PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT
AWARENESS
Dimitrios Georgakopoulos and Donald Baker
Telcordia Technologies, Austin Research Center, 106 E. Sixth Street, Suite 415, Austin, Texas 78701, USA
Keywords: Video surveillance, event processing, situation awareness.
Abstract: The Video Event Awareness System (VEAS) analyzes surveillance video from thousands of video cameras
and automatically detects complex events in near real-time—at pace with their input video streams. For
events of interest to security personnel, VEAS generates and routes alerts and related video evidence to
subscribing security personnel that facilitate decision making and timely response. In this paper we
introduce VEAS’s novel publish/subscribe run-time system architecture and describe VEAS’s event
detection approach. Event processing in VEAS is driven by user-authored awareness specifications that
define patterns of inter-connected spatio-temporal event stream operators that consume and produce facility-
specific events described in VEAS’s surveillance ontology. We describe how VEAS integrates and
orchestrates continuous and tasked video analysis algorithms (e.g., for entity tracking and identification),
how it fuses events from multiple sources and algorithms in an installation-specific entity model, how it can
proactively seek additional information by tasking video analysis algorithms and security personnel to
provide it, and how it deals with late arriving information due to out-of-band video analysis tasks and
overhead. We use examples from the physical security domain, and discuss related and future work.
1 INTRODUCTION
Heightened security demand for physical security
has led to the deployment of large numbers of
surveillance cameras. However, the resulting large
volumes of video will not improve security without
the development of more advanced surveillance
technology and new practices. Although some
surveillance systems currently employ limited video
analysis automation to facilitate detection of simple
video events (i.e., events in the view of a single
camera, such as detected motion in a room or a
security perimeter), virtually all existing surveillance
systems rely mainly on human operators for the
detection of real-world situations of interest that are
complex events involving patterns of simple video
events distributed in time and possibly occurring in
different locations in a facility. Therefore, forensic
event analysis performed by human operators is
currently the main (often unstated) reason for
deploying security cameras.
One possible solution for taking advantage of the
large volume of raw surveillance video is to increase
dramatically the number of human operators
monitoring a facility. However, this solution is
typically beyond existing organizational resources
and budgets. Furthermore, adding cameras and
operators may reduce detection reliability. This is
because it is difficult for humans to reliably relate
events that occur at different times and in different
spaces in a facility, as they are typically shown in
different video displays (as illustrated in
Figure 1)
and at different times.
This work was supported in part by the Advanced Research and
Development Activity (ARDA), Video Analysis and Content
Extraction (VACE) Program under contract HM1582-04-C-0007.
Figure 1: Typical bank of monitors that security personnel
might use for monitoring a facility.
240
Georgakopoulos D. and Baker D. (2006).
PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT AWARENESS.
In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 240-251
DOI: 10.5220/0001573202400251
Copyright
c
SciTePress
The goal of the Video Event Awareness System
(VEAS) is to analyze surveillance videos from
thousand of cameras and other non-video sensors
and automatically detect complex events that
indicate situations of interest, alert humans about
them, provide the evidence that led to the alerts, and
do this in near real-time—at pace with their input
video streams.
To meet these goals, VEAS has been designed to
solve the following engineering problems.
• Correlating information from multiple video
and non-video sources into a coherent picture
of the activities of people within a facility.
• Providing near real-time event detection from
continuously run streaming video analysis
algorithms.
• Proactively seeking additional information
when there is uncertainty or gaps, usually by
tasking video analysis algorithms, and
incorporating this late arriving information into
updated alerts.
• Enabling resource optimization so that
expensive video analysis algorithms are run
only when they are likely to yield useful
information.
• Facilitating situation understanding so that
security personnel can grasp the circumstances
surrounding an alert and react more
appropriately.
• Permitting situation-specific and installation-
specific customization so that the system can
be effectively used in multiple installations
whose needs may change over time.
In this paper, we focus on VEAS’s run-time
architecture and describe its event processing model
and engines. Although we only show how these
innovations enable the automation of surveillance,
the proposed architectures and engines are general
enough to automate event detection and provide
situation awareness in any sensor-based system.
The rest of this paper is organized as follows:
Section 2 provides an overview of the VEAS
architecture and introduces a running example.
Sections 3-6 discuss the main components of the
VEAS architecture and explain how each component
helps to address the engineering problems we
described earlier in this section. Related work is
presented in Section 7, and conclusions in Section 8.
2 VEAS OVERVIEW
In the following sections we describe the main
components in the VEAS architecture and present an
example of a security policy that can be specified
and automated by VEAS.
