`
`Young Francis Day, Serhan DagtaS, Mitsutoshi Iino
`Ashfaq Khokhar, and Arif Ghafoor
`Distributed Multimedia Systems Laboratory
`School of Electrical Engineering
`Purdue University, West Lafayette, IN 47907
`
`Abstract
`
`In this paper, we propose a graphical data model
`for specifying spatio-temporal semantics of video data.
`The proposed model segments a video clip into sub-
`segments consisting of objects. Each object is de-
`tected and recognized, and the relevant information of
`each object is recorded. The motions of objects are
`modeled through their relative spatial relationships as
`time evolves. Based on the semantics provided b y this
`model, a user can create his/her own object-oriented
`view of the video database. Using ihe propositional
`logic, we describe a methodology f o r specifying concep-
`tual queries involving spatio-temporal semantics and
`expressing views f o r retrieving various video clips. Al-
`ternatively, a user can sketch the query, by examplify-
`ing the concept. The proposed methodology can be used
`to specify spatio-temporal concepts at various levels of
`information granularity.
`
`1 Introduction
`
`The key characteristic of video data is the spa-
`tial/temporal semantics associated with it, making
`video data quite different from other type of data such
`as text, voice and images. A user of video database can
`generate queries containing both temporal and spatial
`concepts. However, considerable semantic heterogene-
`ity may exist among users of such data due to differ-
`ence in their pre-conceived interpretation or intended
`use of the information given in a video clip. Seman-
`tic heterogeneity has been a difficult problem for con-
`ventional database [3], and even today this problem
`is not clearly understood. Consequently, providing a
`comprehensive interpretation of video data is a much
`more complex problem.
`
`'This research was supported in part by the National Science
`Foundation under grant number 9418 767-EEC
`
`Most of the existing video database systems ei-
`ther employ image processing techniques for indexing
`of video data or use traditional database approaches
`based on keywords or annotated textual descriptors.
`For indexing, keywords and textual descriptions for
`video data have also been suggested in [6], based on
`a generalization hierarchy in object-oriented realm.
`Video segments can be joined or concatenated based
`on the semantics of this hierarchy. However, this ap-
`proach is very tedious since the perception of video
`contents is done manually by users, not through an
`automatic image processing mechanism. A video sys-
`tem that automatically parses video data into scenes
`using a color histogram comparison routine have been
`proposed in [5]. To locate frames containing desired
`objects, a method based on comparing the color his-
`In [4], a hierar-
`togram maps of objects is used.
`chical video stream model is proposed, that uses a
`template or histogram matching technique to identify
`scene change in a video segment. A video segment is
`thus divided into several subsegments. In each sub-
`segment, a frame-based model is used to index the
`beginning frame of a subsegment from which objects
`are identified. In this system a video stream is parsed,
`and the information is stored in the database. How-
`ever, this system has many limitations. First, the tex-
`tual descriptions are manually associated with video
`segments. Second, t,he system provides parsing only
`for a specific type of video.
`In order to address the issues related to user-
`independent view and semantic heterogeneity, we pro-
`pose a semant,ically unbiased abstraction for video
`data using a directed graph model. The model al-
`lows t,o represent spatic-t,emporal aspects of informa-
`tion associated with objects (persons, buildings, vehi-
`cles, etc.) present in video data. However, such an au-
`tomated video data-base system requires an effective
`and robust recognition of objects present in the video
`database. Due to the diverse nature of the video data,
`we can use various t,echniques currently available ac-
`
`1063-6382/95 $4.00 0 1995 IEEE
`
`40 1
`
`
`
`VSDGCompU.IMobl +5?
`
`b
`
`I
`
`k W v i 0 Data
`Figure 1: System abstraction
`cording to the requirements of different situations that
`may occur in the input. For each input video clip,
`using a database of known objects, we first identify
`the corresponding objects, their sizes and locations,
`their relative positions and movements, and then en-
`code this information in the proposed graphical model.
