I. Detecti on -Trac ki n g Performance with Combined Waveforms CONSTANTINO RAG0 Scientific Systems PETER WILLETT YAAKOV BAR-SHALOM, Fellow, IEEE University of Connecticut It is commonly understood that in radar or active sonar detection systems constant-frequency (CF) pulscs correspond to good Doppler but poor delay resolution capability; and that linearly swept frequency (FM) pulses have the opposite behavior. Many systems are capable of both types of operation, and hence in this paper the fusion of such pulses is examined from a system point of view (i.e., tracking performance) via hybrid conditional averaging (HYCA), a new but increasingly accepted technique for evaluating tracking performance in the presence of missed detections. It is shown that tracking errors are highly dependent on the waveform used, and in many situations tracking performance using a good heterogeneous waveform is improved by an order of magnitude when compared with a scheme using a homogeneous pulse (CF or FM) with the same energy. Between the two types of FM considered (upsweep and downsweep), it is shown that the upsweep is always superior. Also investigated is INTRODUCTION Many active detection systems (radar and sonar) are capable of waveform agility-particularly of constant frequency (CF) and linearly swept frequency (FM, or chirp)-in their transmitted pulses. It appears that such systems presently use their abilities, but do not fuse the results in an optimal way. The purpose of this work is to examine the possible benefits from such fusion. In a recent paper [l] a similar problem was analyzed, but there the system assumed perfect detection and the estimation of the range and range-rate reach the corresponding CRLB (Cramer-Rao lower bound). The importance of waveform design from a tracking point of view is discussed in [2]. We derive the relationship between the pulse shape and the detection performance, and based on that we compute the corresponding measurement noise covariance. To be specific, we assume the following. 1) A Swerling I target and additive white Gaussian noise. A target made up of many reflectors simplifies the analysis, as will be seen. 2) A combined FM+-CF, FM--CF or FM+-FMc transmitted pulse with the two waveforms abutted. FM+stands for FM with a positive sweep rate (upsweep), FM-for FM with a negative sweep rate (downsweep). CF indicates a constant-frequency pulse or subpulse. We do not consider phase-coded or other broadband pulses. 3) A constant target RCS (radar cross section) for the CF and F M subpulses. In the case of our “abutted” pulses this makes sense, as one would expect little change in the observed target scattering characteristics for the duration of a pulse. 4) Known ambient noise power. 5 ) A single target, and no false alarms. an alternating-pulse system, which, while suboptimal, appears to The apportionment of energy between CF and FM or offer robust performance. FM+ and FM- subpulses is a parameter, and as such Manuscript received December 16, 1996; revised April 1 , 1997 IEEE Log No. T-AESI34/2/03204. This research was supported through Rome Laboratory under AFOSR Contracts F30602-94-C-0060 and F49620-95-1-0229, and by the Office of Naval Research Contracts N66604-92-C-1386 and N00014-91-J-l950. Authors’ addresses: C. Rago, Scientific Systems, 400 W. Cummings Park, Ste. 3950, Woburn, MA 01801; P. Willett and Y. Bar-Shalom, University of Connecticut, U-157, Storrs, CT 06269-2157. 0018-9251/98/$10.00 @ 1998 IEEE 612 we include full CF and full FM as special cases. It is well known [3, 41 that a CF pulse has good range-rate (i.e., Doppler frequency) resolution characteristics, but is relatively poor at range resolution; conversely, that an FM pulse can have excellent range, but comparatively poor range-rate resolution. The idea of this research is to investigate whether a combined system (using some apportionment between both kinds of pulses) can perform better than either extreme. Our goal, therefore, is to compare the performances, with respect to tracking error, of a number of schemes: 1) A pure CF waveform, of duration Tp. 2) A pure linear-FM waveform, of duration Tp; sweep rate (denoted by 6 ) is a parameter. As will be shown, the sign of the sweep rate has a significant influence on the tracking performance. IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 2 APRIL 1998 ~ 3) Combined CF-FM- waveform, with the first (1 - y>?:, seconds being of constant frequency, and the remaining yT, of linearly-negatively-swept frequency. As above, the sweep rate is a parameter. 4) A combined FM+-FM- waveform, with the first yT, seconds being of linearly-positively-swept frequency, and the remaining (1 - y)T, of linearlynegatively-swept frequency. As above, sweep rate is a parameter. 5 ) A combined FM+-CF waveform, with the first yT, seconds baing of linearly-positively-swept frequency, and the remaining (1 - y)T, of constant frequency. As above, the sweep rate is a parameter. 