=Paper=
{{Paper
|id=Vol-396/paper-3
|storemode=property
|title=A Framework for Mobile Intention Recognition in Spatially Structured Environments
|pdfUrl=https://ceur-ws.org/Vol-396/paper03kiefer.pdf
|volume=Vol-396
|dblpUrl=https://dblp.org/rec/conf/ki/KieferS08
}}
==A Framework for Mobile Intention Recognition in Spatially Structured Environments==
https://ceur-ws.org/Vol-396/paper03kiefer.pdf
A Framework
for Mobile Intention Recognition
in Spatially Structured Environments
Peter KIEFER and Klaus STEIN
{peter.kiefer,klaus.stein}@uni-bamberg.de
Laboratory for Semantic Information Technologies, University of Bamberg
Abstract. Mobile intention recognition is the problem of inferring an agent’s in-
tentions from the spatio-temporal behavior she shows. We present a framework for
mobile intention recognition that covers all levels, from low-level position data to
high-level intentions, with the detailed structure of the motion track as an inter-
mediate level. Our framework is new in explicitly including the structure of space
into the intention recognition process at each step. We demonstrate the benefit of
adding spatial knowledge by an ex-post interpretation of the behavior tracked in a
location-based game.
Keywords. Spatial cognition, plan recognition, mobile computing
‘There is a great deal of work in low-level perception, and a great deal in high level
recognition - including this thesis. The semantic gap between the output of the low-
level processes and the high-level inference engines remains wide, and few have
ventured to cross it.’ H. Kautz (1987) [12, outlook (p. 127)]
1. Introduction
Suppose you see a car showing a strange driving behavior. It might be traveling at a speed
far below average, slowing down at green traffic-lights, changing its direction frequently,
or stopping at the roadside several times. You probably infer that the driver is foreign
and trying to find his way to some destination. This problem of inferring an agent’s
intentions from her behavior is called intention recognition problem1 . Over many years,
plan recognition has mainly considered problems like language and story understanding,
or the design of intelligent user interfaces (see [7] for an overview).
Mobile intention recognition problems differ from these ‘traditional’ use cases, espe-
cially because mobile behavior happens in space and time. The location where a certain
behavior takes place tells us about possible intentions. The detailed (spatio-temporal)
characteristics of the motion track can also help. In the example above, an insecure driv-
ing style could be a sign for a foreign, drunk, or a novice driver. An experienced observer
(e. g. a police officer) could intuitively distinguish these cases by the motion track de-
1 Also known as plan recognition problem, [22]
28
tails. Automating this kind of inference requires several steps to bridge the gap between
low-level processing and high level intentions. One recent approach for mobile activity
recognition is presented by Liao et al [16]: they use a hierarchical Markov model with the
layers gps measurement, location on a street network, transportation mode, trip segment,
goal, and novelty. Due to the specific motivation of their paper, the semantics of ‘goal’ is
‘target location’ so there is no need for a more complex user intention model.
In this paper, we present a framework for mobile intention recognition that cov-
ers all steps from low-level motion track data to high level intention models. It utilizes
knowledge about the structure of space at each level of the processing hierarchy. A mod-
ular architecture allows to change the methods used for each step, thus making it easily
adaptable to different problem domains. The desktop application INTENSIVE, which
helps in testing and improving mobile intention recognition algorithms, complements
our framework.
In section 2, we explain how space changes the problem of intention recognition,
and introduce the location-based game CityPoker. Section 3 proposes our framework,
and explains each stage of the framework’s processing hierarchy. We show how the in-
tentions of a CityPoker player are recognized using her motion track, and spatial knowl-
edge about the gaming area (real game data, see Fig. 1). It is important to note that the
algorithms presented in section 3 for CityPoker are not claimed to be optimal for any
mobile intention recognition problem. The idea is rather that spatial knowledge can im-
prove intention recognition, independent from the concrete methods used. The intention
simulation environment (INTENSIVE) is outlined in section 4. We give related work in
section 5, and conclude in section 6.
2. Mobile intention recognition
In mobile computing we often find ourselves in situations with restricted possibilities
of human-computer interaction, due to a limitation of the user’s cognitive resources [2].
In these situations, a device should automatically select the information service the user
currently needs, and either present it directly (information-push), or ease the access to
that service (for instance by assigning it to a ‘hot button’).