`The encoded video data may be semantically highly
`rich. Therefore a unified framework is needed for the
`users to express and for the system to process seman-
`tically heterogeneous queries on the unbiased encoded
`data. For this purpose, we propose an object-oriented
`approach that provides an elegant paradigm for rep-
`resenting user’s view of the data. It is a hierarchical
`scheme that provides the necessary framework for a
`user to compose the views of the data with the maxi-
`mum flexibility and at the same time allows processing
`of heterogeneous queries by evaluating the proposed
`graphical abstraction. For this purpose we also de-
`fine an interface between these modeling paradigms.
`Another reason for this modeling approach is to pro-
`vide an efficient indexing mechanism for on-line query
`processing without performing computations on t,he
`raw video data since such computation can be quite
`extensive. The proposed VSDG can be generated off-
`line and subsequently can be used to process user’s
`queries on-line. The architecture of the proposed sys-
`tem is shown in Figure 1.
`The organization of this paper is as follows. Sec-
`tion 2 describes the assumptions, techniques and the
`proposed graphical model. Two examples of video
`clips are used to illustrate the model. In Section 3,
`an object-oriented approach is presented for users to
`specify the perceived view of video data. The use of
`predicate logic for specifying spatio-temporal concepts
`and query types are also presented there. Section 4 de-
`scribes the computer vision/image processing require-
`ments of the proposed model. The paper is concluded
`in Section 5 .
`
`cal objects in the course of time and their relation-
`ship in space. Physical objects may include persons,
`buildings, vehicles, etc. A video database, is a typ-
`ical replica of this wordly environment. In concep-
`tual modeling of video data, it is therefore important
`that we identify physical objects and their relation-
`ship in time and space. Subsequently, we can represent
`these relations in a suitable structure that is useful for
`users to manipulate. Temporal relations among ob-
`jects has been previously modeled by using Petri-Net
`[l], temporal-interval [2], languages, etc. For spatial
`relations, most of the techniques are based on project-
`ing objects on a two or three dimensional coordinate
`system. Very little attempt has been made to for-
`mally express spatio-temporal interactions of objects
`in a single framework. In this section, we propose a
`graphical model to capture both spatial and tempo-
`ral semantics of video data. The proposed model is
`referred as Video Semantic Directed Graph (VSDG)
`model. The most important features of the VSDG
`model is an unbiased representation of the information
`while providing a reference framework for construct-
`ing semantically heterogeneous user’s view of the video
`data. Using this model along with the object-oriented
`hierarchy (discussed in Section 3) the system shown in
`Figure 1 can be used for on-line query processing, with
`performing any computation on the actual raw video
`data. In the following sections, modeling of space and
`time informatlion associated with objects is described
`in detail.
`
`2.1 Spatio-Temporal Modeling over a Se-
`quence of Frames (a Clip)
`
`The spatial attrihute, of a salient physical object
`present in the frames can be extracted in form of
`bounding volume, 2 , that describes the spatial pro-
`jection of an object, in three dimensions. It is a func-
`tion of bounding rectangular ( Y ) , centroid, and depth
`information related to the object. The bounding rect-
`angular is computed with reference to a coordinate
`system with an origin at the lower left corner of each
`frame. Both 2 , and Y are expressed as :
`
`Bounding Rectangular ( Y ) = (width, height, 2 , y)
`
`Bounding Volume (2) =
`
`2 Graph-Based Conceptual Modeling
`
`Generally, most of the wordly knowledge can he
`expressed by describing the interplay among physi-
`
`(Bounding Rectangular, centroid, depth)
`
`Temporal modeling of a video clip is crucial for
`users to constriict, views or to describe episode/events
`in the clip. Episodes can be expressed by collectively
`
`402
`
`
`
`interpreting the behavior of physical objects. The be-
`havior can be described by observing the total dura-
`tion an object appears in a given video clip and its
`relative movement over all the frames. For example,
`occurrence of a slam-dunk in a sport video clip can be
`an episode in a users specified query. The processing
`of this query requires evaluation of both spatial and
`temporal aspects of various objects.