6) FM and CF waveforms, each of duration T,, alternating on successive scans. To be fair, the energy per pulse is independent of the pulse shape employed. Comparison is on the basis of tracking performance as specified by the algebraic Riccati equation (ARE) modified by the hybrid conditional iiveruging (HYCA) technique to account for missed detections [5]-the technique is analytical but approximate, and extensive justification of its use is given in [6]. A major conclusion of this study is that, from a system level (i.e.. tracking) point of view, FM' (upsweep) is alwaq s superior to FM- (downsweep). Overall, a 50%-50% division of energy between FM' and FM- seems the best choice. In Section I1 we give mathematical preliminaries and signal models; naturally, we discuss ambiguity functions and probability of detection. Section I11 details our analysis procedure, specifically relating to our assumptions about measurement accuracy and the use of the HYCA approach. Finally, Section IV gives our results and recommendations. II. BACKGROUND A. Receiver Model Receiver processing is assumed, without loss of generality, to be: performed at "baseband"; that is, subsequent to carrier-removal via mixing. If the transmitted baseband signal is s(t), the matched filter is the one with an impulse response h(t) = s*(-t), i.e., the impulse response of the matched filter is the conjugate time-reversed baseband signal. With ri(t) the received waveform, and r(t) the baseband representation of ri(t), we have, as an output process of our matched fili;er, Envelope L] Matched Filter Fig. 1. Complex envelope matched filter operation. where f, is the carrier frequency, and s(t) is a replica of the transmitted baseband signal. This standard scheme is pictured in Fig. 1. If the return signal is expected to be Doppler shifted (as in the radar case), then the matched filter to a return signal with an expected frequency shift fa has an impulse response h(t) = s*(-t)exp(j2..rrfot), and its output is given by The (baseband) received signal will be modeled as a return from a Swerling I target [3]: = As(t - -r)ej2"fd'Z(target) + v(t) (3) where s(t,-r,fd) = s(t - .)ejzafdt is a delayed ( 7 ) and Doppler-shifted (fd) replica of the emitted baseband complex envelope signal s(t); I(target) is a target indicator. The complex numbers Ai represent the amplitude and phase of each of the target reflectors; for a Swerling I model the assumption is that there are many such reflectors, each one with a random amplitude and phase. Hence A = CiAiapproaches a complex Gaussian random variable with zero mean and variance 2a;. We assume v(t) is complex white Gaussian noise independent of A, with zero mean and variance 2N,. The signal s ( t ) can be any of those described in the previous section. B. Ambiguity Function Of major interest in the sequel is the "ambiguity function" [7], given by (4) As we see next, the ambiguity function specifies the output of the matched filter in the absence of noise; its shape is related to the implied resolution cell (to be defined in Subsection IIIC), and is strongly dependent on the waveform. The relationship between the ambiguity function and resolution cell (which does not appear to be universally well-understood) is discussed in the Appendix. Briefly, the relationship is that the time-reversed ambiguity function defines a RAG0 ET AL.: DETECTION-TRACKING PERFORMANCE WITH COMBINED WAVEFORMS 613 ~ resolution cell "primitive", which, when enclosed as tightly as possible by some tesselating shape such as a parallelogram, defines the resolution cell itself. Examples showing the effect of different energy apportionments between CF and FM pulses are given in Figs. 5 and 8. C. This random variable is still zero mean, with variance given by x [As(a!- T)ej2"f,ln+ v(a!)]*}dXda! Detection Here we discuss the detection of a delayed and Doppler-shifted return. The implied test will use the output of the envelope (magnitude square of the real and imaginary parts according to Fig. 1) matched filter. At time t the magnitude square of the output of a filter matched to a zero delay and a zero Doppler shift is 2 lx(t)I2 = I L r ( X ) s * ( x - t)dAl . (5) Since the delay and Doppler shift are unknown, we , examine without loss of generality this at time T ~and hence the test is given by 2 K d hH l7 I x ( T ~ ) ~ ~ = ~ ~ T o r ( X ) s *-( X~ ~ ) 2 = LT0 v(X)s*(X - To)dX. - (6) with 7 a threshold. We characterize now the random ~ ) the noise-only hypothesis ( H ) and variable ~ ( 7under under the target-present hypothesis ( K ) . Noise-Only Hypothesis: Here we have, according to our formalism, T(target) = 0, and hence X(To) where we have used the definition of the ambiguity function A given by (4)(with the normalization factor to take into account the nonunity energy signals). Recall that the magnitude square of a complex Gaussian random variable x N(0,a?) is exponentially distributed, with density given by (7) We consequently have that the probability of false alarm (P,) is given by and the probability of detection (Pd) by The random variable X ( T ~ )is complex Gaussian, with zero mean and variance given by (13) using (12) we get = 2N0[. The last integral E is the energy of the transmitted pulse. Target Present: In this case we have I(target) = 1, and the matched filter output is thus This relationship has been plotted for a particular probability of false alarm, as a function of 7- and f d and for different FM+-CF apportionments in Fig. 5 and for FM+-FM- in Fig. 8. Observe the pronounced variation in the shape of the "resolution cells" as a function of the apportionment. Ill. COMPARISON OF SCHEMES (9) 614 In this Section we give details relating to our mode of comparison of our various pulse-fusion schemes. IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 2 APRIL 1998 delay and Doppler shift: Of major interest are: the signal model, the target model, the measurement model, and the analytical framework. We devote a subsection to each. A. y(k) = Hx(k) + ~ ( k ) where the measurement covariance E { w ( k ) ( ~ ( k ) >=~ } R is a function of the waveform shape employed. We write Signal Model In our studies we use a baseband signal s(t). This signal consists, in general, of two different subpulses, concatenated in such a way that the phase