Most location-based services directly deduce the information service needed from
the user’s current position (on-enter/on-leave paradigm). This direct mapping from raw
position data to information service does not work satisfactory in many situations. One
prominent example is the room-crossing problem [21]: a user wants to cross a room and
accidentally enters areas around POIs (e. g. exhibits in a museum room) triggering var-
ious information services not needed in that situation (annoying information push). By
mapping the user’s behavior to intentions, before mapping the intention to an appropriate
information-service, we can create an intention-aware information-service. In the mu-
seum example, such an intention-aware information service distinguishes room-crossing
and exhibition-visiting.
Implementing such a service for a mobile device is a challenging task. It is impor-
tant that our service performs well under real-time conditions. This is especially hard
when computing everything on the restricted resources of a mobile device. Server-based
architectures, as an alternative, may suffer from latency, data transfer costs, and network
unavailability. Everything needs to work incrementally, i. e. before we know any future
29
Figure 1. Spatial structure (regions and caches) and motion track of one team in CityPoker
behavior. Another assumption to be made is that we are dealing with keyhole intention
recognition [7], i. e. the user does not change his behavior to manipulate the intention
recognition process.
2.1. Spatial is special
We call an intention recognition problem a mobile one, if (at least some of) the behaviors
used for interpretation are spatio-temporal behavior. This can be any behavior derived
from the user’s position (with timestamp), or position history. The most basic kind of be-
havior are simple enter/leave-region behaviors. More complex types of spatio-temporal
behavior can be derived by motion track analysis if we analyze a number of sequential
motion points (a motion segment). As we will see in section 3, this may include several
steps of preprocessing (section 3).
Not only the positional data make mobile intention recognition special, but also the
spatial context2 . We often have knowledge about the spatial structure, e. g. a number of
points of interest (POI) with circular or polygonal areas around them [1,9]. Others add
a street network to these locations [16], use spatial tessellation [11], or formalize space
2 As most important form of context; we do not lead discussions about the general notion of context here.
30
Information Service
intention-aware
Information User
Database Intentions
non-spatial spatial
behaviour behaviour
other current motion
context pos. track
preprocessing
actions,
position data
events
Figure 2. System architecture.
with Spatial Conceptual Maps [19]. A hierarchical structure between POIs can be created
by ‘assigning to each location a number representing the level of detail that the location
belongs to’ [6, p. 13]. The next step is to use partonomies, i. e. a structure that stores
polygons in a tree with part-of relations between regions [24], or even structures allowing
overlapping regions. Partonomies are often found in the way how human structure space,
e. g. shop ⊂ shopping_district ⊂ city_district ⊂ city ⊂ county ⊂ . . .
The basic intuition of our framework is that not every’thing’ (every intention, be-
havior, action, ...) should be interpreted equal in any region. We will see in section 3
how this exactly works, and how we can make the intention recognition algorithm more
efficient by using spatial knowledge.
2.2. A location-based game: CityPoker
CityPoker, our example for the following section, is a location-based game played by two
adversary teams who try to optimize their poker hand by finding and changing hidden
cards. The game is supported by a mobile assistance system (J2ME on smartphones)
with GPS localization. It is usually played by bike (for a detailed game description see
[20]).
CityPoker imposes the following partonomial structure: the game board (e. g. a
city) contains five rectangular regions (REGION1 , . . . , REGION5 ), each of which contains
31
non-spatial behaviour
information
Segmentation
Classification
and Feature
recognition
extraction feat1 beh1 int1
Selection
Intention
service
feat2 beh2 int2
... ... ...
Spatial Structure
Figure 3. Stream of processing.
three circular regions (CACHE1,1 , . . . , CACHE5,3 ). Figure 1 displays only those regions the
player entered during the game.3
When a player (a team) enters a REGION, the mobile device poses a multiple-choice
quiz with three answers. Each answer will lead the player to one of the three caches.
Arriving at the cache, the device reveals a perceptual hint where to find the hidden cards
(e. g. ‘the cards are hidden under a green wooden object’). This kind of ‘fuzzy’ hint helps
the player to locate the exact place of the cards in the circular cache. The wrong answer
will lead the player to the wrong cache. After some searching time she may correct her
answer and head for another cache.
