Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224341534

Controlling track coalescence with scaled Joint Probabilistic Data Association

Conference Paper · October 2008


DOI: 10.1109/RADAR.2008.4653963 · Source: IEEE Xplore

CITATIONS READS

16 214

1 author:

Hugh Lachlan Kennedy


Technical Knockout Systems Pty. Ltd.
84 PUBLICATIONS   271 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Acoustic Signal Processing View project

Digital Signal Processing (1-D DSP) View project

All content following this page was uploaded by Hugh Lachlan Kennedy on 09 March 2015.

The user has requested enhancement of the downloaded file.


Controlling Track Coalescence with Scaled Joint
Probabilistic Data Association
Hugh L. Kennedy
Technical Knockout Systems Pty. Ltd.
LPO Box 8150, ANU Acton, Canberra 0200
hkennedy@tkosystems.com.au

Abstract— Track coalescence is undesirable when estimating the states of multiple manoeuvring targets, with a less-than-unity
probability of detection, in clutter. Simple and compound forms of coalescence are defined and discussed. Simple coalescence is when
two or more identical tracks follow a single target; compound coalescence is when two or more identical tracks follow the midpoint of
two or more targets. It is shown that the incidence of compound track coalescence in Joint Probabilistic Data Association (JPDA) may
be reduced using a scaling factor to favour the most likely association hypothesis. This prevents multiple hypothesis equivalence when
tracking closely-spaced or crossing targets. The performance of the scaled JPDA (SJPDA) algorithm is compared with Probabilistic
Data Association (PDA) and JPDA using real and simulated data. Larger scaling factors decrease the likelihood and duration of
compound track coalescence; however, they also increase the likelihood of track divergence on clutter or other targets. A value of
unity corresponds to JPDA. The optimal value may be chosen to suit the application. A factor of two was found to give good results in
the test data.

I. INTRODUCTION
In manned and unmanned aerial surveillance missions over populated littoral environments, fast and accurate solutions to the
problem of automatically detecting and tracking multiple surface targets in clutter are required. Joint Probabilistic Data
Association (JPDA) [1] is ideal in sensor systems involving Ground Moving Target Indicators (GMTI) [2], Electro Optic (EO)
and Infra Red (IR) cameras or other high-resolution imaging devices, which typically capture digital representations of highly
complex scenes containing multiple closely-spaced targets of high interest, set against a cluttered background comprised of low-
interest moving objects and correlated noise. JPDA is sufficiently complex to handle cases where peak-to-track assignment is
highly ambiguous; yet, when implemented appropriately, it is sufficiently simple for real-time operation.
JPDA is a multi-target extension of Probabilistic Data Association (PDA) [3],[4] &[5]. PDA considers each track in isolation,
while JPDA updates groups of closely-spaced tracks concurrently; both consider all feasible peak-to-target assignments but
reduce the multiple hypotheses to a single Gaussian estimate for each track. As PDA permits multiple tracks to be updated by
the same dominant peak, they may all converge (or merge, or coalesce) on the same target. As JPDA forces tracks to compete
for the same peak – with most of the peak used by the track with the most likely association probability – this problem occurs
less frequently. To facilitate the relative performance analysis of multi-target trackers, two types of tracking errors are defined
here – simple coalescence and compound coalescence (see Fig. 1). Simple coalescence is when two or more identical tracks
follow a single target; compound coalescence is when two or more identical tracks follow the midpoint of two or more targets.
Coalescence may be brought about by a number of different mechanisms; two typical causes will now be described. Simple
track coalescence commonly occurs when a track on one target is seduced by the peaks of another, during a period of reduced
detectability. A new track then starts on the target when detectability is restored. Compound coalescence occurs when two
tracks are brought together as targets converge and cross. Tracks remain coalesced even when the targets diverge and the peaks
are well separated. The sensor may be unable to resolve the targets as they pass each other; however, this is not a necessary
condition for compound coalescence to occur. Simple coalescence is a common feature of PDA-based trackers but is rare in
JPDA-based trackers; compound coalescence is more likely when JPDA is used.
x

Fig. 1 Illustration of forms of track coalescence: simple (top) and compound (bottom). Targets are depicted in red, tracks in blue.