2.1 VEAS Architecture
The main contribution of this paper is the
introduction of the novel VEAS architecture that
meets the goals of automated near real-time facility
surveillance. As depicted in Figure 2, the
architecture is comprised of a sequence of
information processing components that together
distill vast amounts of data from video and other
non-video sensors (e.g., RFID and badge readers)
into useful information in the form of alerts that are
of interest to security personnel. The arrows in
Figure 2 indicate the primary information flow
between the VEAS components. In addition to
information flow, the arrows indicate (right to left)
Alert & Task
Requests
Video
Meta Data
Facility
Events
Video
Processing
Suite
Event
Extraction
Engine
Static
Facility
Model
Content-
based
Routing
Alerts &
Tasking
Tasking
Surveillance
Event
Ontology
Awareness
Specifications
Video cameras
Awareness
Engine
Video
Segments
Live
Video
Users
Routing
Policies
Task
Specifications
Video
Repository
Non–video Sensors
Live
Non–video Data
Entity
Model
Alert & Task
Requests
Video
Meta Data
Facility
Events
Video
Processing
Suite
Event
Extraction
Engine
Static
Facility
Model
Content-
based
Routing
Alerts &
Tasking
Tasking
Surveillance
Event
Ontology
Awareness
Specifications
Video cameras
Awareness
Engine
Video
Segments
Live
Video
Users
Routing
Policies
Task
Specifications
Video
Repository
Non–video Sensors
Live
Non–video Data
Entity
Model
Figure 2: The VEAS run-time architecture.
PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT AWARENESS
241
client-server relationships and data subscription
relationships between components. The data made
available to each component triggers its
computation, producing data for the next component
in the sequence.
Live video flows from the video sources to both
the Video Repository (VR) component and the Video
Processing Suite (VP Suite) for real-time analysis by
streaming video analysis algorithms. The VR
segments video streams into manageable file sizes
using a time-based video segmentation approach
(e.g., as in (Moser, et al., 1995; Aref, et al., 2002)),
and stores them in the file system. The VP suite
provides video analysis algorithms (e.g., for entity
tracking, face detection) that produce their results
(e.g., entity tracks, video frames with faces) in near
real-time—at pace with their input video stream.
The VP Suite can additionally extract information
from video segments stored in the VR upon request
from its client components. This approach is used
for expensive video analysis algorithms (e.g., face
recognition) that operate on stored video segments
and may not produce their results in near real-time.
The delay caused by tasking video analysis
algorithms causes of information to be processed
(possibly long) after the time of the real-world
situation was captured on video. We call such
information late arriving. Since information from
non-video sensors require no processing, the VEAS
architecture lacks a processing suite for such data.
The purpose of the Event Extraction Engine
(EEE) is to collect video metadata and other data
from non-video sensors, and use these to construct
an Entity Model (EM). The EM is initialized and
subsequently updated by the Static Facility Model
(SFM) as depicted in
Figure 2. The SFM is
maintained by VEAS administrators. It captures site-
specific information that includes:
• Spatial descriptions of the space hierarchy in
the facility and security related attributes for
each space, such as its intended purpose (e.g.,
office, meeting room, record vault) and access
restrictions (e.g., escorted visitors, employees
with a specific clearance).
• Information about known entities (e.g.,
employees, contractors) and their security
related attributes (e.g., name, photo, office,
security clearance).
The Entity Model is an extensible model of the
facility, its occupants, and their activities (e.g.
movements, behaviors, etc.) that is continuously
updated as new information becomes available. To
populate the EM, VEAS’s EEE utilizes the streams
of video data from the VP Suite and non-video
sources (e.g., RFID tags and badge readers).
Additional changes to the EM occur, for instance,
when humans enter, are identified, move around,
and leave the facility. Each of these changes
constitutes a facility event. If a facility event is of
interest to the Awareness Engine, it is forwarded to
the AE for consumption.
The Awareness Engine (AE) performs
continuous detection of complex events as specified
by event patterns called awareness specifications,
and generates an alert when an event pattern in an
awareness specification is matched. Awareness
specifications are authored in concert with the
Surveillance Event Ontology that provides a
semantic type system for events of interest.
Awareness specifications and the Surveillance Event
Ontology are stored in their respective repositories
as illustrated in
Figure 2 and they are described
further in Section 4.
The AE and EEE work together to process
awareness specifications as follows: First, AE
performs a decomposition of event patterns in
awareness specifications to determine their
constituent facility events, and issues subscriptions
to these facility events from the EEE. The EEE
decomposes the subscribed facility events to desired
video metadata over (video source, video analysis
algorithm) pairs, and then performs continuous
analysis of video streams and extraction of the
metadata. The AE consumes facility events to
perform continuous detection of event patterns that
has been specified in awareness specifications.
Finally, the AE generates alert requests when the
patterns are matched, which flow to the Content-
based Routing Engine. In addition to generating
alerts requests, the AE generates task requests for
proactive information gathering tasks, which also
flow to the Content-based Routing Engine. Task
requests are generated only in response to specific
situations as described in an awareness specification.