The information services available are maps of different zoom level (citymap, re-
gionmap, cachemap), perceptual hint, game status, quiz. In case the intention recogni-
tion mechanism guesses wrong, all services are also accessible for the user by manual
selection.
3. A framework for intention recognition in spatially structured environments
Figure 2 gives an overview on our framework. Positional data are the most important
input for mobile intention recognition. From that data we can access the current position
and use it in any processing step. Due to the memory limitations of mobile devices, it is
not possible to keep a complete motion track history. Instead, motion segments are clas-
sified, and kept as sequence of behaviors. Aditionally non-spatial behavior is included:
user actions (in CityPoker: reporting the server when the card has been found), and exter-
nal events (receiving a notification of the other team’s actions). Now user intentions are
inferred and finally the information service selects the suitable data from the database,
possibly parameterized by the current position and other context (e. g. game status). Fig-
ure 3 shows the (simplified) processing pipeline. The single steps are described in the
following sections using the CityPoker example.
3.1. Segmentation and feature extraction
The first steps in processing motion data are segmentation and feature extraction. De-
pending on the concrete environment motion data consists of a sequence of sensor data.
3 The track was recorded during a CityPoker game on Oct. 30, 2006, played by a team of 4 girls of age
between 10 and 14.
32
br
br b0 br
br b0
bcs br
br br
bcs
br br bc
bs
bs
b0
br
br
br
br
bc
Figure 4. Segmented motion track with classified behavior sequence in a cache region. (The player enters
from the right.)
This may include heading, speed, steering angle and so on. In the CityPoker example we
get positional information from a GPS device.4 Now the interesting features are extracted
from the motion data and the track is segmented accordingly. The first important feature
is the region of the partonomy we are in. The second feature interesting in our example
is the current speed. We distinguish low (< 1 m/s), medium (≤ 10 m/s) and high speed
(> 10 m/s), abbreviated as ls, ms, and hs in the following.5 The third feature is the curva-
ture of the motion. To keep things simple we used the average angle of direction change
in the last five points for measuring the curvature (low (lc), if (αk−4 + . . . + αk )/5 < 45◦ ,
high (hc) otherwise).6 The last feature is the maximum diameter of the convex hull of
the segment (i. e. the largest distance between two segment points).
The features used as well as the thresholds can be chosen dependent on the region
entered or left, e. g. while walking down a street other features may be interesting than
while being on a market place. Each change in one of the features (here: region, speed,
curvature) can trigger the start of a new segment (entering/leaving a region, speed drops
from medium to low, curvature gain, etc.). Every time a new segment starts the according
feature vector is passed to the next processing step: classification. Other features (here:
segment diameter) do not trigger a new segment but are subject to change as the user
moves on. In this case the feature vector of the current segment is updated and the change
is propagated to the classification.
4 A sequence of coordinates with timestamps (approx. time between two measures is 2 s). This allows to
compute direction and speed. Depending on data quality preprocessing may be necessary to catch sensor errors
and smoothing.
5 The values are specifically choosen for the example domain: 1 m/s is slow walk (due to data noise the GPS
measured speed nearly never drops to 0, even if the user does not move), and 10 m/s is very fast, at least for a
biker in the city.
6 For more sophisticated motion trach micro structure measures see [23, chapter 12].
33
Production Rules P Grounding
HandleRegion → SelectCache GotoCache GetCard REGION (1)
SelectCache → FindParkingPos SolveQuiz REGION (2)
FindParkingPos → br REGION (3)
SolveQuiz → b0 |bs |bsc REGION (4)
GotoCache → ApproachCache GotoCache | REGION (5)
OrientToCache GotoCache | ω
ApproachCache → (br | bc )+ REGION (6)
OrientToCache → (bs | bcs | b0 )+ REGION (7)
GetCard → SearchCard bChangeCard CACHE (8)
SearchCard → RoamCache SearchCard | CACHE (9)
DetailSearch SearchCard | ω
RoamCache → (br )+ CACHE (10)
DetailSearch → (bc | bs | bcs | b0 )+ CACHE (11)
Figure 5. SGIS production rules for CityPoker (a part of the complete grammar).
