ๐๐ ๐ ๐๐๐๐๐ y(๐) Compressive Radar Imaging (R. Baraniuk and P.Steeghs) Another Example (2-dimensional) The 2D received signal of a point target ๐ ๐ก, η = ๐ด0 ๐๐ (๐ก 2๐ (๐) − ) ๐ −๐4๐๐0 ๐ (๐) ๐ ๐๐ ๐ × ๐๐๐พ(๐ก− 2๐ (๐) 2 ) ๐ ๐ ×๐ If ignore the antenna pattern ๐๐ ๐ =1, ๐ด0 ๐๐ (๐ก − 2๐ (๐) ) ๐ be a constant ๐(๐๐ ) which is the radar cross section of point target ๐๐ Another Example (2-dimensional) The approximate received signal will be −๐4๐๐0 ๐ (๐,๐๐ ) ๐ 2๐ (๐,๐๐ ) 2 ๐๐๐พ(๐ก− ) ๐ ๐ ๐ ๐ก, η, ๐๐ = ๐(๐๐ )๐ × For a measurement scene Ω (๐ = ๐๐ × ๐๐ ) , the recorded echo signal will be ๐ ๐๐ ๐ก, η = ๐ ๐ก, η, ๐ ๐=1 ๐๐ : samples on the azimuth direction (slow time samples) ๐๐ : samples on the range direction (fast time samples) i: the ๐๐กโ point target in the scene Discrete format of the scene ๐ ๐๐ ๐ก, η = ๐๐ ๐ ๐=1 ๐ = ๐=1 −๐4๐๐0 ๐ (๐,๐) ๐ × 2๐ (๐,๐) ๐๐๐พ(๐ก− ๐ )2 ๐ 2๐ (๐,๐) ๐ (๐,๐) ๐๐๐พ(๐ก− ๐ )2 −๐4๐๐0 ๐ ๐๐ ๐ ๐ ๐๐ ๐ −๐∅๐ (๐ก,๐) = ๐=1 where ๐ −๐∅๐(๐ก,๐) =๐ ๐๐๐พ(๐ก− 2๐ (๐,๐) 2 ๐ (๐,๐) ) −๐4๐๐0 ๐ ๐ Discrete format of the scene ๐ ๐๐ ๐ −๐∅๐ (๐ก,๐) = ๐ด(๐ก,๐) ๐ ๐ ๐๐ ๐ก, η = ๐=1 ๐ด(๐ก,๐) ๐ −๐∅1 (๐ก,๐) −๐∅2 (๐ก,๐) ๐ = โฎ ๐ −๐∅๐ (๐ก,๐) ๐ด(1,1) ๐ ๐๐(1,1) ๐ด(2,1) ๐ โฎ ๐ด = ๐ด(๐๐,1) ๐ ๐๐(2,1) โฎ ๐ = ๐๐(๐๐,1) ๐ด(1,2) ๐ โฎ ๐ด(๐๐,๐๐) ๐ ๐๐(1,2) โฎ ๐๐(๐๐,๐๐) Discrete format of the scene ๐ = ๐ด๐ + ๐ ๐ :is additive white noise ๐ด: complete measurement matrix of SAR echo signal According to CS theory, we only need a small set of ๐ to successfully recover the sparse signal ๐ with high probability. Randomly select ๐ = ๐(๐พ๐๐๐(๐/๐พ)) rows of matrix A by using random selection matrix Φ Discrete format of the scene We assume that ๐ have a sparse representation๐ผ in a certain basisΨ (for example, a set of K point targets corresponds to a sparse sum of delta functions as in ๐๐ = ๐พ ๐=1 ๐ผ๐ ๐ฟ(๐ − ๐)), then we have ๐๐ = ๐ท๐ด๐ + ๐ = ๐ท๐ด๐น๐ผ + ๐ = ๐ฉ๐ผ + ๐ Where Θ= ๐ท๐ด๐น ๐๐ , Θ and ๐ผ are complex ๐๐ = ๐ฉ๐ผ + ๐ ๐ ๐ ๐๐ + ๐๐ผ๐ ๐๐ = ๐ ๐ ๐ฉ + ๐๐ผ๐ ๐ฉ ๐ ๐ ๐ผ + ๐๐ผ๐ ๐ผ = Re ๐ฉ Re ๐ผ − Im ๐ฉ Im(๐ผ) +๐ ๐ ๐ ๐ฉ ๐ผ๐ ๐ผ + ๐ผ๐ ๐ฉ ๐ ๐ ๐ผ So we have ๐ ๐ ๐๐ = ๐ ๐ ๐ฉ ๐ ๐ ๐ผ − ๐ผ๐ ๐ฉ ๐ผ๐(๐ผ) ๐ผ๐ ๐๐ = ๐ ๐ ๐ฉ ๐ผ๐ ๐ผ + ๐ผ๐ ๐ฉ ๐ ๐(๐ผ) We define signal๐๐ , ๐ฉ and ๐ผ as ๐ ๐ ๐๐ ๐ ๐ ๐ฉ ๐๐ = ,๐ฉ = ๐ผ๐ ๐ฉ ๐ผ๐(๐๐ ) −๐ผ๐ ๐ฉ ๐ ๐ ๐ฉ ,๐ผ = ๐ ๐ ๐ผ ๐ผ๐ ๐ผ Final Format ๐๐ = ๐ฉ๐ผ + ๐ Sparest solution can be solved by ๐1 norm minimization min ๐ ๐ผ 1 ๐ . ๐ก. ๐ − ๐ฉ๐ผ 2 < ๐ Simulation Results D ERS Ship Image Results SNR=20dB Noise free RD algorithm d CS algorithm CS algorithm SNR=10dB Reference • • • • • • R. Baraniuk and P. Steeghs. Compressive Radar Imaging. IEEE Radar Conference, April 2007. S.J. Wei, X.L. Zhang, J.Shi and G.Xiang. Sparse Reconstruction For SAR Imaging Based On Compressed Sensing. Progress In Electromagnetics Research, P63-81, 2010. J.V. Zyl and Y. Kim. Synthetic Aperture Radar Polarimetry. Dec 2010. G. Cumming and H. Wong. Digital Processing of Synthetic Aperture Radar Data: Algorithms and Implementation. Dec 2007. Radar Basics available at http://www.radartutorial.eu/index.en.html#this Y.K. Chan and V.C. Koo. An Introduction to Synthetic Aperture Radar (SAR). Progress In Electromagnetics Research. P27-60, 2008.