Stefan Growe
Industrial Fairground
Industrial Area : nme a [1] d: [1,1]
P: 0.1 P: 0.2 :0,
[3,00] [2,00]
Fairlnactivity FairConstruction FairActivity FairDismantling
: d: [5, 10] d: [1,11] a: [58] d: [1, 1] d: [2, 5]
to: [0, oo] : 0. to: [0, co] P: 0.9 to: [0, oo] P: 0.8 to: [0, 00]
Hall Parking Lot P: 0.49 P: 0.16 P: 0.19
dH, 1]
P: 0.9
P: 0.16
No Trucks Empty Trucks Filled Persons on
near Halls Parking Lots near Halls Parking Lots Fairground
*t—- part-of *"-— con-of «isa e temp-—rel Time specifications in days
Figure 3: Semantic net for the detection of an industrial fairground with integrated state transition graph.
The relative proportions are sufficient to enable the system to favour the more probable solutions during analysis (see
Chapter 4). For a consistent network it has to be considered, that all state probabilities and the transition probabilities of
each state sum up to 1.
3.2 Use of Temporal Knowledge
During interpretation the temporal knowledge is used to predict successor states for already detected objects. Knowing
the amount of elapsed time A# = t, — t, between two consecutive images the state transition diagram is exploited to
determine possible states of an object in 7, based on its state in /,. To avoid, that a possible candidate is omitted accidently,
the prediction is done in an optimistic way. Starting in the state s; detected for the time-stamp 7, possible successor states
Is) are determined according to the defined transitions. For each candidate s; the earliest start fnr is calculated from the
minimum transition time qnin (Eq. (3)). The latest end te is defined by the maximum remaining time d/^' of state s; ,
the maximum transition time d;"", and the maximum duration d7"* of state s;. (Eq. (4)). State s; is a possible successor
of s; if the current time-stamp f, lies inside the interval [^ t^] and additionally inside to," to, *dj"], i.e. the inter-
val of the permitted occurrence of s; The cases, that a state stays the same or a possible successor is skipped due to a large
time difference, is considered also.
m" - qd (3)
m" ty + d+ I rex (4)
Assuming, that for example the state Fairlnactivity was detected for a region in an image dated March 1, possible succes-
sor states for this region at March 10 are (according to the temporal knowledge in Figure 3): Fairlnactivity, if the state
stays unchanged, FairConstruction, due to its maximum duration of 10 days, and FairActivity, if the construction phase
was shorter than 8 days and therefore not observed. The state FairDismantling can not be reached within 9 days. The sys-
tem AIDA generates hypotheses for each alternative and tries to verify them in a model-driven manner. For the decision,
which solution is investigated first, the competing hypotheses have to be judged and compared. As mentioned in Chapter
2.2.1 the evidence introduced at this stage of interpretation is equal for all alternatives, they differ in new model-driven
346 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