This specificity effectively limits the cases in which
an expensive video analysis task may be run, thus
optimizing computing resources.
The Content-based Routing Engine (CBRE)
interacts with the Awareness Engine to route alerts
to appropriate security personnel. Delivered alerts
come with their pedigree and supporting video
evidence so that users can understand the cause and
greater context of the alert. This facilitates situation
understanding. The CBRE also tasks video
processing algorithms that gather information
whenever an event pattern indicates a situation
requiring further information as indicated by the AE.
Task Specifications are defined by VEAS
SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA
APPLICATIONS
242
administrators and stored in the repository shown in
Figure 2.
The following section presents a security policy
example we use throughout this paper to describe
VEAS’s functionality.
2.2 Example Security Policy
Consider the detection of unescorted visitors within
any space in a facility that requires that visitors be
escorted. For the sake of this example, we’ll divide
people into the following disjoint categories:
• employees, who must either badge-in when
they enter the facility or they must sign-in with
the receptionist to verify their identity and
receive a temporary badge; and
• visitors, who must sign in at the receptionist
desk before they enter the rest of the facility.
Visitors have assigned escorts, who might often
be the employees they are visiting. The receptionist
checks the visitor identity and consults a log of
expected visitors and employees to make the escort
assignment for the visitor.
Our example policy considers a person as
escorted if he/she is in the same space (i.e. room) as
one of their assigned escorts. Not all rooms in our
facility require visitors to be escorted. In particular,
the facility model designates each space in the
facility as either prohibiting visitors, allowing
visitors unescorted, or requiring visitor escort.
Our example security policy defines a visitor to
be unescorted if (1) the visitor is in a room that
requires visitor escorts, and (2) the visitor remains in
this room without any designated escort for longer
than a period of time, say 30 seconds. Introduction
of a time bound tolerance reduces false alarms
because a visitor and his escort may not transition
between rooms at the same time as they move about.
While employees serving as escort know that
they are responsible for escorting their visitor, they
may not always follow the security policy. In a
facility with more than a few visitors, security
personnel may have great difficulty reliably
detecting this security violation as a visitor may be
considered unescorted even if they are in a room
with people who happen not to be designated
escorts. VEAS can reliably detect this real-world
situation and alert security personnel appropriately.
In the following major sections, we will describe
the operation of VEAS’s components and we will
use the unescorted visitor security policy to discuss
each component’s contribution to the detection of
security policy violations.
3 VIDEO PROCESSING SUITE
The Video Processing (VP) Suite is actually a
collection of separate components, all of which
extract video metadata from video streams or stored
clips. Each VP component performs two functions:
• it integrates an existing video analysis
algorithm so that that acts as a consumer of live
video streams or video segments from the
Video Repository; and
• it publishes its results to the Event Extraction
Engine as it becomes available.
Algorithms in the VP Suite can be divided into the
following two broad categories:
• Continuous video analysis algorithms can be
run as video stream filters that perform video
processing in a continuous fashion.
Continuous video analysis algorithms are
directly connected to the live video sources.
• Tasked video analysis algorithms run only in
specific circumstances and they usually take
video segments as input.
Regardless of whether the VP algorithms are
continuous or tasked, they all publish their results to
the Event Extraction Engine as they are produced.
For VEAS to work properly, the VP Suite should
contain at least one continuous video analysis
algorithm for tracking entities (i.e., humans and
other movable objects). The number and location of
the cameras whose video streams are processes by
this continuous algorithm should be sufficient to
enable VEAS to gather basic knowledge of what
entities are in the facility and how they move about.
For the unescorted visitor example, VEAS uses
continuous tracking to follow entities within each
camera’s view. The VP publishes these single
camera tracks as they are collected to the Event
Extraction Engine (EEE). The EEE continuously
stitches these tracks together and correlates them
with identity information gathered from other
sources. VEAS can determine the whereabouts of
any visitor and his/her escorts in near real-time.
Tasked VP Suite algorithms are assumed more
expensive than their continuous algorithm
counterparts. Typically, tasked video analysis
algorithms are run only in situations where the
information potentially gathered by the algorithm
justifies the computational expense of running it.
Therefore, effective tasking a tasked video analysis
requires any clients of the VP suite to first identify
the situation where the algorithm might be useful
and then perform the actual invocation of the
algorithm for that particular situation. The VP Suite
handles the invocation and results publication for the
PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT AWARENESS
243
video analysis algorithm using a web services
platform.
The collection of video analysis algorithms
actually used in the VP Suite may depend on the
security policies of each VEAS installation, the
availability of algorithms and computational
resources needed to run them, and the expected costs
of their invocations. We are currently utilizing
algorithms from the U.S. Government’s DTO
(formerly ARDA) VACE program (VACE 2006).