`Temporal information of objects can be captured
`by specifying the changes in the spatial parameters
`associated with the bounding volume (2) of objects
`over the sequence of frames. At the finest level, these
`changes can be recorded at each frame. Although,
`this fine-grained temporal specification may be desir-
`able for frame based indexing of video data, it may
`not be required in most of the applications. The over-
`head associated with such detailed specification may
`be formidable. Alternatively, a coarse-grained tempo-
`ral specification can be maintained by only analyzing
`frames at 6 distance apart [7]. This skip distance (6)
`may depend upon the complexity of episodes. 6 is
`an interger with frame as its unit. There is an ob-
`vious tradeoff between the amount of storage needed
`for temporal specification and the detail of informa-
`tion maintained by the VSDG model.
`
`2.2 The Proposed Model
`
`Formally, both the spatial and temporal specifica-
`tions of a clip can be represented as a directed graph,
`as shown in Figure 2, that consists of n video seg-
`ments, labeled V I , V - , . . ., V,.
`In this model, time
`spans of physical objects are represented as circular
`nodes in the graph. Salient objects within a segment
`are grouped between two rectangular nodes, whereas
`such a node marks the appearance of a new physical
`object. Within a segment there can be an arbitrary
`number of objects. A video clip may consist of several
`such segments. An object may appear in any number
`of segments. In order to capture temporal semantics
`of object via VSDG, a motion vector can be associated
`with each object, represented by a circular node. The
`formal definition of VSDG is following.
`A VSDG is a directed graph with a set of circu-
`lar nodes (P), a set of rectangular nodes (T), and a
`set of arcs (A) connecting circular nodes to rectangu-
`lar nodes. A circular node has an attribute describing
`the duration for which an object (person, vehicle, etc.)
`appears in a video segment. A rectangular node cor-
`responds to an event in a video clip whenever a new
`physical object appears. In other words, a rectangu-
`lar node marks the start of a new segment, that differs
`from its predecessor segment in terms of appearance of
`
`Figure 2: VSDG representation of a clip
`
`a new physical object. Each circular node has exactly
`one incoming arc and one outgoing arc. The dura-
`is max(duration(O;l), . . .,
`tion of a video segment
`duration(O;,,)), where m; is the number of salient
`objects Oij (1 5 j 5 m i ) appearing in segment V;.
`Each circular node has the following attributes:
`D:P + I,
`is a mapping from t,he set of circular nodes to the set
`of durations in terms of frames.
`0 W : P + (21,. . .,z[;j),
`is a mapping from the set, of circular nodes to the set
`of motion vetors. The element Z; (Vi (1 5 i 5 [:I))
`of t,his vector is the Boiinding Volume at i-2h sampled
`frame. In other words, 2; = (Bounding Boxi, depth;,
`centroid;), and Bounding Box; = (width;, heighti ,
`z;, y i ) , where 1 represents the number of frames as-
`sociated with object. Oi in a given video segment and
`6 is tlhe time granularity for tracking motion of every
`object in a video segment.
`
`From t,he motion vectors, we can derive the rela-
`tive spatial relationship between any two objects for
`any sampled frame. The relative movements between
`two objects can be described as a set of mutual spa-
`tial relationships, a3 in [$]. For any sampled frame,
`the relative positmion between objects 0; and Oj can
`be captured by the spatial relationship between their
`projections on each coordinate axis, I , y, and z . In
`other words, for O;, R;j = ( S R ; j = , SR;j,, S R i j z ) rep-
`resents the spatial relat,ionship of projections of Oj
`wit,h respect, to 0, on each axis. If there are IC concur-
`rent objects, between t,wo rectangular nodes, then IC -
`1 such vectors are generated for each Oi. Therefore,
`in a given segment of a VSDG, if an object 0, appears
`concurrently with 0, for 1 frames, then 0; has a vector
`M;j = (R;j, , . . . , R;j L6, ). Similarly, Oj has a vect,or
`Mji = ( R j i , , . . . , Rji
`). To determine the speed of
`L i J .