3.2. Classification
The classification step is mainly a mapping function from the feature vector to a given
set of user behaviours. This can be realized in various ways, from decision trees and table
lookup to fuzzy or probabilistic approaches. For the CityPoker example we have
{R1 . . . Rn }×{ls, ms, hs}×{lc, hc}×R + → {br , bc , bcs , bs , b0 }
(region, speed, curv., diam.) → behaviour
We distinguish
riding br speed = hs ∨ (speed = ms ∧ curv. = lc)
curving bc speed = ms ∧ curv. = hc
slowcurv. bcs speed = ls ∧ curv. = hc ∧ diam. > 5 m
sauntering bs speed = ls ∧ curv. = lc ∧ diam. > 5 m
standing b0 speed = ls ∧ diam. ≤ 5 m
As the gps data is noisy even a stillstanding subject will show a certain speed and curva-
ture, so the diameter of the region covered in the current segment is used to distinguish
different behaviours. This also means that the classification of the current segment may
change from b0 to bs or bcs with the user moving on. Certainly the classification could
also use the spatial structure, in our example we leave this to the next step. Figure 4
shows a part of the motion track from Figure 1, annotated with classified behaviors.
In some settings it can be efficient to combine segmentation, feature extraction and
classification in one step in the implementation to reduce overhead.
3.3. Intention recognition
The heart of our framework is the intention recognition mechanism itself. We choose
Spatially Grounded Intentional Systems (SGIS) [20] which are context free grammars
(CFG) with intentions as non-terminals, and behaviors as terminals. The applicability of
a production rule in an SGIS depends on the region of their respective terminal behaviors.
Grammars are, in general, intuitive to understand for the knowledge engineer.
34
HandleRegion
SelectCa GoToCa GetCard
FindPP SolveQ Approach SearchCard
DetailS SearchCard
DetailS SearchCard
Roam SearchCard
br b0 br bcs bc br ... ...
HandleRegion
SelectCa GoToCa GetCard
FindPP SolveQ Approach GoToCa ... ...
Orient GoToCa
Orient GoToCa
Approach ...
br b0 br bcs bc br ...
Figure 6. Parsing ambiguity if we had no spatial knowledge.
Figure 5 lists a number of SGIS rules for CityPoker. For reasons of clarity, we se-
lected only rules describing user intentions in a REGION: first, the player finds herself
a good place to stand, before she solves the quiz. She navigates towards her selected
CACHE by target-oriented moving (ApproachCache), interrupted by way-finding (Ori-
entToCache). These non-terminals are mapped to behaviors (rules 6, 7).7 In a similar
way, searching the card consists of (target-oriented) roaming through the CACHE, and
7 These rules are written simplified, and can easily be rewritten as context-free rules.
35
Production Rules P Grounding
GetCard → SearchCard AccLeave SearchCard REGION(12)
AccLeave → (something) REGION (13)
Figure 7. Modifying the CityPoker example to catch an accidental leaving behavior.
a (random appearing) detail search. Finally, our only non-spatial behavior (bChangeCard )
occurs when the players has succeeded in finding the card.
If the classification of the last segment (i. e. the last terminal) changes during pro-
cessing reparsing is necessary. By carefully choosing the production rules one can assure
that this will not change the active intention.
Figure 6 shows two possible parse trees for the start of the behavior sequence from
Fig. 4. Without spatial knowledge, both trees were possible. We cannot decide at which
position in the behavior stream we have to place the cut between GotoCache and Get-
Card, thus dealing with high ambiguity. SGIS reduce ambiguity by restraining the appli-
cability of rules to certain regions. Thus, the rule resolving GetCard is not applicable if
one of its terminals is outside the cache. With that additional information, the top tree is
not possible.
The grammar from Fig. 5 is simplified and does not catch all use cases, for instance,
if the player accidentally leaves the CACHE during searching (as she actually does in
Fig. 4). Also, an unsuccessful search (refine search to another CACHE) has not been
formalized.
The context-freeness of SGIS makes them more expressive than finite-state ma-
chines but some use cases require more context-sensitivity. Context-sensitive grammars
(CSG), on the other hand, are too complex to parse. Researchers have recently argued for
the use of mildly context-sensitive grammars (MCSG) for intention recognition [14,10].