is continuous and the total signal duration is Tp. The extreme cases (Le., when a duration of one of the subpulses is zero) are also included. In the first set of signals, called CF-FM-, each s(t) comprises (1 - y)Tp seconds of CF (which at baseband has constant frequency) followed by yTp seconds of linearly-negatively-swept FM. That is, we have 0 1.t < (1 - y)Tp (1 - y)Tp 5 t < Tp * 0 else (15) Results using other configurations, such as FMfollowed by CF, exhibit different ambiguity functions; however, results are similar. The second set of signals, called FM+-FM-, is described by The last set of signals, called FM+-CF is specified by lo (18) else B. Target Model Our target model is kinematic with rangehange-rate as state variables. That is, we have x(X:+ 1) = Fx(k) + ~ ( k ) where the process noise is white and zero-mean with covariance E { v ( k ) v T ( k ) } = Q ( k ) . Our observations are F= [:4‘1 r-At3 At21 - where At represents the time between samples, co is the wave’s propagation speed in the medium, f , is the carrier frequency, and is the power spectral density of the continuous-time white process noise discretized in (18) [8, p. 2621. Note that for clarity we have chosen to work in one-dimension, range, since it is in this domain that the differences between the various waveform types become apparent. The rms target acceleration corresponding to uq is a,/& [8, p. 2631. cri C. Measurement Model In the most generous world it would be possible to determine an optimal (presumably maximum likelihood) estimate of delay T and shift f, for each detection. However, due to the complexity of such an operation most systems resort to a sampled or quantized approach, which immediately gives rise to the concept of a resolution ceZZ. For us, an ideal resolution cell (to be designated as a resolution cell primitive) is defined as the region in rangehange-rate space where each point, conditioned on the target present hypothesis, has a probability of detection Pd or greater. Mathematically, this resolution cell is defined by the test given in (6) and (13). If the output of the magnitude square matched filter exceeds the threshold 7 ,then the parameters ( 7 , f d ) of the return signal are inside the resolution cell defined by the level P,. If a detection is declared, then the corresponding measurement as supplied to the tracking algorithm is the centroid of that resolution cell (i.e., the matched filter parameters T ~fob>, , with measurement covariance R derivable from the size and shape of the cell, assuming a uniform distribution of the measurements in the cell. Ideally, therefore, resolution cells should be sufficiently regular in geometry that they are mutually exclusive and exhaustive of space (Le., they “tessellate”); and that a target located in a given cell will not effect a detection in any other. Unfortunately this is overly idealized even in the pure CF or FM cases, in which constant-P, contours RAG0 ET AL.: DETECTION-TRACKING PERFORMANCE WITH COMBINED WAVEFORMS 615 1,51 1 I I -1.5‘ -0.08 -0.06 -0.04 -0.02 0 0.02 0 04 0.06 I 0.08 Fig. 2. Resolution cell (shaded area) and corresponding parallelogram that contains it. tend to be elliptical in shape, and it is certainly not true when a combined waveform is used. To handle this problem, we define as a practical resolution cell the parallelogram that contains the resolution cell primitive (see Fig. 2). The average probability of detection is, therefore, the conditional expected value given by Naturally, the area inside the resolution cell but outside the resolution cell primitive contributes to the miss probability. 4) For such a resolution cell, define the measurement covariance by assuming the location of the detection-producing target to be uniformly distributed within the resolution cell primitive, that is, the shaded region in Fig. 2. Our resolution cell primitives do not tessellate; however, the parallelograms that contain each (the practical resolution cells) do. Our model is that at each time sample a “coin” is flipped: with probability (1 - pd) we have a miss, and there is no measurement. And with probability Pd a measurement is generated distributed uniformly within the resolution cell primitive. D. Analytical Approach The state estimation-error covariance matrix is updated by the Kalman filter as follows ([8]): P(k + 1 1 k ) = FP(k 1 k)F’ + Q S(k + 1) = HP(k + 1 I k)H’ + R(k) W(k + 1) = P(k + 1 I k)H’S(k + l)-’ (22) P(k + 1 I k + 1) = P(k + 1 I k) - W(k + 1) where f ( 7 , f d ) is the probability density function of ) the the target being at the coordinates ( ~ , f d and integration is denoted by as being done over the shaded area. Using the previous uniform distribution assumption, this density is equal to the inverse of the area of the parallelogram (Ap): - Our analysis would match reality most closely if we were to assign, at each time, a uniform distribution within each resolution cell to our target, and thence to establish a corresponding measurement error (the difference between the uniformly distributed variate and the resolution-cell center) and probability of detection (via (14) and the ambiguity function). However, in order to fit with our HYCA approach we must quantize Pd, and in the current analysis we quantize to two levels. Our steps are thus as follows. 1 ) Define a resolution cell primitive from the corresponding ambiguity function (see the Appendix) by selecting a “detection threshold” within which a prespecified probability of detection is consistently exceeded. 