The desirable behaviour of JPDA comes at a high computational cost because it involves a combinatorial calculation where
all contested peaks are permuted among all nearby tracks and the likelihood of each association hypothesis evaluated. It is
therefore not usually considered as a candidate tracking algorithm in mission-critical surveillance systems; although, real-time
JPDA implementations are now emerging, as a result of recent advances in the processing capability of digital processors and
hypothesis management strategies [6]. Since its inception in the 80s, simpler and faster approximations to JPDA have been
investigated [7],[8]&[9]. These methods aim to reduce the occurrence of track coalescence, without incurring the computational
overhead of complete JPDA, by applying first-order corrections to the PDA equations. Inevitably, these suboptimal algorithms
do not match the performance of complete JPDA in all cases.

II. THEORETICAL OVERVIEW


The incidence of compound coalescence increases as the target velocity, angle of incidence and the sensor resolution
decrease. As the targets approach each other, the most likely association hypothesis and the penultimate association hypothesis
converge and become more alike. Immediately before the peaks merge, the likelihood of both hypotheses may be numerically
equivalent if the approach has been sufficiently gradual. As a result of this equivalence, the JPDA algorithm does not give one
track priority over the other when common peaks are used; therefore, the peaks are used equally by both tracks and convergence
continues. Hypothesis equivalence persists even after the targets have separated, until random perturbations due to noise cause
the hypotheses to differ slightly. At this point the JPDA algorithm is able to function properly and the tracks quickly separate
and follow different targets.
An exact nearest neighbour version of PDA (EN-NPDA) was proposed in [10], as a way of solving the problem of simple
and compound track coalescence in PDA trackers and compound track coalescence in JPDA trackers. EN-NPDA involves the
enumeration of all peak-to-target association hypotheses, as per JPDA; however, only the most likely hypothesis is retained.
While this approach inhibits compound coalescence, it was found to lead to track over commitment and an increased incidence
of track divergence in the presence of clutter and absent target detections. A less severe approach was later proposed in [11],
through the application of selective pruning, which effectively weights the outcome in favour of the more likely hypotheses.
The problem is also reposed as a linear descriptor system with stochastic difference coefficients. The method is extended to
include integrated track maintenance in [12].
An alternative re-weighting strategy is employed in this paper, with the likelihood of the most likely JPDA hypothesis
multiplied by an arbitrary positive scaling factor, prior to re-normalization. When the scaling factor is unity, JPDA results; as
the scaling factor approaches infinity, EN-NPDA results. Use of a greater-than-unity scaling factor ensures that the most likely
hypotheses are never equivalent. If peak files are pre-generated, or produced using a separate processor, the tracking algorithm
is able to run in real time on a laptop computer. This is achieved through the efficient enumeration of hypotheses. Performance
of scaled JPDA, relative to PDA and JPDA, is assessed in the following sections. Real data are used to support the qualitative
assessment of tracking characteristics. Simulated data are used to quantify differences between the algorithms and to determine
the optimal value of the scaling parameter.
The PDA algorithm is based on the basic design described in [3]; signal strength information is disregarded and clutter is
assumed to be uniform. The target’s state is estimated in four dimensions – position and velocity in x and y coordinates. The
baseline JPDA tracker follows the standard formulation described in [1]; all peak-to-track association hypotheses, with non-
negligible weights, are considered. All trackers are integrated with automatic track initiation and termination logic [13].
III. REAL DATA
Groups of closely spaced and manoeuvring surface targets (a paddling of ducklings on a lake) were observed using an EO
sensor (a digital video camera). The data were down sampled to 7.5 frames per second, 320 by 240 pixels per frame and 8 bits
per colour per pixel. The Red-Blue-Green (RGB) image was converted to a gray scale image and inverted. The inverted image
was then convolved with a filter designed to stop constant background features and pass Gaussian foreground features (see Fig.
2). The filter was tuned to respond to Gaussian blobs with a standard deviation of 1 pixel. The 9x9 convolution mask used in the
filter is given in Table I. Specific masks for pixels near the edges of the image were also synthesised, to ensure that no data
were discarded around the image perimeter as a result of filtering.

Fig. 2 Raw (top) and processed (bottom) image example.

