Stefan Growe
fused using Dempster’s rule of combination. For a more detailed description see (Tönjes, 1999 and 1999b). All node
judgements of the semantic net are subsumed to an overall judgement of the scene description represented by the current
node of the search tree. The possibility of such a search node defines an optimistic estimation of the interpretation quality.
It is used by the system control to decide which alternative is investigated next.
The possibilistic approach supports the integration of uncertainty and imprecision within the interpretation process. The
evidence found in the sensor data is aggregated strictly bottom-up to a final judgement of the scene description. Conse-
quently a set of competing interpretations, that differ solely in model-driven hypotheses and possess therefore the same
evidence, obtain identical values of merit. Hence, the candidate for the further interpretation has to be selected randomly.
But in many cases an expert has prior knowledge which alternative is more probable. A plausible strategy would be to
prefer the most probable alternative. For this reason in Chapter 4 a new judgement calculus based on Bayesian networks
is presented which considers those prior known probabilities.
3 - MULTITEMPORAL ANALYSIS
Applications like environmental monitoring and change detection require the evaluation of images from different acquisi-
tion times. Multitemporal images are also needed for the recognition of complex patterns, that are characterized by a typi-
cal temporal behaviour. Change detection can be carried out at pixel or at object level. Approaches that detect differences
at pixel level require a perfect co-registration of the data sets. Furthermore they are limited to the comparison of images
from the same sensor platform and they are very sensitive to variations in illumination, weather conditions, and perspec-
tive of view. Here, the recognition of changed object semantics is aspired. The scene description derived from the preced-
ing image is used as prior knowledge for the interpretation of the current image. The easiest way to generate a prediction
for the current image from an existing scene interpretation is to assume, that nothing has changed during the elapsed time.
The awarenesses of the last interpretation are transformed unchanged into model-driven hypotheses to guide the analysis
of the current image. But in many cases more reliable hypotheses can be generated, if additional temporal knowledge is
used. Assuming biannual observations, a construction site as an example will probably not be observed at the same loca-
tion again, because the construction has been finished meanwhile. To take advantage of temporal knowledge, it has to
be represented appropriately in the knowledge base so that it can be exploited automatically during analysis.
3.1 Representation of Temporal Knowledge
In this scope temporal knowledge is understood as the knowledge about possible (or probable) transitions between differ-
ent object classes over time. It is represented in a state transition graph which is integrated seamlessly within the semantic
net. The states itself are modelled by concept nodes, the state transitions are defined by a new relation called temporal
relation, which describes the temporal order of the states. For each state s; both, a relative duration d; and an absolute
starting date fy, can be specified. To consider the uncertainty of the knowledge, time intervals are used for all temporal
declarations. It is possible to define a prior probability P(s;) for each state, which represents the relative frequency of its
occurrence. By default the probabilities are postulated to be equally distributed. For each state transition connecting two
states s; and s, its duration d; and its conditional probability P(sls;) can be defined also. Temporal relations are established
exclusively between objects with a symbolic meaning, because no general statements about temporal changes of geomet-
ric objects, for example, can be made. In contrast to hierarchical relations like part—of and con—of, the start and end node
of a temporal relation may be identical — forming a loop — to represent that the state stays unchanged over time. The men-
tioned representation scheme combines aspects of classical state transition graphs and markov chains known from the
system theory, temporal constraint networks (Dechter et al., 1991) used in Al, and planning or activity networks known
from the operations research field.
Figure 3 shows a semantic net for the detection of an industrial fairground. From a single aerial image a number of halls
and parking lots can be recognized, sufficient to classify the site as an industrial area. For the decision, whether it is a
fairground or not, a complete cycle of inactivity, construction of booths, fair activity, and dismantling of booths has to
be observed in a sequence of multitemporal images. The mentioned cycle is represented in the semantic net. Each of the
four states can be recognized by specific features. During the construction and dismantling phase the parking lots are
empty, but there are a lot of trucks next to the halls, that keep the equipment for the booths. During a fair the fairground
itself is free of cars, but the parking lots are crowded. If the image resolution is high enough, visitors can be detected walk -
ing on the fairground. If no fair takes place, there are few or even no cars neither on the parking lots nor on the fairground.
Typical durations are defined for the states: The inactivity might last the whole year represented by the interval [1, 365].
For the other phases durations of 2 to 10 days are assumed. Absolute starting dates are not defined, because there are no
seasonal restrictions for fairs. The state transitions in this example can occur from one day to the next. Hence their dura-
tions are specified by one day. For the states and state transitions prior probabilities are estimated due to general experi-
ences. The inactivity state is the most probable of the four. The conditional probabilities express that this state is stable,
while the others are more or less transient. An absolute precision of the probabilities is not necessary (even not possible).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 345