2) Form the corresponding (practical) resolution cell as the parallelogram which most snugly circumscribes the resolution cell primitive, 3) For such a resolution cell, define the probability of detection to be its conditional average value over the parallelogram that contains the cell (see (21)). 616 x S(k + l)W’(k + 1). With R(k) = R this evolves to a steady state which characterizes the performance of the filter, and this solution can be computed from the associated ARE. However, in the case that detections can be missed, the measurement noise has a stochastic covariance (R(k)), and there is no steady-state, but a stationary matrix-valued random process. Therefore, it is legitimate to seek the expected value of the error covariance. Let us assume that the measurement noise covariance matrix can be modeled as detection R(k) = where, as described in the previous subsection, R is derived from the size and shape of the resolution cell and the probabilities of detection and miss are the corresponding averaged values. The method here employed is the HYCA method of [SI. (For justification of the approach, and comparison of HYCA results to simulation, the reader is directed to [6].) Under this scheme we embed a binary random process { b ( k ) } into the Riccati equation such that b(k) = P { b ( k ) = i} = 1 detection 2 miss i=l (24) IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 2 APRIL 1998 Tx Pl(k targets), the other applicable to the radar situation (high frequency, short pulses, fast-moving targets). Both cases use a Swerling I target and additive white Gaussian noise. We expect that results using other models would give rise to similar conclusions. + l/k + I) 1 - I’d pl(klk) P2(k/L) P?(k+l/b+l) L 1 - P,i Fig. 3. Illustration of procedure for finding steady-state covariance for Kalman filter operating in stochastic environment Y i a HYCA procedure. Then, with Pij(k + 1 1 k + 1) defined as the estimation error covariance given b(k) = i and b(k + 1) = j (obtainable from (22) with R(k) according to (23)), we get L Pj(k + 1 I k +. 1) = x P i j ( k I k)P{b(k) = i}. i= 1 (25) The state probabilities at time k , pi(k) = P { b ( k ) = i } , i = 1,2 are derived from the state probabilities at time k - 1 and the Mark.ov transition matrix: The matrices P,(k 1 k ) reflect the state error covariance given that s(k) = i. The mean state covariance matrix (i.e., what we want) is given by 2 ~ ( I kk) = C P j ( k I k ) ~ j ( k ) . (27) .j=1 This can be evaluated when this matrix Markov process reaches steady state. Fig. 3 shows a flow-diagram that reflects the mechanics of the iterations. IV. RESULTS We have developed two sets of representative plots: one with parameters suitable for sonar (Le., low carrier frequency, long pulses, slow-moving A. Sonar Case For the (active) sonar case we assume a pulse length of Tp = 1 s, a scan-to-scan time of At = 30 s, and a carrier frequency of f , = 3.5 kHz. In each case here we have SNR = 20 dB and K, = 50 Hz/s or )i = -50 Hz; this corresponds to a maximum frequency sweep of 50 Hz, found in case y = 1. The first waveform analyzed was the combined CF-FM- (a CF subpulse followed by downsweep (see Fig. 4(a)). For various values of y we show in Fig. 5 representative constant-Pd contours for the CF-FMwaveform. As discussed earlier, the “resolution cell primitive” is determined by the volume enclosed by one of these contours, and for the CF and FM cases these are as expected: in the former we have good Doppler and comparatively poor range resolutions (Fig. 5(a)); and in the latter the reverse (Fig. 5(c)). Intermediate values of y give more complicated Pd contours. Here the resolution cell primitive is determined by the Pd = 0.80 contour. In Fig. 6 we show the corresponding measurement errors; these are the diagonal elements of the measurement covariance matrix R (corresponding to the delay variance and the Doppler-shift variance) and the correlation coefficient between delay and Doppler-shift measurements. (Note that a negative correlation between delay and Doppler-shift corresponds to a positive correlation between range and range-rate measurements. Negative measurement correlation is usually more favorable for tracking .) Fig. 6 deals with measurement covariances. As discussed earlier, this corresponds to the covariance of a uniformly distributed target within the resolution cell primitive, for example the shaded region of Fig. 2. It is worth noting here that an earlier contention, that FM provides good delay information, does not seem at first to be borne out by this figure. The 2- -12 3 0 0 1 0 2 0 3 0 4 0 5 06 0 7 08 09 i time time time (a) (b) (C) Fig. 4. Frequency variation for waveforms used in analyses. (a) CF-FM-. (b) FM+-FM-. (c) FM+-CF. RAG0 ET AL.: DETECTION-TRACKING PERFORMANCE WITH COMBINED WAVEFORMS 617 0 -0.5 -- -0.5 : -0.5 0 0 : o , o 0.02 -1 0.5 -0.5 0.5 -0.2 0 I 0,061 1 10 0 20 30 40 50 60 70 80 90 100 0.2 0 (C) I 10‘ I 1 0 lo-10 -1 -0.2 0 :Kl‘:\I 0 (e) 0.2 -10 -0.4 -0.2 0 (f) 20 30 40 50 60 70 80 90 100 (b) 0.2 0 (d) -2 -0.2 10 0.2 20 -0 5 0 -1 -20 -0.5 1 1 10 10 20 30 40 50 60 70 BO 90 100 (4 0 FM- pulse 0.5 (9) Fig. 5. Resolution cells (probability of detection contours) and parallelograms that contain them in delaylDoppler space for CF-FM- waveforms, ’P = 0.001 and SNR = 20 dB, frequency sweep factor IC. = 50 H d s and probability of detection threshold of 0.80. (a) y = 0% (CF). (b) y = 20%. (c) y = 40%. (d) y = 50%. (e) y = 60%. (f) y = 80%. (g) y = 100% (FM-). Horizontal and vertical axes are in seconds and Hertz, respectively, a 1 s pulse with 50 H d s is assumed. reason is that in Fig. 6(a) we are plotting the delay measurement error alone, and with reference to Fig. 5 this is as large as for the