4 EVENT EXTRACTION ENGINE
The purpose of the Event Extraction Engine (EEE)
is to perform continuous analysis of video streams
(as well as data streams from non-video information
sources) and extract simple events about the facility
being monitored, called facility events. To do this,
the EEE maintains as its primary information asset
an Entity Model (EM). The EM is a continuously
extended model of the facility, its occupants, and
their activities.
The EM is a key VEAS innovation. In
particular, EM permits simple, high-level awareness
specifications, since these are built on top of the EM
which acts as a “world” model of the facility. In
addition, EM provides a simple, yet effective,
strategy for video analysis efficiency, since it avoids
unnecessary video analysis by combining and
caching all known information for a facility.
The specific information kept in the EM is
described in Section 4.1. The construction of EM is
described in Section 4.2. The use of EM by the
Awareness Engine is discussed in Section 5.
4.1 Information in the Entity Model
The Entity Model maintains a dynamic record of
entities (i.e., humans and other movable objects) and
their movements in space and time. In particular,
the EM contains the following tracking and
identification information: spaces (e.g., rooms,
portals, etc.), entities, real world identities of people,
current and historical relationships between these,
and time intervals when these relationships occur.
The EM fuses information from various video
analysis algorithms. More importantly, EM
continuously builds three dimensions of facility
information whose purpose is to allow the
Awareness Engine (AE) to determine
instantaneously the following:
• the location of everyone in the facility at a
particular time (e.g., 1:22pm);
• the pattern of use in a particular room (e.g.,
Room 4); and
• how a particular person (e.g., John Smith)
moves around the facility over time.
Based on the availability of more advanced video
analysis algorithms in the VP Suite the EM can be
extended easily to include additional details to make
a broader or more detailed model of the facility.
Possible extensions include the following:
• Extending the spaces to larger or more
complex sites and areas, such as battle zones,
cities, etc.
• Providing state information related to
individuals under surveillance (Hongeng, et al.,
2003), e.g., body position, gestures, gaze, etc.
• Tracking inanimate entities, such as briefcases,
packages, vehicles, shipping containers, etc.
• Tracking entity-entity interactions, such as
conversations, hand-shakes, carried bags,
vehicle collisions, etc.
• Incorporating additional sensor information,
such as temperature, sound, chemical or
radiation detection, heart rate, etc.
• Modeling normalcy such as patterns of
individuals’ behavior, room usage statistics,
social network analysis, etc.
These EM extensions are beyond the scope of this
paper, and they are not discussed further.
4.2 Entity Model Management
To construct the EM, the EEE does the following:
• pulls information from the Static Facility
Model (SFM) that has been populated by
surveillance administrators;
• utilizes the video metadata (produced by both
streaming and tasked video analysis algorithms
in the VP Suite);
• utilizes the streams of information produced by
non-video sources; and
• integrates these in an incremental fashion to
make a coherent picture of what is currently
known about the facility and its occupants.
The SFM (illustrated in the VEAS architecture
in
Figure 2) is the source of many installation-
specific, space- and people-related facts in the EM.
In particular, the SFM captures information about
the facility, its hierarchy of spaces, boundaries
between those spaces, camera locations and
orientations, and identities of known people.
Additional SFM facts include the type of each space
(e.g., office, conference room, lab, hall way), the
association of office spaces to individuals and shared
SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA
APPLICATIONS
244
spaces to organizations, and the spaces that are off-
limits to each class of entities. The SFM also
provides facts about known persons. Examples of
information about a person include his/her name,
employee ID (if he/she is an employee), photos for
face recognition, etc.
Non-video sensors (e.g. RFID readers, badge
scanners) typically provide information as to the
identity of moving entities and their position in the
facility as they move about in time. Such
information streams can be correlated in the EM
with similar information collected from video.
EEE fuses all its input information in facility-
wide entity tracks, and inferences cross-track entity
IDs even in situations where entities move through
areas that have no camera coverage. EEE also
incorporates the real-world identities of the entities.
To perform facility-wide entity tracking, the EEE
keeps video source independent information in the
EM concerning how entities move through the
facility. This information is camera independent and
makes no assumptions about the real-world identity
of persons. Tracking algorithms generate entity IDs
when they detect a moving entity. Facility-wide
tracking can be performed at different, predefined
levels of spatial granularity. Supported granularities
in VEAS include buildings, rooms (both captured in
the static facility model), and global (grid-based)
coordinates. For a given entity ID and interval of
time, the EEE records in the EM a sequence of
tuples indicating that the person spent a particular
interval of time in a particular space. The entire
sequence also has a pedigree summarizing the
sources considered in generating the sequence and a
certainty measure. Each sequence is contiguous and
has a unique entity ID as long as the Event
Extraction Engine has a high confidence of the
continuity of identity.
