Instrumentation & Measurement Magazine 25-1 - 42

Fig. 7. A sample LiDAR measurements from an extended target with false alarms.
on the sensor field of view
by distributions like Poisson
point process [13]. 3)
The shape of the extended
target is modelled into
simple or complex geometrical
shapes like sticks,
rectangle, ellipse, etc. [14].
The data association is a
very important aspect in
all of the extended target
tracking algorithms. In
Probabilistic Multiple Hypothesis
tracker (PMHT), it
termination. Hence, we may end up with many false tracks or
missed tracks.
Point vs. Extended Target
Theoretically, high-resolution sensors should provide better
tracking performance. However, it brings additional
challenges to the tracking algorithms, like increased computational
power, etc., and sometimes it may provide bad tracking
performance if all of the issues are not addressed properly or a
simpler approach is taken to deal with the computational complexity.
Most of the tracking algorithms consider the targets
as a point target, i.e., at most, one measurement is generated
for each target. With lower resolution sensors, only one measurement
is generated from the target, which makes it a point
target. However, with higher resolution sensors like radars,
LiDAR or cameras, multiple measurements can be generated
from various parts of a single target. Tracking a target with
multiple detections is called extended target tracking. Fig. 7
shows an extended target with LiDAR measurements from
various shape points. Green dots indicate the measurements,
and the red rectangle indicates a track formed with these
measurements.
It is not possible to estimate the size of a target when a target
appears as a point target for a sensor. Note that 'point' is
defined for sensor-target pair, not for the target. A target can
be a point for one sensor and be extended for another sensor.
With high-resolution sensors, it is easier to estimate the
size of a target and its class. However, only a part of the target
may be visible in one scan due to self-occlusion. Hence, detections
over multiple scans need to be considered to estimate the
whole shape.
The major challenge with the extended targets is the data association
of the measurements to the target while ignoring the
false alarms and outliers. If the scenario has multiple, possibly
closely spaced, extended targets, the data association becomes
even more difficult. The extended target tracking problem is
addressed in three major paths: 1) A preprocessing algorithm is
used to group the multiple measurements into one single measurement.
Averaging is one of the methods among many. With
this technique, the problem is simplified, but it loses important
information like size, shape and orientation. 2) By using
the random finite set concept, the measurements are modelled
42
assumed that multiple measurements from one target are possible
[15]. In [16], a PMHT based approach is proposed to track
extended targets by considering the targets as rectangles. The
probability hypothesis density (PHD) filter can keep a count
of the number of tracks in a multitarget scenario and can track
multiple target states without the data association between
measurement and tracks, but a common PHD filter assumes a
single measurement from each target.
Fusing information from multiple high-resolution sensors
optimally is an open problem. Generating tracks from each
sensor's measurements separately and fusing the extended
target tracks is a feasible, suboptimal approach.
Multipath
Signals coming from a target sometimes bounce from a nearby
surface before reach the sensor. In this case, the measured
range or angle of the target can be far off from the actual values.
This problem is called multipath problem. The multipath problem
can be classified into two categories: unresolved multipath
and resolved multipath. In an unresolved multipath case, we
will get only one measurement from a target, but the measurement
values might have a bias. In the resolved multipath, we
will get more than one measurement from a target. Note that
this is different from the extended target tracking problem,
even though we get multiple detections from a target.
The signal from a low altitude target (aircraft or missile)
might have multipath due to the reflection from the sea
surface. Fig. 8 shows four paths of the signals [6]. This is an example
of an unresolved multipath problem. These multipath
measurements induce the problem of erroneous elevation angle
estimation. To address this problem, first, an assumption
is considered that the average signal power is known and follows
Rayleigh distribution. The sum signal and the difference
signal of the in-phase and the quadrature components are calculated
using all the echos of the multipath, and a Maximum
Likelihood-based angle extractor algorithm is proposed in [6].
This problem is again extended for unknown signal strength
where an estimator of the average power of the signal is used
in angle extractor algorithm.
Another multipath problem is the appearance of ghost
tracks in a complex urban environment. Autonomous vehicles
and other assisted driving systems use LiDAR and radar
IEEE Instrumentation & Measurement Magazine
February 2022
Instrumentation & Measurement Magazine 25-1

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-1