pure-CF case. What is not plotted in Fig. 6(a) (but is available in Fig. 6(c)) is that delay and Doppler measurements are strongly correlated, meaning that if something is known about the range-rate (for example, that it is small) then delay can be inferred quite accurately. In Fig. 7 we show steady-state tracking errors for the CF-FiW waveform obtained via our HYCA procedure, for various values of the target motion uncertainty (quantified by the process noise power spectral density a t ) . Note the difference between Figs. 6 and 7: the former deals with measurements as supplied to the tracker, and the latter to the posterior tracking errors using these measurements. There are a number of items of note as follows. duration ( 7 . in X of the total pulse duration) Fig. 6. Measurement error variances for CF-FM- waveform as function of y;same parameters as Fig. 5. (a) Delay variance. (b) Doppler-shift variance. (c) Correlation coefficient between delay and Doppler shift. 1 0’ 0 10 20 30 40 50 80 70 80 90 100 (4 I=--+ [(,,se,)’] lo-‘ ‘no 10-2----=,---== zII=:::::____-_-__-: _..-_____--- _ _ _ - 1) Comparing Figs. 6 and 7 it is clear that system-level (i.e,, tracking) performance cannot be inferred directly from measurement-level (Le., resolution-cell shape) performance. 2) For low motion uncertainty, corresponding to nearly constant velocity target motions (process noise uq = lop2 ds’.’), FM-outperforms CF. However, for more-maneuvering targets (oq= 0.3 m / ~ ‘ . ~ ) the opposite is true. This underscores the need for system-level comparison. In the former case, the tracking and measurement uncertainties are complementary; with respect to Fig. 5(g) their overlap 618 (essentially a “stripe” across this resolution cell) is small, and the performance is good. In the latter case the tracking uncertainty and filter gains are large; hence the controlled joint rangelrange-rate uncertainty of Fig. 5(a) is preferable to the large (albeit strongly correlated) joint uncertainties of Fig. 5(g). 3 ) In all cases shown here a value y = 50% (half CF, half €34-) is close to optimal. The improvement IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 2 APRIL 1998 1 'Fi:Fi 20 0 -10 -20 -0.5 [set?] 0 0.5 o':mi -0.2 -0.5 -0.1 0 0.2 -5 -0.1 0 0.1 IW vi2' 0 ' 10 ' 20 ' 30 ' 50 ' ' I 60 70 80 90 100 60 70 80 90 100 (b) 0.1 0 ' 40 / OI -1 0 10 20 30 40 50 (C) -* -0.1 0 (e) 0.1 -0.2 0 (f) 0.2 -0.5 0 0.5 (9) Fig. 8. Resolution cells (probability of detection contours) and parallelograms that contain them in delay/Doppler space for FM+-FM- waveforms,'P = 0.001 and SNR = 20 dB, frequency sweep factor IF. = 50 Hds and K, = -50 Hds, respectively, and probability of detection threshold of 0.80. (a) y = 0% (FM-). (b) y = 20%. (c) y = 40%. (d) y = 50%. (e) y = 60%. (0 y = 80%. (8) y = 100% (FM'). Horizontal and vertical axes are in seconds and Hertz, respectively, a 1 s pulse with 50 Hz/s is assumed. FM' pulse duration (7. in o/r of the total pulse duration) Fig. 9. Measurement error variances for FM+-FM- pulse as function of y,same parameters as Fig. 8. (a) Delay variance. (b) Doppler-shift variance. (c) Correlation coefficient between delay and Doppler shift. in range estimation, as compared with FM-only or CF-only can be as high as an order of magnitude. 4) The alternating-pulse scheme, pictured as the horizontal lines in Fig. 7, has excellent performance for nearly constant velocity targets. 10 For the FM+-FW waveform, the pure FM' and the 50%-50% apportionment are the most appealing. RAG0 ET AL.: DETECTION-TRACKING PERFORMANCE WITH 30 40 50 60 70 80 90 100 50 60 70 80 90 100 (a) [(m/sec)'] The lack of smocithness in these figures is attributable to the sometimes weird resolution-cell shapes encountered. The second waveform analyzed was FM+-FM(upsweep followed by downsweep, see Fig. 4(b)). For various values of y we show in Fig. 8 representative constant-P, contclurs for this case. It should be noted that even when the measurement errors in range and range rate are symmetric around the 50%-50% pulse (see Fig. 9), the steady-state tracking errors are not (see Fig. 10). This is because as we go from a pure EM- waveform to a pure FM+, the resolution cell rotates, going from a positive correlation between range and range-rate errors to a negative correlation, as illustrated in Fig. 8, case (a) and (c). At the tracking level, negatively correlated measurements errors tend to compensate (as opposed to the positively correlated ones), leading to smaller estimation errors than the ones corresponding to the positive correlated case (see [8, p. 1191). We can summarize this as follows. 20 10" l0-'O 10 20 30 40 (b) FM' pulse duration (y. in % of the total pulse duration) Fig. 10. Steady-state estimation errors as function of y,same parameters as Fig. 8 (FM+-FM-), carrier frequency f, = 3.5 kHz. (a) Range variance. (b) Range-rate variance. Dashed lines: uz = 0.1; solid lines: u2 = 0.01; dash-dot lines: u,' = 0,0001. For each case horizontal lines refer to case of alternating FM+/FMpulses. In view of the last conclusion, we revisit the CF-FM case, but now with a positively swept FM pulse (FM+-CF (see Fig. 4(c)). The results in this case are shown in Fig. 11 and 12. As we can note now, only for high uncertainty motion targets the 50% combined FM+-CF waveform is optimal. For the other cases, the pure positively swept FM pulse performs as good as or better than the combined one, however, not by much. Our findings are summarized in Table I. The target acceleration corresponding to highest value of COMBINED WAVEFORMS 619 I 0.06, 0 10 20 40 30 50 60 70 80 90 100 TABLE I Estimation Range Error Variances in m2, for Various Schemes in Sonar Case I 9 x lo3 6 x lo4 1 x 104 pure CF pure FM- I I0 ' FM+ D -I I~ r -P - it 50% CF. 50% FM- H 50% CF. 50% FM+ 10 i0 30 20 11 /I I I 6 x 10' 3 x 10' 2 x io1 3 x io* 3 x 101 4 x lo1 1 ti 1 1oo to-zO I1 2 x lo3 2 x lo4 1 x io3 1 ! 