Cross-track entity ID inference is necessary
when blind spots exist in the facility, e.g., rooms
without cameras. In this case, there may be
circumstances where it is not known with certainty
whether the entity ID associated with one track is the
same as that of another track. To infer the real-
world identity of entities the EEE may use one of the
following solutions:
• Correlate a person seen in a video segment
with the time and place of a badge scan or
RFID reading at that location.
• Task face recognition or other identification
algorithms in the VP Suite to determine the
identity of persons appearing in a video
segment through comparison to a set of known
faces (such information for employees may be
included in the Static Facility Model).
• Utilize assessments made by security personnel
tasked with determining the identity of a
person through viewing a video segment or via
direct interaction with the person in question.
In the last two solutions, tasking is initiated by
the Awareness Engine, as described in Section 4.
For the unescorted visitor example, the EEE
tracks the locations and identities of all people as
they move around the facility. The EM utilizes
information from the Static Facility Model to
determine how various spaces are designated as to
whether visitor escort is required within them, and
the real world identities of known entities, such as
the employees. As people enter the facility, badge
scans and receptionist records published into the
EEE are correlated with the video tracks of entities
as they move around the facility. The EM is also
capable of recording the designated escorts for each
visitor. This information is provided to the EEE via
the receptionist.
The Event Extraction Engine publishes the EM
incrementally, i.e., as video and non-video
information is streamed to the EEE and is published
within in the EM. The Awareness Engine can
register interest in any particular combination of
persons, spaces, time intervals and receives updates
in the form of facility events.
The EM is maintained as a 3-tiered storage
solution. In particular, the EM is kept in a main
memory data structure that serves as a cache for an
EM database that is also periodically archived.
Transition of information between the EM storage
tiers is based on installation-specific policies that
consider the age of the information and whether it
contributes to alerts that are still of interest to
security personnel.
5 AWARENESS ENGINE
The Awareness Engine (AE), illustrated in Figure 2,
consumes facility events published by the Event
Extraction Engine and detects complex events.
When it detects such events, the AE publishes alert
requests and task requests that flow to the Content-
based Routing Engine for execution. To detect
complex event and compute corresponding alert and
task requests, the AE utilizes user-authored
awareness specifications that describe how to
compute complex events of interest and who should
receive an alert or perform a task.
PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT AWARENESS
245
In Section 5.1, we introduce VEAS’s event
ontology and awareness specifications. Section 5.2
describes the types of event stream operators that
exist in VEAS and discusses how event operator
types can be created and extended. Section 5.3
describes the alert and tasking event operators. In
these sections, we provide a detailed example of an
awareness specification for the unescorted visitor
security policy.
5.1 Awareness Specifications
Awareness specifications are comprised of
customized, interconnected computational units
called event operators.
Figure 3 shows such a
specification for detecting unescorted visitors.
Event operators in
Figure 3 appear as boxes with
a descriptive name. Ovals in the figure number the
event operators for reference in this paper. Event
operators are connected into computational pipelines
that operate in parallel. Within each pipeline, events
flow from the bottom to the top. The inputs and
outputs of event operators are typed streams of
events (i.e., packets of information indicating real-
world situations).
Event types of interest, their relationships, and
the information computed for each event type is
described in the Surveillance Event Ontology (not
shown in Figure 3). VEAS’s Surveillance Event
Ontology (SEO) acts as a type system describing the
events of interest for a specific installation of VEAS.
For example, the Zone Visit event type described in
the SEO includes information about the space within
the facility being visited; the entity ID of the visitor;
the real-world identity of the visitor, should the
visitor be a person; and the time interval of the visit.
Event operators are computational units that
recognize more complex real-world events
(produced on their output stream) from simpler real-
world events (consumed on their input stream(s)). A
connection between the output of one operator and
the input of another is only allowed if they have
semantically compatible event types, as described by
the Surveillance Event Ontology. This restriction
helps ensure information compatibility between
producing and consuming event operators, but it also
ensures semantic compatibility, so that the overall
awareness specification computes alert and task
requests relative to the author’s understanding of the
real-world event being specified.
To meet the surveillance needs of a particular
installation (e.g., to support the specific security
policies that need to be enforced), VEAS provides a
tool called the Awareness Specification Editor that
permits authorized VEAS users to:
• Customize the Surveillance Event Ontology by
creating new/refining existing event types.
• Author awareness specifications by inter-
connecting and customizing event operators via
dialog boxes.
The awareness specification in Figure 3 is shown
in this tool.
To aid in the development of new operators that
cannot be produced by customizing existing ones,
VEAS provides a programming interface. With
proper training, new event operator types can be
developed by a programmer in a matter of hours.
VEAS permits the creation of installation-
specific operator palettes that are aimed to simplify
authoring of awareness specification in each specific
VEAS deployment. Such pallets typically include
custom operator types with names closely
resembling the main concepts in the security policies
of a specific installation. In Sections 5.2 and 5.3, we
discuss various operators and give examples.