MCSGs have been developed in the natural language processing (NLP) community and
fall in complexity between CFGs and CSGs. As one of their properties they allow for
cross-dependency relations between non-terminals in the parse tree. Constraints concern-
ing the applicability of production rules can be formulated for these dependencies. De-
pendencies with spatial constraints are the heart of Spatially Constrained Context-Free
Grammars (SCCFG), and Spatially Constrained Tree-Adjoining Grammars (SCTAG),
both introduced in [13].
In CityPoker, one such constraint arises if we try to catch the accidental leaving
behavior by adding the rules from Fig. 7. To formulate rule (12) correctly, we would
need to state that the first SearchCard must happen in the same CACHE like the second
SearchCard. In general, the production rule is applicable if the region R1 of intention I1
has a certain spatial relation r (e.g. identical, touches, contains, ...) to another region R2
where intention I2 happens. This kind of dependency cannot be formulated with SGIS,
but with SCCFGs. If our use case requires to cross such dependencies, SCTAGs are the
best choice.
3.4. Information service selection
The task here is to find an information service for the active intention. That is, the in-
tention closest to the current terminal behavior. For instance, in Fig. 6 at the second be-
havior, in both trees the intention SolveQ is the active one. Also, by climbing up the tree
36
to the more high-level intentions, we can easily create an ordered queue of information
services by selecting one service for each intention up to the root node.
Finally, finding the information service can be solved very simple by a mapping
mechanism, like SolveQ → quiz. We include space into the selection mechanism by
parameterizing the information service. Thus, the information service quiz will not just
show any quiz, but the one relevant in the current REGION.
3.5. Backpropagation
The processing as shown so far uses a strictly linear pipeline (Fig. 3). For many scenarios
this is most efficient regarding memory and time, but there are settings where information
from higher steps can help on lower steps. For example, the intention recognition reveals
which intention the user currently has (or may have). By propagating this information
back to the classification step it can be used to choose the most efficient classification
algorithm (e. g. if SearchCard applies a more sophisticated and therefore more expensive
classification is needed while otherwise a simple table lookup is sufficient).
As the classification step may give any arbitrary sequence of behaviours, the inten-
tion recognition production rules need to have some kind of “catch all”, i. e. at any time
there must be at least some rule applicable for any behaviour. One way to solve this is
to tell the classification step that the last behaviour is not applicable and should be re-
classified. This will certainly not work if classification is done with a deterministic ta-
ble lookup algorithm (as there is exactly one behaviour matching) but works fine with
probabilistic approaches, where different behaviours apply with different probability.
4. INTENSIVE: a simulation and testing environment
Our framework is flexible in four ways: (1) it can be used in different use cases by
choosing appropriate intentions, behaviors, and information services. (2) It allows to
change the algorithm chosen on each layer. (3) It allows to parameterize each algorithm.
(4) It allows to change the geographic model, i.e. to port the use case to a new area.
It is part of this research to identify the best configuration of algorithms and param-
eters ((2) and (3)) for use cases and spatial models of different complexity ((1) and (4)).
While some decisions on the configuration can be made analytically (e. g. regarding the
complexity of the formal grammar), others need to be made empirically using real mo-
tion track data. Optimally, we would like to test any possible configuration outdoors with
real users, and see if the information services proposed are perceived as helpful. This is
certainly not feasible due to the high effort associated with extensive outdoor testing.
The INTention SImulation enVironmEnt (INTENSIVE) addresses this problem.
With the desktop application INTENSIVE we can easily configure an intention recog-
nition mechanism (design time) and test it by playing back previously recorded motion
track data (runtime). At design time, the INTENSIVE user specifies a geographic model,
an intention processing model, and a motion track log file which contains a sequence of
latitude/longitude pairs with timestamp. At runtime, the system reads the log file, pro-
cesses the gps data according to the intention processing model, and outputs the behav-
iors, intentions, and information services that occur.
An INTENSIVE intention processing model consists of a number of algorithm
blocks which do the computations. Algorithm blocks have input and output ports over
37
Figure 8. Part of an intention processing model in INTENSIVE (preprocessing, feature extraction, and seg-
mentation).
Figure 9. Monitoring the agent in the geo visualization screen.