1 104 I 1 103 I 1 104 1 x 103 6 x lo3 7 x 10' 2x10~ 4x10~ 1 x io3 1 x.104 4 x 103 2 104 60 50 70 80 90 100 - 50% FM+. 50% FMalt CF. FMalt CF. FM+ alt FM+. FM- L 1x10~ 3 x 10' 7 x 10' Note: In each column, the best is in boldface. FM' pulse duration (7.in % of the total pulse duration) Fig. 11. Measurement error variances for FM+-CF waveform as function of y,same parameters as Fig. 5 but with positive FM swept. (a) Delay variance. (b) Doppler-shift variance. (c) Correlation coefficient between delay and Doppler shift. 1 0 10 20 30 40 50 60 70 80 90 100 10'1 0 10 ' ' 20 30 40 50 60 70 60 90 100 1 (b) FMt pulse duration ( 7 . in 9% of the total pulse duration) 10-31 0 10 3 20 i 30 40 50 60 70 80 90 Fig. 13. Steady-state estimation errors as function of y (FM+-FM-) for radar case described in Section IVB, same parameters as Fig. 8), carrier frequency f, = 40 GHz, pulse length T = 100 ,us. (a) Range variance. (b) Range-rate variance. Dashed lines: o2 = 10; solid lines: 'e = 100; dash-dot lines: e: = 1000. For zach case horizontal fines refer to case of alternating FM*/FM- pulses. 100 (4 FM' pulse duration (7:in % of the total pulse duration) Fig. 12. Steady-state estimation errors as function of y (FM+-CF), carrier frequency ,& = 3.5 kHz. (a) Range variance. (b) Range-rate variance. Dashed lines: e: = 0.1; solid lines: e,"= 0.01; dash-dot lines: e2 = 0,0001. For each case horizontal lines refer to case of alternating CF/FM+ pulses. oq considered is 0.06 d s 2 , typical of the turn of a submarine. 8. Radar Case In this case we have assumed a relatively long radar pulse, Tp = 100 p s and a carrier frequency off, = 40 GHz. This pulse combined with the high carrier frequency gives a frequency resolution well below the expected Doppler shift. Given the 620 previous results for the sonar case, we investigate the FM+-FM- and FM+-CF cases only. The FM modulation used corresponds in this case to IC. = +500 kHz/s. The scan-to-scan time is At = 1 s; the SNR is 20 dB. The first waveform we analyzed was the FM'-FM-. The steady-state tracking errors for three different process noise levels are shown in Fig. 13, and as we can see, the pure FM' pulse performs the best. Next we analyzed the FM+-CF waveform, for the same radar conditions. As in the previous example, the pure FM' waveform performs the best in terms of steady-state estimation errors, but in this case the 50%-50% pulse performs nearly optimally (see Fig. 14). These results are strongly dependent on the pulse duration (or the range and range-rate resolution). An interesting point to note here is that as we increase the pulse length, the pure FM' is no longer optimal; IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 2 APRIL 1998 "[I 1' 0 1 I (m/sec) *] 1 0 0 , 10 k 20 30 50 40 60 70 I 1W 90 80 (b) FMt pulse duration (7,in % of the total pulse duration) Fig. 14. Steady-state estimation errors as function of y for radar case described in Section IVB (FM+-CF). (a) Range variance. (b) Range-rate variance. Dashed lines: g : = 1000; solid lines: g,' = 100; dash-dot lines: 0: = 10. For each case horizontal lines refer to case of alternating CFIFM+ pulses. ,* , . . 30 do TABLE I1 Estimation Range Error Variances in m2, for Various Process Noise Intensities and Waveform Schemes, in Radar Case (Pulse Length 100 p s ) ----- f 'Q1o Fig. 16. Steady-state estimation errors as function of y for same case as Fig. 14 (FM+-CF) but with pulse length of Tp = 200 p s . (a) Range variance. (b) Range-rate variance. Dashed lines: 0," = 1000; solid lines: 0,'= 100; dash-dot lines: of = 10. For each case horizontal lines refer to case of alternating CFlFM' pulses. A 10 50 $0 A 80 90 pure i o FM- (a) alt CF. FM+ _. H nlt FM+. FM10 20 30 SO 40 60 70 80 90 100 (b) FM' pulse duration (y in I I I 2 x 104 I 5 x 103 Nore: In each column, the best is in boldface. ,--_ "'0 ~ I1 7% of the total pulse duratlon) Fig. 15. Steady-state estimation errors as function of for same case as Fig. 13 (FM+-FM-) but with pulse length of Tp = 200 p s . (a) Range variance. (b) Range-rate variance. Dashed lines: o2 = 1000; solid lines: ut = 100; dash-dot lines: u,' = 10. For each case horizontal lines refer to case of alternating FM+/FM- pulses. the 50%-50% waveform replaces it. Also worth mentioning here is the fact that the alternating scheme can perform slightly better than the 50%-50% pulse for low motion uncertainty targets. The steady-state tracking errors for a pulse length Tp = 200 ps (the SNR and P ' are assumed the same as with the 100 ps waveform, i s . , the RCS is smaller) are shown in Figs. 15 (FM+-FM-) and 16 (FW-CF). Our findings are summarized in Tables I1 and 111. The target acceleration corresponding to the highest value of crq is 32 d s 2 . II 2.5 x 103 i TABLE I11 Estimation Range Error Variances in m2, for Various Process Noise Intensities and Waveform Schemes, in Radar Case (Pulse Length 200 p s ) 1) 0,"= 1000 I 0II e.mire CF -~ purc FMpure FM+ 50% CF, 50% FM+ 50% _.FMt. 50% FMalt CF. FM+ alt FM+, FM- 11 8 x 10' I1 4.5 x 106 G x lo4 2.5 x 104 1 x lo4 II 2 5 x 104 II 104 Note: In each column. the best is in H H It I u,"= 100 8 x lo5 I 3 x 106 3.5 x lo4 7 x 103 4 x lo3 I I 4.5 x 103 I 5.5 103 boldface. 1 o,' = 10 1 I 8 x lo5 flU I 7 x lo5 8 x lo3 I I I 5 x 103 4 x lo3 2 . 