TABLE I
FILTER CONVOLUTION MASK

-0.0292 -0.0292 -0.0292 -0.0291 -0.0291 -0.0291 -0.0292 -0.0292 -0.0292


-0.0292 -0.0292 -0.0287 -0.0267 -0.0250 -0.0267 -0.0287 -0.0292 -0.0292
-0.0292 -0.0287 -0.0223 0.0017 0.0218 0.0017 -0.0223 -0.0287 -0.0292
-0.0291 -0.0267 0.0017 0.1094 0.1992 0.1094 0.0017 -0.0267 -0.0291
-0.0291 -0.0250 0.0218 0.1992 0.3474 0.1992 0.0218 -0.0250 -0.0291
-0.0291 -0.0267 0.0017 0.1094 0.1992 0.1094 0.0017 -0.0267 -0.0291
-0.0292 -0.0287 -0.0223 0.0017 0.0218 0.0017 -0.0223 -0.0287 -0.0292
-0.0292 -0.0292 -0.0287 -0.0267 -0.0250 -0.0267 -0.0287 -0.0292 -0.0292
-0.0292 -0.0292 -0.0292 -0.0291 -0.0291 -0.0291 -0.0292 -0.0292 -0.0292

Approximately 5000 frames of data were processed by the trackers. The output was inspected visually. The data selected for
presentation in this section are typical examples that qualitatively demonstrate that the PDA, JPDA and scaled JPDA methods
have very different tracking characteristics.
In Fig. 3 and Fig. 4, tracks are set against the underlying filtered EO sensor data. Each track is shown in blue with – the
numerical track ID; the current estimated target position as a dot; the track history as a solid line (in Fig. 3 only); and the track
estimate uncertainty as a blue square, centred on the current estimated target position, with sides set so the area enclosed by the
square equals the area enclosed by the 98% probability ellipse of the spatial state covariance matrix. The peaks produced by the
detector in the current frame are plotted as green dots surrounded by green circles with a fixed and arbitrary radius. The target
features underlying the peaks are also visible as white blobs, with clutter as gray texture. The detection threshold has been set to
optimize the probability of target detection and minimize the probability of clutter detection. While clutter uniformity was
assumed in the tracker models, the actual density was highly spatially and temporally correlated. The peaks produced by the
detector were then processed using PDA, JPDA and the scaled JPDA algorithm described above, with a scaling factor of two.
A. Low target-density example
In Fig. 3 all trackers merge tracks 0 and 1 to varying degrees. The PDA tracks merge early (simple track coalescence); they
soon become, and remain, identical. The merge is triggered by one of the targets temporarily becoming unobservable as a
duckling dives. In the current frame, compound coalescence is evident; later, as target separation increases, both tracks move
from the target midpoint and follow the same target. The JPDA tracks merge only when the targets cross paths, producing a
single detection; the tracks stay merged for several frames (compound coalescence), even as the targets become resolvable
again. The tracks only begin to separate in the current frame, several frames after target separation. The scaled JPDA tracks
only merge briefly while the targets are unresolvable in the raw sensor data.
Fig. 3 Tracker performance comparison. PDA (top), JPDA (middle) and scaled JPDA (bottom). Real data containing three targets.

B. High target-density example


In Fig. 4, the PDA tracker presents a confused picture with multiple tracks spanning multiple targets. Several cases of simple
and compound coalescence are exhibited. The picture is improved considerably using JPDA. The only significant difference
between JPDA and scaled JPDA is in the behaviour of the pair of tracks on the closely-spaced targets at the left edge of the
image – they are merged by JPDA (compound coalescence) but not merged by scaled JPDA.
Fig. 4 Tracker performance comparison. PDA (top), JPDA (middle) and scaled JPDA (bottom). Real data containing twelve targets.

