Compressed Sensing in MIMO Radar Chun-Yang Chen and P. P. Vaidyanathan California Institute of Technology Electrical Engineering/DSP Lab Asilomar 2008 Outline Review of the background – Compressed sensing [Donoho 06, Candes&Tao 06…] • Compressed sensing in radar [Herman & Strohmer 08] – MIMO radar [Bliss & Forsythe 03, Robey et al. 04, Fishler et al. 04….] Compressed sensing in MIMO radar – Compressed sensing receiver – Waveform optimization – Examples Conclusion Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 2 Review of the keywords: Compressed sensing, MIMO Radar 3 Brief Review of Compressed Sensing y Φ s dim( y ) dim( s ) Goal: Reconstruct s from y. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 4 Brief Review of Compressed Sensing y Φ s dim( y ) dim( s ) Goal: Reconstruct s from y. Incoherence: max i j φi,φ j is small. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 5 Brief Review of Compressed Sensing y Φ s Incoherence: max i j φi,φ j dim( y ) dim( s ) Goal: Reconstruct s from y. Sparsity: is small. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 i | s i 0 is small. 6 Brief Review of Compressed Sensing y Φ Incoherence: max i j φi,φ j is small. s dim( y ) dim( s ) Goal: Reconstruct s from y. Sparsity: i | s i 0 is small. Given y and F, s can be perfectly recovered by sparse approximation methods even when dim(y)<dim(s). Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 7 Brief Review of Compressed Sensing y Φ Incoherence: max i j φi,φ j is small. s dim( y ) dim( s ) Goal: Reconstruct s from y. Sparsity: i | s i 0 is small. Given y and F, s can be perfectly recovered by sparse approximation methods even when dim(y)<dim(s). This concept can be applied to sampling and compression. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 8 Review: Compressed Sensing in Radar [Herman & Strohmer08] Range u y Doppler targets Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 9 Review: Compressed Sensing in Radar [Herman & Strohmer08] Range u y y Φ Doppler targets * * s * * si: target RCS in the i-th Range-Doppler cell. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 10 Review: Compressed Sensing in Radar [Herman & Strohmer08] Range u Doppler y y Φ targets * * s * * si: target RCS in the i-th Range-Doppler cell. F is a function of the transmitted waveform u. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 11 Review: Compressed Sensing in Radar [Herman & Strohmer08] Range u Doppler y y Φ targets * * s * * F is a function of the transmitted waveform u. si: target RCS in the i-th Range-Doppler cell. Assumption: s is sparse. Transmitted waveform u can be chosen such that F is incoherent. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 12 Review: Compressed Sensing in Radar Target scene s can be reconstructed by compressed sensing method. High resolution can be achieved. [Herman & Strohmer08] y Φ * * s * * F is a function of the transmitted waveform u. si: target RCS in the i-th Range-Doppler cell. Assumption: s is sparse. Transmitted waveform u can be chosen such that F is incoherent. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 13 Brief Review of MIMO Radar MIMO Radar Each element can transmit an arbitrary waveform. u2(t) u1(t) u0(t) Phased array radar (Traditional) Each element transmits a scaled version of a single waveform. w2u(t) w1u(t) w0u(t) Advantages – Better spatial resolution [Bliss & Forsythe 03] – Flexible transmit beampattern design [Fuhrmann & San Antonio 04] – Improved parameter identifiability [Li et al. 07] Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 Compressed Sensing in MIMO Radar 15 MIMO Radar Signal Model (p,t, fD) t:delay fD :Doppler p: direction … u0(t) u1(t) uM-1(t) Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 16 MIMO Radar Signal Model (p,t, fD) (p,t, fD) t:delay fD :Doppler p: direction … u0(t) u1(t) … uM-1(t) y0(t) y1(t) Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 yN-1(t) 17 MIMO Radar Signal Model (p,t, fD) (p,t, fD) t:delay fD :Doppler p: direction … u0(t) u1(t) … uM-1(t) M 1 y n (t ) u m (t t ) e y0(t) y1(t) j 2 yN-1(t) T p (xm y n ) e j 2f D t m 0 Received signals Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 18 MIMO Radar Signal Model (p,t, fD) (p,t, fD) t:delay fD :Doppler p: direction … u0(t) u1(t) … uM-1(t) M 1 y n (t ) u m 0 m (t t ) e y0(t) y1(t) j 2 yN-1(t) T p (xm y n ) e j 2f D t Range Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 19 MIMO Radar Signal Model (p,t, fD) (p,t, fD) t:delay fD :Doppler p: direction … u0(t) u1(t) … uM-1(t) M 1 y n (t ) u m (t t ) e m 0 xm: location of the m-th transmitter yn: location of the n-th transmitter y0(t) y1(t) j 2 yN-1(t) T p (xm y n ) e j 2f D t Cross range Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 20 MIMO Radar Signal Model (p,t, fD) (p,t, fD) t:delay fD :Doppler p: direction … u0(t) u1(t) … uM-1(t) y0(t) y1(t) M 1 y n (t ) u m (t t ) e j m 0 xm: location of the m-th transmitter yn: location of the n-th transmitter 2 j e Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 2 yN-1(t) T p (xm y n ) e sin ( x m y n ) j 2f D t for linear array 21 MIMO Radar Signal Model (p,t, fD) (p,t, fD) t:delay fD :Doppler p: direction … u0(t) u1(t) … uM-1(t) M 1 y n (t ) u m (t t ) e y0(t) y1(t) j 2 m 0 xm: location of the m-th transmitter yn: location of the n-th transmitter Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 yN-1(t) T p (xm y n ) e j 2f D t Doppler 22 MIMO Radar Signal Model M 1 y n (t ) u m (t t ) e j 2 sin ( x m y n ) e j 2f D t m 0 Discrete Model: 0 L t M 1 yn IL m 0 0 ( L ' L t ) L 1 j 2 e D L Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 2 j ( xm yn ) NM um e j 2 D ( L 1 ) L e 23 MIMO Radar Signal Model M 1 y n (t ) m 0 u m (t t ) e j 2 sin ( x m y n ) e j 2f D t Range Discrete Model: 0 L t M 1 yn IL m 0 0 ( L ' L t ) L Range Cell: 1 j 2 e D L t 0 ,1, 2 L 1 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 2 j ( xm yn ) NM um e j 2 D ( L 1 ) L e L: Length of um 24 MIMO Radar Signal Model M 1 y n (t ) u m (t t ) e j 2 sin ( x m y n ) m 0 e j 2f D t Doppler Discrete Model: 0 L t M 1 yn IL m 0 0 ( L ' L t ) L 1 j 2 e D L Range Cell: t 0 ,1, 2 L 1 Doppler Cell: D 0 ,1, 2 L 1 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 2 j ( xm yn ) NM um e j 2 D ( L 1 ) L e L: Length of um 25 MIMO Radar Signal Model M 1 y n (t ) u m (t t ) e m 0 j 2 sin ( x m y n ) e j 2f D t Angle Discrete Model: 0 L t M 1 yn IL m 0 0 ( L ' L t ) L 1 j 2 e D L Range Cell: t 0 ,1, 2 L 1 Doppler Cell: D 0 ,1, 2 L 1 Angle Cell: 0 ,1, 2 NM 1 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 2 j ( xm yn ) NM um e j 2 D ( L 1 ) L e L: Length of um M: # of transmitting antennas N: # of receiving antennas 26 MIMO Radar Signal Model 0 L t M 1 yn IL m 0 0 ( L ' L t ) L 1 j 2 e D L 2 j ( xm yn ) NM um e j 2 D ( L 1 ) L H e nm ( H t ) HD Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 27 MIMO Radar Signal Model 0 L t M 1 yn IL m 0 0 ( L ' L t ) L 1 j 2 e D L ( H t Overall Input-output relation: 2 j ( xm yn ) NM um e j 2 D ( L 1 ) L H e nm ) HD y0 y 1 y H H H D u t y N 1 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 u0 u 1 u u N 1 28 MIMO Radar Signal Model Overall Input-output relation: y0 y 1 y H H H D u t y N 1 Hα Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 u0 u 1 u u N 1 α ( t , D , ) 29 MIMO Radar Signal Model Overall Input-output relation: y0 y 1 y H H H D u t y N 1 Hα t D u0 u 1 u u N 1 α ( t , D , ) Range Cell: t 0 ,1, 2 L 1 Doppler Cell: D 0 ,1, 2 L 1 Angle Cell: 0 ,1, 2 NM 1 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 30 Compressed Sensing MIMO Radar Receiver y H α u sα α α ( t , D , ) t D Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 31 Compressed Sensing MIMO Radar Receiver y H α u sα α α ( t , D , ) y Received waveforms t D Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 32 Compressed Sensing MIMO Radar Receiver y H α u sα α α ( t , D , ) y u Received waveforms Transmitted waveforms Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 t D 33 Compressed Sensing MIMO Radar Receiver y H α u sα α α ( t , D , ) t y Received waveforms u Transmitted waveforms H α Transfer function for the target in the cell Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 D 34 Compressed Sensing MIMO Radar Receiver y H α u sα α α ( t , D , ) t y Received waveforms u Transmitted waveforms H α Transfer function for the target in the cell s α RCS of the target in cell Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 D 35 Compressed Sensing MIMO Radar Receiver y H α u sα φα α ( t , D , ) α φ α sα Φ s α t y Received waveforms u Transmitted waveforms H α Transfer function for the target in the cell s α RCS of the target in cell Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 D 36 Compressed Sensing MIMO Radar Receiver y H α u sα φ φα α ( t , D , ) α α sα Φ s α t s is sparse if the target scene is sparse. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 D 37 Compressed Sensing MIMO Radar Receiver y H α u sα φ φα α ( t , D , ) α α sα Φ s α t s is sparse if the target scene is sparse. D Compressed sensing algorithm can effectively recover s if F is incoherent. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 38 Waveform Optimization y H α α u sα φα φ α sα Φ s α Goal: Design u such that max α α ' H α u , H α 'u t D is small. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 39 Waveform Optimization α TX α' α RX … u α' … s α H α u s α ' H α 'u Goal: Design u such that max α α ' H α u , H α 'u is small. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 40 Waveform Optimization α TX α' α RX … u α' … s α H α u s α ' H α 'u Goal: Design u such that max α α ' H α u , H α 'u Small Correlation is small. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 41 Waveform Optimization: Dimension Reduction H α u , H α 'u u ( H α ' H αt ' H α D ' ) ( H α H αt H α D ) u H H Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 42 Waveform Optimization: Dimension Reduction H α u , H α 'u u ( H α ' H αt ' H α D ' ) ( H α H αt H α D ) u H H H H H u ( H α ' H α H α D ' H αt ' H αt H α D ) u H Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 43 Waveform Optimization: Dimension Reduction H α u , H α 'u u ( H α ' H αt ' H α D ' ) ( H α H αt H α D ) u H H H H H u ( H α ' H α H α D ' H αt ' H αt H α D ) u H H u ( H α ' H α C αt αt ' H α D α D ' ) u H CK Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 k 1 1 1 44 Waveform Optimization: Dimension Reduction H α u , H α 'u u ( H α ' H αt ' H α D ' ) ( H α H αt H α D ) u H H H H H u ( H α ' H α H α D ' H αt ' H αt H α D ) u H H u ( H α ' H α C αt αt ' H α D α D ' ) u H k ( α , α ' , αt , α D ) CK Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 1 1 1 45 Waveform Optimization: Dimension Reduction H α u , H α 'u u ( H α ' H αt ' H α D ' ) ( H α H αt H α D ) u H H H H H u ( H α ' H α H α D ' H αt ' H αt H α D ) u H H u ( H α ' H α C αt αt ' H α D α D ' ) u H ( α , α ' , αt , α D ) k Goal: Design u such that max ( α , αt , α D ) ( α ', 0 , 0 ) ( α , α ' , αt , α D ) is small. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 CK 1 1 1 46 Waveform Optimization: Beamforming To concentrate the transmit energy on the angles of interest, we want the following term to be small α B ( α , α ,0 ,0 ) B: the set consisting of angles of interest. Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 47 Waveform Optimization: Beamforming To concentrate the transmit energy on the angles of interest, we want the following term to be small B: the set consisting of angles of interest. ( α , α ,0 ,0 ) α B To uniformly illuminate the angles of interest, we want the following term to be small 2 α B ( α , α ,0 ,0 ) 1 B ( α , α ,0 ,0 ) α B Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 48 Waveform Optimization: Cost function Incoherent max ( α , αt , α D ) ( α ', 0 , 0 ) Stopband ( α , α ' , αt , α D ) ( α , α ,0 ,0 ) α B 2 Passband ( α , α ,0 ,0 ) α B Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 1 B ( α , α ,0 ,0 ) α B 49 Waveform Optimization: Cost function max ( α , αt , α D ) ( α ', 0 , 0 ) ( α , α ' , αt , α D ) + ( α , α ,0 ,0 ) α B + 2 (1 ) ( α , α ,0 ,0 ) α B Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 1 B ( α , α ,0 ,0 ) α B 50 Waveform Optimization: Cost function Incoherent Stopband max ( α , α ' , αt , α D ) ( α , α ,0 ,0 ) (α(α , ',α0t, 0, ) α D ) α B min 2 u 1 ( α , α ,0 ,0 ) (1 ) ( α , α , 0 , 0 ) B α B α B Passband Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 51 Phase Hopping Waveform Consider constant-modulus signal: u m (l ) e Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 j 2 ml 52 Phase Hopping Waveform Consider constant-modulus signal: u m (l ) e j 2 ml Consider