5 x 103 2.5 103 nH C. Discussion of Results The previous two sections have dealt with three scenarios: sonar, shorter pulse radar, and longer pulse radar. We cannot hope to be exhaustive in our study; but there is some variety here, and we can claim some insight. RAG0 ET AL.: DETECTION-TRACKING PERFORMANCE WITH COMBINED WAVEFORMS 62 1 1) Downsweep FM (m-), in its pure form or in combination with CF7is almost never a good idea. Intuitively, the problem is that range and range-rate measurement errors are positively correlated, and this is undesirable given a kinematic target motion model. 2) Pure upsweep FM (EM’) performs well in almost all cases studied. This is due to the negative correlation of the range and range-rate measurement uncertainties, which yields a partial cancellation of the errors [8]. It is occasionally outperformed (for example, in the longer pulse radar case) by combination CF/FM+ and FM+/F’Iv- waveforms. 3) The “alternating” schemes, in which a pulse of one type is followed by a pulse of another type, appear to offer a very good suboptimal performance. It is our judgment that FM+/FM- is the most robust performer, followed closely by CF/FM+. V. SUMMARY In this paper we have investigated the performance of combined constant- and swept-frequency waveform fusion systems. Our comparison is in terms of steady-state tracking performance, via the ARE as modified to account for missed detections. Fundamental to our analysis is the resolution cell primitive (defined for our purposes as the volume of rangekange-rate space wherein the probability of detection exceeds a given threshold) and its implied measurement error. Our results indicate that the overall detection-tracking performance is strongly dependent on the waveform used, and that the use of the optimal waveform can lead to dramatic improvement in tracking error (as much as an order of magnitude range-estimation error variance reduction). A major conclusion of this study is that, from a system level (i.e., tracking) point of view, EM+ (upsweep) is always superior to FM- (downsweep). Overall, a 50%-50% division of energy between EM+ and FM- seems the best choice. However, a scheme using alternating CF and FM+ waveforms, available through only minor modifications to conventional signal processing, works well also. 4 Y(‘ 1 f0) fd I I2 X (s(i - T o ) e > 2 ~ f o [ ~ - T o ) )* Fig. 17. Magnitude square matched filter operation. We begin by repeating the definition (4) (28) of the normalized ambiguity function, and to introduce the magnitude-square matched-filter operation of Fig. 17. With reference to this latter, the signal y(7,fd; 7-o,fo) is that which would appear as output for a filter matched to a signal having delay T~ and Doppler shift f o , when the actual signal has delay T and Doppler shift fd. Our goal is to relate y to the ambiguity function, and this turns out to be straightforward: Y(T>>,f3,.;df IL 11:” +a3 As(t - 7 ) e l z T f d ( t - T ) s * ( t - ,o)e-J2nfO(L-TO) = As(X)s*(X = =4 7 0 - 7-3 fd - fo) + - & ),lza(k-f0)X,-12nf0(T~T0) I’ dX )2 (29) in which the effect of noise has been ignored. We next turn to the concept of the resolution cell. It is assumed that ( 7 ,fd)-space must be (uniquely and exhaustively) tesselated into regions, usually parallelograms,’ whose target occupancies will be indicated by a threshold-exceedance at a filter matched to the centroid ( 7 ,fd) value.2 Without loss of generality we focus attention on the resolution cell matched to zero delay and Doppler shift. Ideally , is this cell ought to match a region where y ( ~fd;O,O) large; however, since we can write APPENDIX A. MATCHED FILTERS, AMBIGUITY FUNCTIONS, AND RESOLUTION CELLS we have that this cell ought to be a parallelogram into which can be inscribed a time-reflected ambiguity function contour, where the level provides some minimum Pd value of interest (for much of this work we have chosen this as Pd = 0.80). Thus we take this element A(-7,,fd)to be, for want of a more widely accepted term, a “resolution-cell primitive,” from which a tesselating resolution cell grid can be determined. As there appears to be some confusion both in common reference and in the relevant literature, we offer this brief primer on the relationship between the above concepts. ‘We are here assuming that only range and range-rate are of interest, an inherently one-dimensional case; more generally we have parallelopipeds. ’This amounts to sampling in (r,,fd)-space. ACKNOWLEDGMENTS Thanks to E E. Daum for the suggestion to investigate the FM+-FM- cases; and to W. Cao for his comments on the manuscript. 622 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34. NO. 2 APRIL 1998 As a penultimate note, we recall that while range is directly proportional to delay T , range-rate is proportional to the negative Doppler frequency shift -jd. Thus, due to symmetry of the ambiguity function, the resolution cell primitive in rangehange-rate space has the same shape (although different scale) and orientation as does the ambiguity function in delay/Doppler space. Since it appears that d(-7,fd)rather than d ( T , f d ) is the quantity of interest, it is reasonable to question the motivation for defining the ambiguity function in the way it is defined. As our final note, we respond to this. The answer is that many current systems do not attempt to tesselate delay/Doppler space with resolution cells, but use a single filter matched to f d = 0 for continuous (or finely sampled) values of 7 . The quantity of interest is in this case ~ ( 7 , f d ; ~ o , O=) 4 7 0 - 7 , f d ) (31) meaning that a target with Doppler frequency shift f d produces a matched-filter trace whose shape is as of a horizontal “slice” of the ambiguity function of (4) at frequency fd. This idea and its effects on tracking are well discussed in [2]. Daum, F. (1992) A system approach to multiple target tracking. In Y. Bar-Shalom (Ed.), Multitarget-Multisenor Tracking: Applications and Advances, Vol. 11. Artech House, 1992, 149-181. [3] Levanon, N. (1988) Radar Principles. New York: Wiley, 1988. [4] Skolnik, M. L. (1990) Radar Handbook. New York: McGraw-Hill, 1990. [SI Li, X.