IV. SIMULATED DATA


A. Synthesis
Simulation allows track estimates to be compared with target ground truth. The output of the detector was emulated directly.
Sixteen clutter peaks were uniformly and randomly distributed, in x-y coordinates, over the 320x240 surveillance region. The
origin of the coordinate system is in the upper left-hand corner of the image, with the x and y axes running in the horizontal and
vertical directions, respectively. Eight targets were injected. The simulation scenarios were contrived to promote target crossing
events. The targets all originate at the same x coordinate, with the y coordinate drawn randomly from a uniform distribution
between 0 and 239. The initial x and y velocities of all targets were generated according to Gaussian distributions. The x
distribution had a mean of 2.0 and a standard deviation of 0.0, pixels per frame; the y distribution had a mean of 0.0 and a
standard deviation of 0.1, pixels per frame. The targets accelerated randomly in the y dimension. This process noise was
distributed as a Gaussian variable, with a mean of zero and a standard deviation of one pixel. The x velocity was held constant.
Zero-mean Gaussian measurement noise, with a standard deviation of one pixel, was also added to the x and y coordinates of the
target peaks. The probability of target peak detection was 0.9.
One hundred simulation scenarios were generated and processed using PDA and scaled JPDA (SJPDA). Scaling factors of
1.0, 1.5, 2.0, 4.0 and 8.0 were used. A SJPDA scaling factor of 1.0 is equivalent to JPDA. Each scenario is terminated when no
targets remain in the surveillance region.
B. Analysis
A number of measures were taken to facilitate the analysis process – target peaks were always detected during the first frame,
tracks were only initiated on target peaks, tracks were only initiated in the first frame, tracks were confirmed immediately and
tracks were never deleted. These simplifications ensure that track-to-target assignment is one-to-one and unambiguous. The
distance (i.e. the error) between the estimated target position and the corresponding true target position was determined for
every track within sensor coverage, in every frame.
A typical frame from one of the scenarios is shown in Fig. 5. Ground truth is plotted in red, with target IDs, target peaks and
target history displayed. Divergent tracks due to target swapping have been produced by all processing methods. Nearby target
peaks cause a PDA track to diverge during track establishment. PDA also yields several cases of simple track coalescence.
There are no cases of simple track coalescence in JPDA or SJPDA. JPDA exhibits many cases of compound track coalescence,
evident in both the track history and the current frame. Compound coalescence is absent in the SJPDA output.
Error statistics are summarized in Fig. 6 and Table II. In Fig. 6, 99 bins of equal width were formed over the interval from 0
to 16 pixels. Long featureless tails extend beyond the upper limit; therefore, these data were omitted from the plot. In this figure,
SJPDA with high scaling factors appears to cause a slight contraction in the error distribution around zero. In an attempt to
quantify the differences in tracking performance, Table II contains error distributions collated using three coarse bins of unequal
width. Each bin is assumed to represent different modes of tracking behaviour. The first bin, containing the smallest errors (0 to
1.5 pixels), is dominated by steady-state tracking errors. Errors falling in the second bin (1.5 to 16 pixels) are thought to be
predominantly due to compound coalescence. The contribution from divergent tracks, due to target swapping, dominates the
third bin (more than 16 pixels). Binning the data in this way indicates that PDA has the greatest number of tracks in the middle
bin. It also shows that increasing the value of the scaling factor in SJPDA tends to shift errors from the middle bin into the inner
and outer bins, which suggests that the incidence of complex coalescence decreases and that the incidence of track divergence
(due to track swapping) also increases. These findings were consistent with a qualitative assessment of tracking behaviour.
Fig. 5 Tracker performance comparison. PDA (top), JPDA (middle) and SJPDA, with a scaling factor of 2.0 (bottom). Simulated data containing eight targets.
Fig. 6 Truncated error distribution – 99 bins of equal size.

TABLE II
COMPLETE ERROR DISTRIBUTION – THREE BINS OF UNEQUAL SIZE

Tracker Bin Limits (pixels)


0 to 1.5 1.5 to 16 16 to Inf.
PDA 68940 123172 710887
SJPDA, fac. = 1.0 (JPDA) 68206 120646 714148
SJPDA, fac. = 1.5 82562 98963 721475
SJPDA, fac. = 2.0 87606 92527 722867
SJPDA, fac. = 4.0 93177 84172 725651
SJPDA, fac. = 8.0 91348 81047 730605

V. DISCUSSION
If a Bayesian tracker does not know how to correctly assign tracks to closely-spaced targets as they separate and diverge,
then in the absence of further information, the most conservative course of action is for all tracks to follow the midpoint of all
targets. From a statistical standpoint, the resulting compound coalescence minimizes the expected error; from an operational
standpoint however, this outcome may be undesirable. In surveillance applications, if use of a scaling factor forces the tracker to
make the wrong track-to-target assignment decision then the operator may be able to manually re-assign the track IDs. In
interdiction applications, it is better to aim and shoot directly at two hostile targets than it is to fire at the empty space in
between. Use of a scaling factor forces the tracker to make an assignment decision, even if there is a high probability that it may
be wrong. The magnitude of the scaling factor determines the amount of pressure exerted. A factor of approximately two is
appropriate in the real and simulated data.