`an object,, we t,ake t,he difference Z, - 2 0 and divide it
`For O;, vectors R and M are derived infor-
`by [:J6.
`mation and are additional attributes associated with
`circular node represent,ing 0;.
`
`403
`
`
`
`3 A VSDG-Based Object-Oriented
`Model
`
`As mentioned earlier, the objective of the system,
`shown in Figure 1, is to process video queries on-line.
`We, therefore, need to represent the rich semantics of
`video data using a suitable data model. In this sec-
`tion, we propose an object-oriented model which can
`be interfaced with the VSDG model. The objective
`of proposing an object-oriented paradigm is two folds.
`First, this can be used by the users to define their
`conceptrial queries involving motion in a systematic
`way. Second, it can allow processing of user’s con-
`ceptual queries by evaluating VSDG. Therefore, the
`system processes users’ queries with the assistance of
`the object-oriented views. In other words, an object-
`oriented view serves as a ”knowledge base” for the sys-
`tem. In Section 3.1, we describe how a user can con-
`struct. views about. the video data consisting of various
`objects.
`
`3.1 An Object-Oriented Based User’s De-
`fined View
`
`As we have mentioned video data is represented
`by three entities, spatial, temporal and physical ob-
`jects. For user to query video data, we propose an
`object-oriented approach which provides an elegant
`paradigm for user’s view representation. It is a hi-
`erarchical schema that. allows a user to compose their
`views about, the video data. The objective is to offer
`the maximum flexibility to a user to represent their
`semantics and at the same time allow processing of
`heterogeneous query by executing the proposed VSDG
`model. Generally, in any worldly knowledge three en-
`tities of interest are: spatial, temporal, and physical.
`The spatial entity called Conceptual Spatial Object
`(CSO) is the spatial concept associated with an object
`which can be extract,ed at the frame level. For exam-
`ple, Presiding a meeting attaches a meaning to some
`spatial area. For this concept, a person in a frame
`may be itlent,ified such that he/she is either stand-
`ing or sitting on a chair in the center of a meeting
`room. Another example of a spatial concept is ‘sit-
`ting’. A person may be sitting on a chair or some
`physical object. In this case, we have a conceptual
`spatial object, ‘sitting’ with attributes ‘a physical ob-
`ject which can sit.’ and ‘a physical object being sit
`on’, and they are relat.etl by the ‘sitting’ relationship.
`Conceptwil Temporal Object (CTO) defines the con-
`cept t<hat extends over a sequence of frames. It may
`involve several CSOs or CTOs (may be combination
`of both). There can be several temporal relationships
`
`Figure 3: Snapshots of two example video clips
`
`2.3 An Example of VSDG-Based Model-
`ing
`
`In this section, we use a video clip shown in Figure
`3 to illustrate the proposed model. In the example
`video clip (Figure 3(a)), a car (object 2) and a per-
`son (object 1) appear first, then the camera moves
`toward the right and two persons (object 1 and object
`5) are walking toward each other and shake hands.
`Assuming that proper object recognition methods are
`used to identify these objects, we can appropriately
`define the bounding volumes information for the ob-
`jects. The complete VSDG model, for the example
`video clip is given in Figure 4, which describes the
`information about various objects and their temporal
`behaviors. The VSDG in Figure 4, has four rectan-
`gular nodes which correspond to three different scene
`changes. The first rectangular node ( t o ) marks the
`start of video clip, tl indicates the appearance of ob-
`jects 0 5 , t 2 indicates the appearance of object OS, and
`t 3 indicates the end of the video clip. There are a to-
`tal of six objects, 01, 0 2 , 0 3 , 04, 0 5 , OS, and some
`objects appear in multiple scenes. For example, 01,
`and V,.
