It
| to
ing
trol
> tO
ults
nce
line
)3).
n a
we
line
lem
on,
; in
1ity
fa
ent
95
107
of time ¢ — 1 is exactly assigned to one line segment of time t. In our approach we consider this constraint in a
restrictive manner. For example in case of splitting of line segments we also tolerate multiple assignments.
The similarity (d) of two line segments g? ! and 9j is determined based on a scoring function G, that measures
the similarity in the line parameters length, orientation and position (see also (McIntosh et al., 1988).
The similarity in length (b;) is determined as follows:
— mén(i^1, 1%)
1 maz(l‘”!, 1%) :
(1)
Here I!~! denotes the length of the line segment g*-l.
To vote the similarity in orientation (ba) the absolute difference between the orientation a”! and a; of the two
line segments have to be compared with a maximum difference ©;:
©; — [of^! — ot|
bp = —— 21, 2
2 0, (2)
The similarity of two line segments gi
and gj in position (bs) is defined by:
9s - |i - 8j|
£' denotes the predicted position Z of gi. The prediction is based on linear regression of measured displacements
in the past. The absolute difference between predicted and measured position must be compared to a maximum
difference (02).
The matching function G is given by the following weighted sum:
3 ;
ki
After determining the correspondences between consecutive frames, a displacement field is computed. A robust
analysis requires a significant displacement field. Therefore we connect the last 5 measured displacements for
each line segment to one displacement vector.
The following segmentation of the displacement field is done based on an algorithm to cluster analysis. Beginning
with one of the displacement vectors as center of the first cluster, we try to classify the other displacement vectors
to existing clusters. Therefore orientation, length and position of the displacement vectors are evaluated. If a
vector cannot be attached to one of the existing clusters, a new cluster will be generated. The segmentation
process is finished, when all displacement vectors are attached to clusters. Because of the sequential processing
of displacement vectors a further step is required, which merge similar clusters.
Each of the resulting regions contains a set of line segments moving in a similar way.
Now a decision is necessary, if the features of a region belong to background or possible vehicles. This is
done based on perceptual grouping strategies. We consider parallelism, orthogonality and proximity as basic
relationships.
The surrounding rectangles of features belonging to possible vehicles result in the attention fields.
5. RESULTS
Vision based driver assistance requires a processing in real time. MOSAIK was developed on a HP-9000 /735
workstation. On this station we also measure the computing time shown in Table 1. The measured times contain
all steps of recognition respectively attention control (not including the time to read a frame from harddisk).
In the case of integrated attention control a decrease of computing time is yield. This is also seen in the left
of Tab. 2. The edge detection (bottom) is a lower boundary of the computing time in MOSAIK. In case of
pure recognition (top) the maximal computing time is spent. The advantage of integrated attention control and
interaction between the states illustrates the curve in the middle. The changing between the levels of abstraction
in case of interaction is shown in the right of Tab. 2. Analogous to the perception of stimuli in humans, the
recognition can be initiated by attention control only after a period of 5 time frames. This mechanism avoids
a permanent initiation of the recognition by attention control.
The robustness of the attention control can be measured in the rate of false alarms. Today we yield about 16
%.
An example of interaction between recognition, tracking and attention control is shown in Fig. 4. The left
image in the first row illustrates the tracked vehicles at time t. The new incoming vehicle (middle, first row) is
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995