VI. CONCLUSION
The investigations described in this paper indicate that it is possible to tune the behaviour of the JPDA tracking algorithm, to
reduce the incidence of compound track coalescence, by simply using a scaling factor to favour the most likely association
hypothesis. This prevents multiple hypothesis equivalence when processing closely-spaced or crossing targets. In the
simulations used in this study, the optimal value of the scaling factor was found to be around two. Smaller values increase the
likelihood and duration of compound track coalescence, while larger values increase the likelihood of track divergence on
clutter or other targets. The scaling factor may be empirically adjusted to yield the desired tracking behaviour in a given
application.
REFERENCES
[1] T. Fortmann, Y. Bar-Shalom and M. Scheffe, "Sonar tracking of multiple targets using joint probabilistic data association," IEEE Journal of Oceanic
Engineering, vol. 8, no. 3, pp. 173-184, Jul. 1983.
[2] W. Koch, R. Klemm and U. Nickel, "Tracking of multiple ground moving targets with adaptive monopulse radar Part II: The tracker," in Proc.
International Radar Symposium, pp. 1-4, May 2006.
[3] Y. Bar-Shalom and E. Tse, “Tracking in a Cluttered Environment with Probabilistic Data Association,” Automatica, vol. 11, pp. 451-460, Sep. 1975.
[4] T. Kirubarajan and Y. Bar-Shalom, “Probabilistic data association techniques for target tracking in clutter,” Proceedings of the IEEE, vol. 92, no. 3, pp.
536-557, Mar. 2004.
[5] Y. Bar-Shalom, T. Kirubarajan and X. Lin, “Probabilistic data association techniques for target tracking with applications to sonar, radar and EO sensors,”
IEEE Aerospace and Electronic Systems Magazine, vol. 20, no. 8, pp. 37-56, Aug. 2005.
[6] P. Horridge and S. Maskell, "Real-Time Tracking Of Hundreds Of Targets With Efficient Exact JPDAF Implementation," in Proc. 9th International
Conference on Information Fusion, pp. 1-8, Jul. 2006.
[7] J.A. Roecker and G.L. Phillis, "Suboptimal joint probabilistic data association," IEEE Transactions on Aerospace and Electronic Systems, vol. 29, no. 2,
pp. 510-517, Apr. 1993.
[8] G. Gavriloaia, A. Sperila and A. Stoica, "An Ad-Hoc Method for Avoiding Tracks Coalescence in Pdaf for Tracks Fusion," in Proc. 7th International
Conference on Telecommunications in Modern Satellite, Cable and Broadcasting Services, vol. 2, pp. 579-582, Sep. 2005.
[9] H.L. Kennedy, "Detecting and Tracking Moving Objects in Sequences of Color Images," in Proc. IEEE International Conference on Acoustics, Speech
and Signal Processing, vol. I, pp. 1197-1200, Apr. 2007.
[10] R. J. Fitzgerald, “Development of practical PDA logic for multitarget tracking by microprocessor,” in Multitarget-Multisensor Tracking: Advanced
Applications, Y. Bar-Shalom, Ed. Reading, MA: Artech House, 1990, pp. 1–23.
[11] H.A.P. Blom and E.A. Bloem, "Probabilistic data association avoiding track coalescence," IEEE Transactions on Automatic Control, vol. 45, no. 2, pp.
247-259, Feb. 2000.
[12] H.A.P. Blom, E.A. Bloem and D. Musicki, "Joint Integrated PDA Avoiding Track Coalescence under Non-Homogeneous Clutter Density," in Proc. 9th
International Conference on Information Fusion, pp. 1-8, Jul. 2006.
[13] S.B. Colegrove, Advanced Jindalee Tracker: Probabilistic Data Association Multiple Model Initiation Filter, Defence Science and Technology
Organisation Technical Report No. DSTO-TR-0659, Jun. 1999. (http://www.dsto.defence.gov.au/publications/2121/DSTO-TR-0659.pdf).

View publication stats

You might also like