`0 2 , 0 3 , and 0 4 appear in video segments
`
`404
`
`
`
`Figure 5: Fan’s view
`Michael Jordan (i.e., a PO) has a slam-dunk (CSO +
`CTO)’. The system first finds the image of Michael
`Jordan from the picture attribute of player class, and
`selects a set of video clips, then the slamdunk CTO
`is involved with its inputs supplied by the places in
`VSDG of selected video clips. On the other hand,
`a coach would like to identify those video segments
`where pass (CTO) occurs around the right sideline of
`the front court. The system invokes methods associ-
`ated with a CTO called pass and with a CSO called
`right sideline and uses them to evaluate VSDG.
`The definit,ions of some of the classes used in this
`example are given in Table 2. The entries are self-
`explanatory. The methods are coded based on predi-
`cate logic and describe the spatio-temporal processing
`related to the event of that object. In class NBA,
`”Setof’ is iised t,o specify t,he association abstraction.
`The method ”slamdunk” will be defined later.
`
`3.2 Predicate Logic for Spatio-Temporal
`Semantics
`
`In this section we describe a methodology to express
`concept,s to express queries for video data using CSO
`and CTO and rules based on predicate logic.
`
`3.2.1 Spatial Predicate
`Here we define two sets of spatial predicates. The
`first set specifies predicates involving absolute mea-
`sures in video data, while the second set represents
`spatial predicates based on relative positions among
`objects in video d a h . These sets are defined on two
`arbitrary physical objects ”a” and ”b”. Similar rela-
`tions are described in [ll, 9, lo].
`Set I :
`D ( t , a , b ) : Dist,ance bet,ween a and b at time 1 (may
`not, need t)
`- = J(xat - z b t ) ’ + (yat - Y b t ) ’ + ( t a t - %at)’
`DP(t1, t z , a): Displacement of a between 11 and 12
`= - J ( ~ a t 1 - X a t z ) ’ + (yat1 - ~atz)’ + ( t a t 1 - %at’)’
`RC(t, a , b): The relative coordinate of b to a at time t
`- = ( z b t - r a t , Y b t - !/at, Z b t - t a t )
`
`t - Y + Y + Y +
`
`Figure 4: VSDG representation of the example clip
`
`among physical objects (described below). Examples
`are ‘slamdunk’, ‘walking’, etc. A formal definition of
`CTO is given in Section 3.2. Lastly, Physical Objects
`(PO) are the physical objects described above which
`correspond to places in VSDG. Some examples are
`persons, tree, houses, et c .
`Both CSO and CTO can be called logical objects.
`In general, any concept or semantic view of a user
`can be expressed using a set of rules from predicate
`logic that operate upon CSOs, CTOs, and POs. The
`foundation of these rules is given in Section 3.2. The
`objects and rules together can describe complex be-
`havior involving multiple objects.
`For video data, a user can use combination of vari-
`ous abstractions to construct his/her view of the video
`data. The important feature of this hierarchy, and in
`general for any object-oriented abstractions, is that
`each terminal node is either a CTO, a CSO, or a PO.
`Any complex video query is expressed as a function
`of these terminal nodes and processing of such query
`requires execution of some CTO and CSO over the
`specified PO’s. As an example, consider a sports video
`database which can be used by multiple users with
`widely different interests. Figure 5 describes an ob-
`ject hierarchy of view/knowledge which a user would
`like to construct.
`A fan may view the video data as the collec-
`tion of players, event, and teams. Furthermore, in
`his/her view, there are three types of players, for-
`ward, guard, and center. There are two types of
`event, individualxvent and team-event. Teams con-
`sist of those from NBA and NCAA. A sports fan can
`generate a query such as ‘Give the video clips where
`
`405
`
`
`
`EC(a, 6): a externally connects with 6.
