![](bg5.png)
deviations of the measurement values from the shape,
the object state can be corrected.
First the position difference is calculated by the
mean perpendicular distance of all corresponding
measurement values to one side, as shown in figure
3. In this example, the expected shape is too far away
because the velocity of the object was underestimated.
Moving the shape by the mean of the pink, solid dif-
ference vectors leads to an average match of the mea-
surement points. Note that most measurement values
correspond to only one side and therefore only those
difference vectors to the corresponding side are ac-
counted for. By using the perpendicular distance con-
sequently, only motion perpendicular to the car’s side
is considered. Therefore the same procedure is used
on the other visible side to get the measured object po-
sition. Obviously this is problematic for the one-sided
I- and IS-shapes. Especially for IS-shapes, movement
parallel to the shape is expected and is therefore mea-
sured through the motion of the shape ends. However,
this is not as precise as the difference measurement
because the beam directions of the lidar scanner lead
to a discrete measurement accuracy perpendicular to
the beams. Compared to the I-, IS- and L-shapes, the
position update for C- and E-shapes is relatively easy.
If the measured values match the dimension thresh-
olds defined for pedestrians or bicycles, the calculated
center of gravity is used as the position.
Second, the orientation of the measured shape is
determined. In case of an I-, IS-, L- or E-shape the
orientation is calculated using regression lines fitted
into the measurement points (cf. (Lindl, 2008)). To
avoid corruption through short sides and round object
shapes, some additional characteristics, such as side
lengths and matching quality of the corner point, are
considered. Especially the determination of the orien-
tation of bicycles is often problematic. In such cases,
the orientation is not measured using the lidar points
but determined during the filtering process.
Finally, the dimensions of the objects are adapted.
Therefore the dropped perpendicular bases of the end
points are taken into consideration. These are deter-
mined anyway at the same time as the heading cal-
culation. As the lidar scanner cannot detect the com-
plete dimensions of an object in every measurement
step because of occlusion or deficient remission, the
measured dimensions vary over time. To overcome
these deviations, the dimensions could be filtered as
in (Fayad and Cherfaoui, 2007). However, to pro-
hibit the shrinking of an object over time when it is
only partly detected, we simply keep the maximum of
measured and previous dimensions like in (Fuersten-
berg et al., 2003).
3.2.5 Changes in Representing Shapes
An important phenomenon when tracking dynamic
objects with lidar measurements is that the detected
shape varies depending on the orientation and dis-
tance to the lidar scanner. Since objects are repre-
sented here using five characteristic I-, IS-, L-, C- and
E-shapes, it is important to detect when another shape
fits better. This enables optimal interpretation of the
measurement values. A good example for chang-
ing representations is an overtaking car, whose shape
changes from “IS” to “L” and finally to “I”. To detect
such transitions, special areas around these shapes are
defined, e.g. for the IS-shape the area where the sec-
ond side is expected. If measurement points exist in-
side these regions, the shape is switched. This also
holds for the C- and E-shape, which normally change
when the dimensions of the tracked object exceeds the
thresholds for pedestrians or bicycles.
3.3 Detection of New Objects
At the beginning of the tracking process or if an ob-
ject enters the surveillance zone of the sensor, there is
no previous information available. Therefore the pre-
sented approach has to distinguish between the track-
ing of known objects, as presented in the previous sec-
tion, and the detection of new objects.
The sequence of the object processing is known
from figure 2. First, all known objects are processed
and only the unassociated segments are passed onto
the detection step. Here, possible traffic participants
have to be identified. This is done using the common
approach (Kaempchen et al., 2005), (Lindl, 2008),
where the measurement data is searched to identify
characteristic shapes. We try to find segments which
match the introduced C-, I-, IS- and L-shapes. The
first is characterized by a small expansion of 0.9m at
most and possibly represents a pedestrian. The lat-
ter have to be approximated with one or two regres-
sion lines respectively and the regression error has to
be below a threshold. Finally the dimensions of the
shapes have to match the tolerated dimensions of traf-
fic participants shown in table 1. Note that the bi-
cycle shape “E” is missing for the creation of new
objects. Since this shape is very unspecific, moti-
vated by the varying measurements of bicyclists, there
are not enough special characteristics to search for.
To avoid numerous false detections, this shape has
been excluded and bicycle objects are created through
shape changes of dynamic objects.
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
798