5.2 Event Stream Operators
VEAS provides two generic spatio-temporal
operator types that use the Entity Model as their tacit
input. These operators (and customized operators
based on them) commonly appear as event stream
sources within virtually all awareness specifications.
The Qualified Zone Visit operator type allows the
specification of simple events qualified by time,
space, identity, or attributes thereof for instances of
people known to be in spaces throughout the facility
and produces events of the Zone Visit event type,
described earlier. A similar event operator type,
Figure 3: Example awareness specification utilizing
operators 1-10.
SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA
APPLICATIONS
246
Qualified Portal Traversal, recognizes people
traversing specific portals within the facility.
There are four Qualified Zone Visit event
operators appearing in the unescorted visitor
awareness specification in Figure 3. These operators
have been customized to recognize (from top to
bottom and left to right):
• situations where an unidentified person is in
the lobby of the facility where the receptionist
sits (operator 5);
• situations where a visitor is known to be
somewhere in the facility (operator 6);
• situations where a visitor is known to be in a
zone where visitor escort is required (operator
9); and
• situations where an employee is known to be in
a zone where visitor escort is required
(operator 10).
The title of the event operator is displayed and each
operator is customized via dialog boxes. Events
meeting a specification flow (upward in the
specification) via event operator interconnections.
VEAS provides a few set operator types that
perform set functions over their input operator
streams. The Or event operator type computes a
union of its input event streams. The Difference
event operator type computes a set difference of two
input streams. These operators are polymorphic
with respect to their input and output event types.
The remaining event stream operator types can
be categorized in broad functional categories:
• joining – combining related events from
multiple input streams into composite events
on the output stream;
• filtering – culling of uninteresting events from
the input in the output; and
• grouping and aggregation – grouping of
multiple input events from a single stream into
aggregated output events.
Join event operators behave like a join operator
in a stream query in that they merge information
from multiple sources. The Accompanied Zone Visit
operator type, for example, joins two input streams
of Zone Visit events. The first input stream
describes people in zones for which accompaniment
information is desired. The second input stream
describes people in zones who are possible
accompanists. For each event on its first input, the
operator outputs an event that adds information
about whether the person was accompanied, by
whom, during what time intervals, and the intervals
during which he or she was unaccompanied.
For the unescorted visitor example, we created a
variant of Accompanied Zone Visit operator type.
The Escorted Zone Visit (operator 8) is an
installation-specific join operator customized to find
time intervals where both a visitor and one or more
of his escorts are in the same zone at the same time.
The operator computes the associated escorts, the
escorted time intervals, and the unescorted time
intervals, and the longest unescorted time interval as
part of its output. These values are computed
regardless of whether any escorts are found for the
visitor. Visitor zone visits to be considered are
provided on the first input from event operator 9.
Possible escort zone visits to consider are input as
the second input from event operator 10.
Filtering
operator types are the simplest category
in that they output events from their input event
stream that meet the filter criteria as specified by the
event operator’s customization. In the unescorted
visitor example, the Unescorted Visitor (operator 7)
is an installation-specific filter event operator that
looks for escorted zone visits on the input with a
longest unescorted interval that is greater than the
specified threshold. In this specification, the
threshold is set to 30 seconds via a customization
dialog box.
An example of an aggregation is the Fleeing
operator type (not used in the awareness
specification in Figure 3). The operator produces
output events that represent episodes of people
exiting a particular space meeting the criteria of a
minimum count of people exiting and a minimum
sustained exit rate over the episode. The operator is
used to look for mass exodus events in a facility that
may or may not be accompanied by alarms or other
signals known to security personnel. The Fleeing
event operator type is an example of an event
operator that embodies a computation that is difficult
or impossible to express in a stream query language.
Its notion of grouping is beyond the capability of
standard “group by” constructs as the group is not
readily indexed. It would be impossible to capture
this notion of fleeing in a simple streaming query.
5.3 Alert and Tasking Operators
The Alert Delivery and Information Gathering
operator types submit prioritized alert and
information gathering task requests, respectively, to
the Content-based Routing Engine. Operators of
these types appear as event stream sinks in
awareness specifications. Both alert requests and
tasks requests have a priority value that is
incrementally computed in these two operators by a
threat analysis module based on the input event type
and event instance information.
PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT AWARENESS
247
The Alert Delivery operator converts its input
event stream into prioritized alert requests to be sent
to the Content-based Routing Engine. User
customization of this operator effectively provides
delivery instructions that include the target delivery
role (of users who should receive the alert), alert
name, and alert description. In the unescorted
visitor awareness specification in Figure 3, escorted
zone visits matching the criteria (of having a
suitably long unescorted period) are output from the
Unescorted Zone Visit (operator 7) and flow to Alert
Delivery event operators 3 and 4, which generate
alerts for unescorted visitors found in the facility,
but to different roles. One operator alerts the
security guard and the other operator alerts those
people playing the situation-specific role of the
escorts who are failing to escort their visitor.