`m(a,, 6,) V m(ay, 6,) V m ( a z , bz) V m-'(az, 6,)
`Vm-'(ay, by) v m - ' ( a 2 , 6,)
`We define a predicate called INTERSECT(ai, 6 i ) as
`o(ai,bi) V s ( a i , b i ) V d ( a i , b i ) V f ( a i , b i ) V o-'(ai,bi)
`Vs-'(ai,bi) v d - ' ( a i , b i ) Vf-'(a;,bi)
`PO(a, 6): a partially overlaps 6 .
`E INTERSECT(a,, bZ) A INTERSECT(ay , 6,)A
`INTERSECT(a,,b,) A -.PP(a,b) A yPP(6,a) A
`Table 2: Class definitions
`6 )
`+(a,
`a no
`DC(a, 6): a is discrete from 6 .
`E -INTERSECT(a,, 6,) A -INTERSECT(a,, by)
`A-INTERSECT(a,,b,)
`
`search(), . . .
`
`pl.).e,
`
`=.me, te.m
`birthdate
`
`In specifying the spatial predicates for Set 11, thir-
`teen spatial relations between two objects are adopted.
`The thirteen spatial relations and their notations are
`shown in Table 11. Here, we use the same notations
`as we used earlier in our temporal modeling [2]. The
`relations in Set I1 can be graphically illustrated in Fig-
`ure 6. Note that "a" and "b" can be either 2D or 3D
`objects.
`Table 3: Spatial/temporal relations and notations
`
`mt
`
`Temporal Notation
`bi
`
`Relation
`Spatial Notation
`before
`b
`after
`b-'
`meets
`m
`met - by
`m-'
`ov erlaps
`overlapped - by
`0
`0-1
`starts
`8
`started - by
`8 - 1
`during
`d
`during-'
`d-' f
`d;'
`finishes
`r;'
`f-'
`finishes-'
`eauals
`e
`e,
`The spatial predicates for Set I1 are represented as fol-
`lows :
`E(a, 6): a is equal to 6.
`= e(a,, 6,) A e(ay, 6,) A e(a,, 6 2 )
`Note that ai is a's projection on i-axis.
`PP(a,b): a is a proper part of 6.
`[ ~ ( a z 6%) V d ( a z , a,)
`V f(az, bo)] A
`M a y , 6,) v 4 a Y , 6,) v f(aY, 6, 11 A
`[S(%, 6 2 ) v d ( % , 6,) v f(a2, bz)]
`
`Macro
`I3 (Before)
`0 (Overlap)
`
`definition
`b t v mt
`V f;' V et
`0 1 V s t V d;'
`
`3.2.3 Expressing Queries Using Predicate
`Logic
`In the following section, we classify and characterize
`different types of queries and explain with examples
`how the spatio-temporal logic predicate can be used
`to describe t,hese qiieries. Generally video queries
`can be classified into two categories: spatial queries
`and spatia-temporal queries. Spatial queries deal only
`
`406
`
`
`
`with the spatial attributes of objects, including their
`appearances. Following are few queries of this type :
`
`0 Querying whether or not an object/person is
`present in a video clip(s):(x IN v ) .
`0 Identifying the relative position of object/person.
`For example, search for those video clips where
`Mr. X appears with Mr. Y, with X standing in
`front of Y. The predicate for such a query is :
`3 f E frame, 3 x , y E person
`z IN f A y IN f A x.z.depth < y.z.depth
`Here x, y are circular nodes in VSDG and r.depth
`is the depth of a bounding volume z of a circular
`node associated with x or y .
`
`The second type of queries, called spatietemporal
`queries, involve temporal dimension also and generally
`require abstraction of information over a large group
`of frames. Some typical queries which fall into this
`category are:
`
`Finding the duration of an object. For example,
`how long has person X appeared on a certain
`video clip. This query can be expressed as :
`X.duration A X IN v
`
`0 Estimating the speed of an object. For example,
`how fast is object X walking in a certain clip.
`x IN A D P ( t l , t a , z )
`t a - t i
`Here, tl and 22 are two variables denoting frame
`numbers assigned by the system.
