DETECTION ALGORITHMS AND TRACK BEFORE DETECT ARCHITECTURE BASED ON NONLINEAR FILTERING FOR INFRARED SEARCH AND TRACK SYSTEMSThe algorithms, software and experiments developed by Skirmantas Kligys, Boris Rozovskii and Alexander Tartakovsky. Click here to watch Detection and tracking in real time. Cruise missiles over land and sea cluttered background are serious threats to Infrared Search and Track Systems (IRST). In general, these threats are stealth in both the infrared and radio frequency bands. That is, their thermal infrared signature and their radar cross section can be quite small. Future predicted threats, i.e. the next generation of cruise missiles, will be even more difficult to detect at a sufficient range to counter. Further, low elevation trajectory objects, such as sea skimming missiles, have radar signals with large amounts of temporally and spatially correlated interference called multipath. This multipath problem remains an enormous obstacle to existing detection and tracking algorithms. Hence, new technology is needed which will allow for the timely detection, tracking, and identification of such threats. IRST systems are one component of a multisensor suite which can meet the technical challenge of the timely detection/track/identification of low signal-to-(noise+clutter) ratio (S(N+C)R) targets. The multisensor suite should include an IRST, Radar, and a coherent laser (Lidar). We envision a cueing hierarchy where the IRST can cue the Radar or the Radar cues the IRST. Once a candidate track is established the Lidar can be used to identify the target by its micro doppler signature. In this project, we have developed computationally efficient algorithms and adaptive architecture with optimized overall performance (statistical and computational) for real-time reliable detection and tracking of low-observable targets in IRST systems. Despite the fact that we focus on an IRST against cruise missiles over land and sea cluttered backgrounds, the results are equally applicable to other sensors (e.g., Radar, Lidar) and other kinds of targets (e.g. ballistic missiles). In the research we concentrated on the three interrelated problems:
The generalized block-diagram of the system under investigation is shown in Figure 1. We develop both the signal processing architecture (clutter removal algorithms and TbD algorithms) and track detection algorithms.
Figure 1. Generalized block-diagram of the developed system New algorithms are developed to meet the following important practical requirements.
Appearance/Disappearance Detection and Tracking of a Skimming Missile Below we present the results of application of developed detection and nonlinear filtering (ONF) algorithms for detection and tracking of multiple targets that appear and disappear at unknown points in time. The algorithms were tested with the use of the real IR background obtained from SPAWAR Systems Center, San Diego, CA. The figures below contain the results of testing for these real cluttered images. We have applied adaptive sequential algorithms that allow for timely detection of both track appearance and disappearance. These algorithms exploit the results of TbD - the estimates of target’s spatial location based on ONF (see below) are used in the decision statistics. In other words, we use the system that is built upon simultaneous, joint detection/estimation/tracking algorithms: detection depends on preliminary tracking while tracking depends on the results of detection. Optimal Spatio-Temporal Nonlinear Filtering for Track-Before-Detect Decision statistics used for target track appearance/disappearance detection exploit the estimates of target location. In turn these estimates are built upon preliminary tracking – the so called track-before-detect procedure. The TbD algorithm provides coherent processing in detection stage and is based on the optimal spatio-temporal nonlinear filtering. The measurements
where The expected range of possible "behaviors" of the target is modeled as a Markov process. Often this process may be well described by a randomly perturbed multi-dimensional linear or nonlinear dynamical system
where The output of the ONF is the functional time series For computing the optimal nonlinear filter (JPD) we use the Spectral Separation Scheme (the SSS algorithm). This algorithm does not involve solving of PDE's on-line and is recursive in time and space. These properties are very important since they allow adaptive sequential multi-resolution filtering to be performed. We note that the conventional nonlinear filtering procedures, such as the Extended Kalman Filter, do not work in our applications, since the posterior distribution is typically multi-peak and hence is far from being Gaussian. Figure 2 shows the results of tracking before detection of a low observable surface skimming missile (SNR=-3.1dB). It is seen that the optimal nonlinear filter exactly follows the target after processing of 6 frames. The results of filtering (maximum posterior estimate and posterior mean) are then used in adaptive detection algorithms to form the decision statistics. Target Detection and Tracking Figure 3 illustrates the behavior of the adaptive CUSUM (cumulative sum) statistic U*n for different S(N+C)R (SNR=-0.25dB, -3.62dB, and SNR=-6.6dB, respectively) . It is seen that most of the time this statistic is close to zero or negative when the target is absent. Sometimes, however, peaks arise. These peaks may be identified with target presence. But they are typically short and may be easily distinguished from the peaks due to target. One possible method to discriminate between false alarms and true target is to make the decision on target presence if there are several subsequent exceedances of the threshold. Otherwise the decision on target absence is made (i.e. a single exceedance is identified with false alarm). In our experiments we observed visible difference in behavior of the decision statistics when the target appears compared to the case of its absence up to the SNR -6.6dB. In the pictures the first target appears at time n=1 and disappears at n=28. The second target appears at time n=39 and never disappears. Figure 4 shows the results of detection of
track appearance and disappearance by the adaptive detection algorithm for SNR=-0.25dB and
SNR=-6.6dB. The detection of tracks occurs when the adaptive CUSUM exceeds the threshold a
(the upper one) and track disappearance is detected when another adaptive statistic When the detection occurs the target trajectory is estimated by the ONF algorithm, which is identical to described above TbD procedure. The results of tracking are shown in the figure by asterisks. |
are collected at discrete time moments 
and the
relationship between the observation and target location is modeled by a nonlinear
measurement equation of the form
represents the spatial coordinate in the plain, 
is the true location of the target at time 
, 
is a signal amplitude, 
is a signal function (response of the sensor to target) and 
is the measurement noise process. 
is a
noise process that describes uncertain and unpredictable target motion; f is a
known function. 
of joint posterior densities (JPD)
which measure the likelihood at the time moments 
that the vector of target features parametrized by 
is in close proximity of the grid
point 
. Computation of the JPD
splits into two separate procedures: at every time step, spatio-temporal filtering for
clutter removal is done first, and then the spatio-temporal nonlinear filtering is
performed to estimate the target location.
exceeds the threshold b
(the lower one). When the decision on target disappearance is made, the statistic U*n
is renewed form zero, i.e. the detection algorithm is prepared to detect another
target. It is seen that the algorithm is able to detect even very low SNR targets. The
thresholds were chosen such that the false alarm rate and false detection of target
disappearance would be guaranteed at the specified levels. There are two targets in the
pictures: the first target appears at time n=1 and disappears at n=28 while the second one
appears at time n=39 and does not disappear. The algorithm detects the first target with
the delay about 20 seconds (20 frames) for SNR=-6.6dB and 4-6 frames for SNR=-0.25dB. Then
the fact of target's disappearance is detected with very small delay. Finally the second
target is detected with the delay about 5-6 frames for SNR=-6.6dB and 2-3 frames for
SNR=-0.25dB.