The Information Gathering event operator issues
tasks (via the Content Routing Engine) to security
personnel and the VP Suite. User customization of
the operator adds reference to a task specification
and a role of users to who can perform the task. The
special role “VP suite” is used to task the VP suite to
run a video analysis algorithm. The input event
stream to an Information Gathering operator
effectively describes real-world events that indicate
that more information is needed by VEAS. The task
specification describes an information gathering task
to be executed for each such event.
The Information Gathering event operator turns
each of its input events, which describe a real world
situation, into a corresponding request to gather
needed information via an information gathering
task, such as the running of video analysis
algorithms or consulting security personnel for their
assessments. The eventual execution of these
information gathering tasks by the Content-based
Routing Engine both consume resources and create
delays. Thus, the Information Gathering event
operator is a departure from the strictly incremental
processing approach used by other event operators.
The results produced by information gathering
tasks do not directly flow to the Awareness Engine.
Instead, the information flows to the Awareness
Engine indirectly via published changes in the Entity
Model. There are two Information Gathering event
operators along the top of the specification in Figure
3. Event operator 1 tasks the receptionist to identify
each unknown person in the lobby. For visitors
known to be in the facility, event operator 2 will
generate an automatically executed task that will
look through an expected visitor log, determine the
appropriate escort(s) for the visitor, and tag the
visitor identity attribute in the Entity Model with this
information. This information is used by event
operator 8 in determining whether the visitor is
escorted.
6 CONTENT-BASED ROUTING
ENGINE
The Content-based Routing Engine (CBRE)
combines the coordination capabilities of workflow
systems (Georgakopoulos, et al., 2000; BEA;
TIBCO; Vitria; Georgakopoulos, 2004) with the
content routing and syndication capabilities of
existing content management systems (EMC;
FatWire; FileNet; Georgakopoulos 2004). CBRE-
provided workflow and content management
capabilities are needed to dynamically service the
following types of requests from the AE:
• prioritized alert requests for alerts that must be
delivered to security personnel; and
• prioritized task requests for tasks that need to
be performed for gathering more information.
To service prioritized alert requests, the CBRE
supports content and alert routing policies that
determine the appropriate user (or subset of users) to
route specific alerts and related evidence (e.g., video
segments). In addition, CBRE provides a mechanism
and a client tool that allows users to view alerts, and
drill down into the video evidence for each alert.
CBRE processes prioritized task requests by
automatically invoking video analysis algorithms
from the VP Suite, assigning tasks to users playing
security personnel roles, and coordinating the
execution of all tasks that are required by an
awareness specification. These CBRE functions are
discussed in more detail in the following sections.
6.1 Processing of Tasks
To support task assignment, coordination, and
automation, the CBRE includes workflow
management functionality. In particular, CBRE’s
workflow capabilities are similar to those provided
by CMI (Georgakopoulos, et al., 2000).
Just like many other workflow management
systems, CBRE’s workflow component provides for
the automation/enactment of workflow processes.
Workflow processes typically consist of recursively
defined activities, and specify control flow and
dataflow between them. Basic activities, the leaves
of the process specification tree, may be assigned to
humans playing various roles, or be automatically
executed using available computing resources.
SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA
APPLICATIONS
248
The Task Specifications in Figure 2 are actually
workflows specifications that have been defined by
VEAS administrators to automate the tasking of
video processing algorithms and security personnel.
CBRE selects, instantiates, and enacts task
specifications in response to requests from AE.
Although CBRE’s workflow component offers
general workflow capabilities, the tasks, data flow,
and control flow performed by CBRE (and VEAS in
general) are typically centered on the orchestration
of information gathering (involving VEAS video
analysis and people). This permits pre-integration of
video of the finite set of video analysis algorithms in
the VP Suite, and the definition of typical human
tasks for providing expert opinion and performing
security-related activities (e.g., ask a person for an
ID, secure a space in the facility, etc.) in a palette of
reusable task specifications that require little or no
customization and that can be readily utilized in
awareness specifications.
Tasks to be executed are specified through the
Information Gathering event operator, described in
Section 5.3. In the example in Figure 3, the
Awareness Engine issues a Determine Identity task
request to CBRE to collect additional information.
This task is pre-specified and implemented by a
workflow processes consisting of several subtasks.
The first subtask queries the Entity Model in the
EEE to determine the time the person under
investigation was in the restricted room. This is
followed by a task that involves the execution of a
face recognition algorithm on video collected during
this time interval. If the face recognition algorithm
fails to find at least one video frame with the
person’s face, a security guard is tasked to go to the
restricted room and question the unidentified person
as a last resort. Results of the Determine Identity
flow back to the Entity Model via the Event
Extraction Engine. If such a task determines the
identity of the person and his presence in this room
is a security threat (e.g., he is a visitor alone in a
room that requires visitors to have escorts), the
Awareness Engine will issue an alert request to the
CBRE which then routes the alert to the specified
security guard (e.g., the security guard who is closer
to the visitor’s current location).