`
`Several complex queries can be constructed using
`the above basic query types. As mentioned earlier, all
`these queries generally require processing of various
`combination of objects which are the leaf nodes of the
`object hierarchy (shown in Figure 5). Such process-
`ing in turn requires performing operations of predicate
`logic on CSO, CTO, and PO of the VSDG.
`For example, the concept slamdunk can be de-
`scribed by the following two concepts.
`
`0 Slam-dunk is a scored basket where a player's
`hand is top of the basket rim and grabs the rim
`instantaneously after the basket is successful.
`0 A scored basket is categorized as slam-dunk if
`the acceleration of the ball towards the basket
`is greater than g , the gravitation force, i.e, it is
`not a free fall.
`
`In order to construct the logic of this query, we can
`define a logic function called Ondop as follows.
`EC(a,b) A (yea > y c g )
`OndoP(a,b)
`
`E
`
`trt,
`
`Figure 7: QTVE sketches for slam-dunk
`
`Using this function and the previous defined pred-
`icates, slamdunk can be described as following;
`3 x E basketball, 3 y E basketrim, 3 z E basketnet,
`3 hl, h2 E hand
`mt (On-top(h1, x ) V Ondop(h2, x ) ,
`m z , Y ) , P P ( X , z ) , OnJop(h1, Y ) v On-top(h2, Y ) )
`Similarly, t,o specify the concept like walking. Two
`possible ways are as follows.
`(1) Assume that there is a fixed (not moving) object
`on the scene (e.g., a home). Then a person is walk-
`ing if the relative distances between the person and
`the house at. two different instants is greater than a
`threshold. Accordingly, the following expression can
`be used to describe this event.
`3 x E person, 3 y in house, 3 t l , t 2
`( t i < t2)A
`(D(RC(t1, y, x), RC(t2, y, x)) > Threshold)
`( 2 ) Assume there is no reference object. In this case
`the expression is :
`3 x E person, 3 t l , t 2
`(tl < 2 2 ) A ( D P ( t l , t 2 , z) > Threshold)
`Similarly, the concept, 'pass' can be described as
`following.
`E player, 3 y E basketball
`3 ~ 1 ~ x 2
` ( E C ( z 1 , Y), W X l
` Y ) A DC(z2, Y ) , EC(z2, Y ) )
`m t
`,
`Theoretically, any concept that requires expression
`of spatio-temporal interactions among objects can be
`specified by predicate logic expressions. We have pro-
`vided only a limited number of examples and even for
`those examples, only a few possible ways of specify-
`ing them have been discussed. Occasionally, there are
`situations where some qiieries cannot be processed by
`the proposed models and we need to process the raw
`video data. These queries may contain some objects
`which are not already identified in the VSDG model.
`
`407
`
`
`
`3.3 Query by Temporal Video Example
`
`An alternative way to express queries for video
`databases is what we call “Query by Temporal Video
`Example” (QTVE). In this method, queries can be en-
`tered by drawing sketches that examplify the concept
`in the video clips to be searched. Possible sketches
`that a user can draw for “slam-dunk are shown in
`Figure 7.
`In this method, main objects of interest and their
`positions are roughly sketched by the user in a se-
`quence of frames and the temporal information is also
`specified by the time differences, Ati’s. The number
`of frames may vary. No time difference is required for
`the single frame case. As for the video data, sketches
`are first passed through a recogniton process. The
`method used for this purpose should be different than
`that for video data, because the visual characteristics
`of sketches are quite different from those of real video
`scenes. Once the objects and their positions are iden-
`tified, a sub-VSDG can be formed the same way the
`main VSDG for data is built. The last step is a simi-
`larity matching between two VSDG’s to determine the
`video clips that contain the concept represented by the
`query VSDG. This matching process has to take two
`important issues into consideration:
`First, the shapes and positions of the sketched ob-
`jects is very approximate. Therefore, this information
`must have a high degree of tolerance in the match-
`ing process. For example, a player approaching the
`rim from right and left are two pictorially different,
`cases but conceptually identical. Second, the tempo-
`ral information must also be used with a high toler-
`ance. The user cannot be expected to specify time
`differences exactly, a range should rather be sufficient.