38
which they exchange information with other algorithm blocks. The user adds, connects,
and parameterizes algorithm blocks easily using the graphical user interface (Fig. 8). IN-
TENSIVE comes with a number of predefined algorithm blocks containing a set of im-
portant algorithms for each processing step. The plugin architecture of intensive allows
developers to implement new algorithms and add them to the system. Geographic mod-
els are imported from KML-files. During runtime, the user can watch his agent8 moving
in the geo visualization screen (Fig. 8). To speed up testing, it is possible to adjust the
speed of the runtime. By going through a cycle of manual model adjustment and testing,
the user hopefully reaches an intention processing model that perfectly handles the use
case.
At time of writing the INTENSIVE tool is still under development. The current
version of INTENSIVE targets at researchers and experienced application developers.
The idea is to try out configurations on the desktop and only implement the best ones
for a mobile device. Automatic code generation which converts the intention processing
model to code for the target platform will be part of future work.
5. Related Work
Besides Liao et al [16] (see section 1), Bui chooses another probabilistic network based
approach with several layers [5]. In both papers, computations are done on a server. Prob-
abilistic grammars were discussed in the areas of pattern recognition, machine vision,
and action recognition [3,17]. An overview on grammars in machine vision is given in
[8]. Probabilistic state dependent grammars [18] constrain the applicability of a rule de-
pendent on a general state variable. The generality of this state variable leads to an ex-
plosion in symbol space if trying to apply a parsing algorithm, so that an inference mech-
anism is chosen which translates the grammar into a Dynamic Bayes Network (DBN).
Geo-referenced DBN are proposed by [4] to fuse sensory data and cope with the problem
of inaccurate data.
Learning the locations relevant for a mobile user, instead of modeling them by hand,
is done in [1,16]. This can be reasonable in domains where the spatial structure is not
fixed. In CityPoker, we could learn the size of the circular caches. Also the steps of
our processing hierarchy could replace modeling by learning, e.g. the classification. We
could also try to detect the types of behavior interesting for a domain automatically with
methods from the spatio-temporal data mining community [15].
6. Conclusion and outlook
We have explained the characteristics of mobile intention recognition, and presented an
appropriate framework for including space in the intention recognition process. For the
intention recognition mechanism itself we chose a grammatical formalism, enhanced by
spatial knowledge. The analysis of a real CityPoker motion track has exemplified how
the spatio-temporal details of a trajectory can be exploited.
8 In general, INTENSIVE is also open to simulate team intentions by using several motion track logs. Log
files may also contain non-spatial behavior like interaction with the device.
39
In our future research, we will further explore the use of MCSGs for mobile intention
recognition. Different MCSG formalisms (e.g. Tree Adjoining Grammars, Head Driven
Phrase Structure Grammars) need to be considered for different use cases and tested
in INTENSIVE. INTENSIVE itself will be extended to cope with maps and motion
tracks from different sources and different code generation backends for different mobile
platforms.
An implementation on a mobile device will help to explore the feasibility of parsing
MCSGs on restricted resources. Interpreting behavior that happens in a spatial structures
with overlapping regions is another interesting issue.
References
[1] Daniel Ashbrook and Thad Starner. Using gps to learn significant locations and predict movement across
multiple users. Personal and Ubiquitous Computing, 7(5):275–286, 2003.
[2] J. Baus, A. Krueger, and W. Wahlster. A resource-adaptive mobile navigation system. In Proc. 7th
International Conference on Intelligent User Interfaces, pages 15–22, San Francisco, USA, 2002. ACM
Press.
[3] A.F. Bobick and Y.A. Ivanov. Action recognition using probabilistic parsing. In Proc. of the Conference
on Computer Vision and Pattern Recognition, pages 196–202, 1998.
[4] Boris Brandherm and Tim Schwartz. Geo referenced dynamic Bayesian networks for user positioning
on mobile systems. In Thomas Strang and Claudia Linnhoff-Popien, editors, Proceedings of the Inter-
national Workshop on Location- and Context-Awareness (LoCA), LNCS 3479, pages 223–234, Munich,
Germany, 2005. Springer-Verlag Berlin Heidelberg.
[5] Hung H. Bui. A general model for online probabilistic plan recognition. In Proceedings of the Interna-
tional Joint Conference on Artificial Intelligence (IJCAI), 2003.
[6] Hung Hai Bui, Svetha Venkatesh, and Geoff A. W. West. Tracking and surveillance in wide-area spatial
environments using the abstract hidden markov model. IJPRAI, 15(1):177–195, 2001.