phase on a lattice: ml C ml K , C ml 0 ,1, 2 , K 1 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 53 Phase Hopping Waveform Consider constant-modulus signal: u m (l ) e j 2 ml Consider phase on a lattice: ml min C ml C ml K , C ml 0 ,1, 2 , K 1 2 max ( α , α ' , αt , α D ) ( α , α ,0 ,0 ) (α(α , ',α0t, 0, ) α D ) α B 2 1 2 2 ( α , α ,0 ,0 ) (1 ) ( α , α , 0 , 0 ) B α B α B Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 54 Simulated Annealing Algorithm min f ( C ) subject to C C Simulated annealing – Create a Markov chain on the set A with the equilibrium distribution T (C ) ZT f (C ) exp ZT T 1 C’ C … … f (C ) exp T C – Run the Markov chain Monte Carlo (MCMC) – Decrease the temperature T from time to time Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 55 Example: Histogram of correlations Alltop Sequence # of (,’) pairs 300 200 100 0 0 2 4 6 8 10 12 14 16 18 20 16 18 20 Parameters: Uniform linear array # of RX elements N=10 # of TX elements M =4 Signal length L=31 # of phase K=15 Angle of interest ALL HProposed u , H α 'u Method , ' α 300 200 100 0 0 2 4 6 8 10 12 14 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 56 Example: Histogram of correlations Alltop Sequence # of (,’) pairs 300 200 100 0 0 2 4 6 8 10 12 14 16 18 20 16 18 20 Parameters: Uniform linear array # of RX elements N=10 # of TX elements M =4 Signal length L=31 # of phase K=15 Angle of interest ALL Proposed Method 300 200 100 0 0 2 4 6 8 10 12 14 H α u , H α 'u , ' Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 57 Example: Histogram of correlations Alltop Sequence # of (,’) pairs 300 200 100 0 0 2 4 6 8 10 12 14 16 18 20 16 18 20 Parameters: Uniform linear array # of RX elements N=10 # of TX elements M =4 Signal length L=31 # of phase K=15 Angle of interest ALL Proposed Method 300 200 100 0 0 2 4 6 8 10 12 14 H α u , H α 'u , ' Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 58 Example: Recovering Target Scene 3 Range Target Scene 2 1 0 10 20 30 200 400 600 800 1000 1200 10 20 Cross Range 30 40 10 20 30 40 10 20 Cross Range 30 40 60 Matched Filter 40 10 20 20 0 30 200 400 600 800 1000 1200 30 20 Range Compressed Sensing 10 0 10 20 30 200 400 600 800 1000 1200 SNR=10dB Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 59 Example: Recovering Target Scene 3 Range Target Scene 2 1 0 10 20 30 200 400 600 800 1000 1200 10 20 Cross Range 30 40 10 20 30 40 10 20 Cross Range 30 40 60 Matched Filter 40 10 20 20 0 30 200 400 600 800 1000 1200 30 20 Range Compressed Sensing 10 0 10 20 30 200 400 600 800 1000 1200 SNR=10dB Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 60 Example: Recovering Target Scene 3 Range Target Scene 2 1 0 10 20 30 200 400 600 800 1000 1200 10 20 Cross Range 30 40 10 20 30 40 10 20 Cross Range 30 40 60 Matched Filter 40 10 20 20 0 30 200 400 600 800 1000 1200 30 20 Range Compressed Sensing 10 0 10 20 30 200 400 600 800 1000 1200 SNR=10dB Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 61 Example: Recovering Target Scene 3 Range Target Scene 2 1 0 10 20 30 200 400 600 800 1000 1200 10 20 Cross Range 30 40 10 20 30 40 10 20 Cross Range 30 40 60 Matched Filter 40 10 20 20 0 30 200 400 600 800 1000 1200 30 20 Range Compressed Sensing 10 0 10 20 30 200 400 600 800 1000 1200 SNR=10dB Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 62 Conclusion Compressed sensing based receiver – Applicable when the target scene is sparse – Better resolution than the matched filter receiver Waveform design – Incoherent – Beamforming – Simulated annealing Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 63 Thank You! Q&A Any questions? Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 64 Simulated Annealing Algorithm min f ( C ) subject to C C Simulated annealing – Create a Markov chain on the set A with the equilibrium distribution T (C ) ZT f (C ) exp ZT T 1 Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 … … f (C ) exp T C C’ C 65 Simulated Annealing Algorithm min f ( C ) subject to C C Simulated annealing – Create a Markov chain on the set A with the equilibrium distribution T (C ) ZT f (C ) exp ZT T 1 … … f (C ) exp T C C’ C – Run the Markov chain Monte Carlo (MCMC) Chun-Yang Chen, Caltech DSP Lab | Asilomar 2008 66