`In Figure 7, for example, a reasonable range for At1
`would be 0.2 - l.0sec.
`In addition to sketches, real video frames sampled
`from the database can also be used for querying. The
`advantage of this is less reliance on the ability of the
`user to draw proper figures, but the flexibility is re-
`duced in the sense that only predefined and stored
`concepts can be queried.
`
`4 Conclusion
`
`We have proposed a graphical data model called
`Video Semantic Directed Graph (VSDG) for unbiased
`abstraction and representation of video data. The
`model extracts spatial and temporal information of
`objects in a video clip and represents it in the form of
`a directed graph. A user can use this this graphical
`
`representation to construct its own view of the data.
`For this purpose, an object oriented data model is de-
`scribed. Using propositional logic described in the pa-
`per, a user can specify queries and hence can retrieve
`corresponding video clips without ever processing raw
`video data. The proposed methodology employs com-
`puter vision and image processing (CVIP) techniques
`to automate the construction of the video database
`based on the VSDG model.
`
`References
`
`[l] T . D. C. Little, and A. Ghafoor, ”Synchronization and
`Storage Models for Multimedia Objects,” IEEE Journal
`on Selected Areas In Communications, Vol. 8 , No. 3, April
`1990, pp. 413-427.
`[2] T . D. C. little, and A. Ghafoor, ”Interval-Based Concep
`tiial Models for Time-dependent Multimedia Data,” IEEE
`Tranaactions on Knowledge and Data Engineering, Vol. 5,
`No. 4, Angust 1993, pp. 551-563.
`(31 C. J. Date, An Introduction i o Database Systems, Vol. 1 ,
`5th edition, Addision Wesley, 1990.
`
`[4] D. Swanberg, C.-F. Shu, and R. Jain, “Knowledge Guided
`Parsing in Video Databases,” Proceeding of SPIE, San
`Jose, California, January 1993.
`
`[5] A. Nagasaka, and Y. Tanaka, “Automatic Video Indexing
`and Full Video Search for Object Appearances,” in 2nd
`Working Conference on Visual Database Systems, Bu-
`dapest, Hungry, October 1991, IFIP WG 2.6., pp.119-133.
`
`[6] E. Oomoto, and K . Tanaka, “OVID: Design and Imple-
`mentation of a Video-Object Database System,” IEEE
`Transactions on Knowledge and Data Engineering, Vol.
`5, No. 4, August 1993, pp. 629-643.
`
`[7] M. Ioka, and M. Kurokawa, “A Method for Retrieving
`Sequences of Images on the basis of Motion Analysis,”
`Proceeding SPIE, Image Storage and Retrieval Systems,
`San Jose, 1992, pp. 35-46.
`
`[B] A. Del Bimbo, E. Vicario, and D. Zingoni, “A Spati-
`Temporal Logic for Image Sequence Coding and Re-
`trieval,” Proceeding of IEEE VL’ 92 Workshop on Visual
`Languages, Seattle WA, September 1992, pp. 228-230.
`[9] D. A. R. and A. G. Cohn, “Exploiting Lattice in a theory
`of Space and Time,” Computer Mathematics Applications,
`Vol. 23, No. 6-9, 1992, pp.459-476.
`[IO] A. Abella, and J . R. Kender, “Qualitative Describing Ob-
`jects Using Spatial Prepositions,” in AAAI-93, Proc. of
`1 1 t h National Conference on Artificial Intelligence, pp.
`536-540.
`[ll] D. A. Randell, Z . Cui, and A. G . Cohn, “A Spatid Logic
`Based on Regions and Connection,” in Proc. of 3rd Intl.
`Conf. on Principles of Knowledge Representation and
`Reasoning, Cambridge, MA, 1992, pp. 165-176.
`
`