[7] Sandra Carberry. Techniques for plan recognition. User Modeling and User-Adapted Interaction, 11(1-
2):31–48, 2001.
[8] Gaurav Chanda and Frank Dellaert. Grammatical methods in computer vision: An overview. Technical
Report GIT-GVU-04-29, College of Computing, Georgia Institute of Technology, Atlanta, GA, USA,
November 2004. ftp://ftp.cc.gatech.edu/pub/gvu/tr/2004/04-29.pdf.
[9] H.M. Dee and D.C. Hogg. Detecting inexplicable behaviour. In Proceedings of the British Machine
Vision Conference, pages 477–486. The British Machine Vision Association, 2004.
[10] Christopher W. Geib and Mark Steedman. On natural language processing and plan recognition. In
Proceedings of the 20th International Joint Conference on Artificial Intelligence (IJCAI), pages 1612–
1617, 2007.
[11] Björn Gottfried and Jörn Witte. Representing spatial activities by spatially contextualised motion pat-
terns. In LNAI 4434 G. Lakemeyer et al., editor, RoboCup 2007, International Symposium,, pages
329–336. Springer, 2007.
[12] Henry A. Kautz. A Formal Theory of Plan Recognition. PhD thesis, University of Rochester, Rochester,
NY, 1987.
[13] Peter Kiefer. Spatially constrained grammars for mobile intention recognition. In Christian Freksa et al.,
editor, Spatial Cognition VI, Lecture Notes in Computer Science. Springer, 2008. accepted.
[14] Peter Kiefer and Christoph Schlieder. Exploring context-sensitivity in spatial intention recognition. In
1st Workshop on Behavior Monitoring and Interpretation, KI 2007, 2007.
[15] Patrick Laube, Marc van Krefeld, and Stephan Imfeld. Finding remo - detecting relative motion patterns
in geospatial lifelines. In Developments in Spatial Data Handling, Proceedings of the 11th International
Symposium on Spatial Data Handling, pages 201–215, 2004.
[16] Lin Liao, Donald J. Patterson, Dieter Fox, and Henry Kautz. Learning and inferring transportation
routines. Artificial Intelligence, 171(5-6):311–331, 2007.
[17] Dimitrios Lymberopoulos, Abhijit S. Ogale, Andreas Savvides, and Yiannis Aloimonos. A sensory
grammar for inferring behaviors in sensor networks. In IPSN ’06: Proceedings of the fifth international
40
conference on Information processing in sensor networks, pages 251–259, New York, NY, USA, 2006.
ACM Press.
[18] David V. Pynadath and Michael P. Wellman. Probabilistic state-dependent grammars for plan recog-
nition. In Proceedings of the 16th Annual Conference on Uncertainty in Artificial Intelligence, pages
507–514, 2000.
[19] Nancy Samaan and Ahmed Karmouch. A mobility prediction architecture based on contextual knowl-
edge and spatial conceptual maps. IEEE Transactions on Mobile Computing, 4(6):537–551, 2005.
[20] Christoph Schlieder. Representing the meaning of spatial behavior by spatially grounded intentional
systems. In GeoSpatial Semantics, First International Conference, volume 3799 of Lecture Notes in
Computer Science, pages 30–44. Springer, 2005.
[21] Christoph Schlieder and Anke Werner. Interpretation of intentional behavior in spatial partonomies. In
Spatial Cognition III, Routes and Navigation, Human Memory and Learning, Spatial Representation
and Spatial Learning, volume 2685 of Lecture Notes in Computer Science, pages 401–414. Springer,
2003.
[22] C.F. Schmidt, N.S. Sridharan, and J.L. Goodson. The plan recognition problem: An intersection of
psychology and artificial intelligence. Artificial Intelligence, 11(1-2):45–83, 1978.
[23] Klaus Stein. Qualitative Repräsentation und Generalisierung von Bewegungsverläufen. PhD thesis,
Institut für Informatik der Technischen Universität München, 2003.
[24] Klaus Stein and Christoph Schlieder. Recognition of intentional behavior in spatial partonomies. In
ECAI 2004 Worskhop 15: Spatial and Temporal Reasoning (16th European Conference on Artificial
Intelligence), 2005.
41