-R., and Bar-Shalom, Y. (1995) Stability evaluation and track life of the PDAF for tracking in clutter. IEEE Transactions on Automatic Control, 36, 5 (May 1991), 588-602. [6] Li, X.-R., and Bar-Shalom, Y. (1995) Performance prediction of hybrid state algorithms: The scenario conditional averaging approach. In C. T. Leondes (Ed.), International Series on Advances in Control and Dynamic Systems, Vol. 72. New York: Academic Press, 1995. [7] Stein, S . (1981) Algorithms for ambiguity function processing. IEEE Transactions on Acoustic, Speech and Signal Processing, ASSP-29, 3 (June 1981), 588-599. [8] Bar-Shalom, Y., and Li, X.-R. (1993) Estimation and Tracking: Principles, Techniques and Software. Boston: Artech House, 1993. [2] REFERENCES [l] Kershaw, D. J., and Evans, R. J. (1994) Optimal waveform selection for tracking systems. IEEE Transaction7 on Information Theory, 40, 5 (Sept. 1994), 1536-1550. Constantino Rag0 received his Bachelor Degree from the University of La Plata, Argentina, in 1986, and the M.Sc. and Ph.D. degrees from the University of Connecticut, Storrs, in 1994 and 1995, respectively, all in electrical engineering. From 1986 to 1991 he was affiliated with the “Laboratorio de Electrbnica Industrial, Control e Instrumentacibn (LEICI)” at the Department of Electrical Engineering, University of La Plata, working in the areas of digital signal processing and spectral estimation. From 1991-1995 he was a graduate researchheaching assistant in the Department of Electrical and Systems Engineering, University of Connecticut. Since 1995 he has been with Scientific Systems Company, Inc., Wobum, MA, as a research engineer. His research interests include detection and estimation theory, failure detection and identification, and data fusion and tracking. Peter Willett received his B.A.Sc. in 1982 from the University of Toronto, Toronto, Canada, and his Ph.D. in 1986 from Princeton University, Princeton, NJ. Since 1986 he has been with the University of Connecticut, Storrs, where he is an Associate Professor. His interests are generally in the areas of detection theory and signal processing. RAG0 ET AL.: DETECTION-TRACKING PERFORMANCE WITH COMBINED WAVEFORMS 623 Yaakov Bar-Shalom (S’63-M’66-SM’80-F’84) was born on May 11, 1941. He received the B.S. and M.S. degrees from the Technion, Israel Institute of Technology, in 1963 and 1967 and the Ph.D. degree from Princeton University in 1970, all in electrical engineering. From 1970 to 1976 he was with Systems Control, Inc., Palo Alto, CA. Currently he is Professor of Electrical and Systems Engineering and Director of the ESP Lab (Estimation and Signal Processing) at the University of Connecticut. His research interests are in estimation theory and stochastic adaptive control and has published over 200 papers in these areas. In view of the causality principle between the given name of a person (in this case, “(he) will track”, in the modern version of the original language of the Bible) and the profession of this person, his interests have focused on tracking. His other interests are stochastic control of vertical airfoils and of pairs of inclined foot supports on crystals. He coauthored the monograph Tracking and Data Association (Academic Press, 1988), the graduate text Estimation and Tracking: Principles, Techniques and Sojhvare (Artech House, 1993), the text Multitarget-Multisensor Tracking: Principles and Techniques (YBS Publishing, 1995), and editcd the books Multitarget-Multisensor Tracking: Applications and Advances (Artech House, Vol. I, 1990; Vol. 11, 1992). He has been elected Fellow of IEEE for “contributions to the theory of stochastic systems and of multitarget tracking”. He has been consulting to numerous companies, and originated the series of Multitarget-Multisensor Tracking short courses offered via UCLA Extension, at Government Laboratories, private companies and overseas. He has also developed the commercially available interactive software packages MULTIDATTMfor automatic track formation and tracking of maneuvering or splitting targets in clutter, PASSDATm for data association from multiple passive sensors, BEARDATTMfor target localization from bearing and frequency measurements in clutter, IMDATTMfor image segmentation and target centroid tracking and FUSEDATTMfor fusion of possibly heterogeneous multisensor data for tracking. During 1976 and 1977 he served as Associate Editor of the IEEE Transactions on Automatic Control and from 1978 to 1981 as Associate Editor of Automatica. He was Program Chairman of the 1982 American Control Conference, General Chairman of the 1985 ACC, and Co-Chairman of the 1989 IEEE International Conference on Control and Applications. During 1983-1987 he served as Chairman of the Conference Activities Board of the IEEE Control Systems Society and during 1987-1989 was a member of the Board of Governors of the IEEE CSS. In 1987 he received the IEEE CSS Distinguished Member Award. Since 1995 he is a Distinguished Lecturer of the IEEE AESS. He is co-recipient of the M. Barry Carlton Award for the best paper in the IEEE Transactions on Aerospace and Electronic Systems in 1995. 624 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 2 APRIL 1998