6.2 Alert Delivery and Drill Down
Just like many other workflow engines
(Georgakopoulos, et al., 2000; BEA; TIBCO; Vitria;
Georgakopoulos, 2004), CBRE provides a worklist
mechanism and a corresponding tool. Alerts are
delivered in the worklist of VEAS users. Users may
select a specified alert to view its properties, play
video segments with evidence, and drill down to
determine its causes and view related video
evidence. This CBRE drill-down mechanism
involves the display of details of how the triggering
awareness specification decided the occurrence of a
complex event, images and animations of entities
moving within the facility, annotated or raw videos,
and other relevant information.
Figure 4 shows an alert in the worklist tool in the
backmost window. The next window shows the
detail of the unescorted visitor alert, which includes
the pedigree and video evidence for the alert. The
unescorted visitor is seen in the right of the video
frame within a video segment that was given as
evidence for the alert.
6.3 Routing Policies
The workflow component in the Content-based
Routing Engine automatically routes alarms and
tasks to specific roles. Therefore, the CBRE needs a
mechanism to associate users and roles. This is
accomplished by defining CBRE-supported routing
policies that describe how specific users of VEAS
correspond to the roles they might play in viewing
specific security alerts and/or giving VEAS expert
feedback as part of an information gathering task.
Routing policies include:
• Alarm delivery polices: Definition of these
policies involves the association of VEAS
users with specific target roles (e.g., assign
user “John Smith” to the role “security officer
on duty in entrance B”). These policies are
meaningful in a specific facility.
Figure 4: Alert drill down and video evidence.
PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT AWARENESS
249
• Expert feedback polices: These associated
VEAS users with “expert” and “security” roles
(e.g., to review a video segment, secure a
space, etc.).
• Delegation and escalation polices: These
policies define what to do when an alert or a
request has been delivered, but it cannot
“consumed” by a user in the assigned role.
Policies in this category permit a user to
delegate an alarm directed to him/her to
another role, and redirect alarms to other roles
when a specified timeout expires.
7 RELATED WORK
The most advanced commercial surveillance
systems, e.g., (GVI), utilize multi-camera motion
detection to detect simple predefined events, and
they are useful for guarding a perimeter (e.g., track
potential intruders along a fence). Unlike VEAS,
such surveillance systems cannot detect user–
specified complex events, cannot utilize installation-
specific knowledge (e.g., know who is a visitor or an
employee, or be aware of spaces requiring escort and
those that do not), and cannot combine and/or task
sophisticated video analysis algorithms.
Early event processing systems, such as Snoop
(Chakravarthy, 1994) developed event algebra based
models, with generic event operators such a filter,
sequence, and count. CEDMOS (Baker, et al., 1999)
moved toward self-contained events and the
computation of event parameters for complex
events. Although these systems explored ideas that
have been adopted by VEAS and stream databases,
usability and efficiency were not addressed.
Stream databases, e.g., STREAM (Stanford
University), Aurora (Abadi, et al., 2003), TinyDB
(UC Berkeley), Borealis (Borealis Project; Abadi, et
al., 2005), and Streambase (Streambase System),
utilize a relational model with SQL enhanced with
time-based windows for data streams. Operators are
based on generic relational operators, i.e., selection,
projection, join, aggregates, and group-by.
The VEAS model and language, which builds on
ideas from our earlier awareness work (Baker et al.,
2002), includes surveillance-specific operators are
built on a dynamic Entity Model. They are
specializations spatial operators (e.g., in Meeting
room 2), temporal operators (e.g., workweek,
holiday, 3rd shift), entity identification operators
(e.g., visitor, employee in a specific organization),
and relational algebra operators.
To perform event processing, stream databases
(e.g., STREAM (Stanford University), Aurora
(Abadi, et al., 2003), TinyDB (UC Berkeley), and
Borealis (Borealis Project; Abadi, et al., 2005))
require use of time windows for stream queries and
assume that there is no late arriving information.
Optimization is performed assuming that input
information is readily available and no information
extraction cost is considered.
In contrast, VEAS requires no time windows
(which is a requirement in video surveillance due to
arbitrarily late arriving information). It performs
incremental computation, and deals with information
arriving late. These capabilities permit awareness
specifications to take into account and reduce the
cost of video analysis tasks.
Stream databases have no tasking and
information gathering capabilities. In contrast,
VEAS proactively gathers information that is
missing to confirm or refute a partially matched
event pattern within an awareness specification.
VEAS is capable of tasking and executing video
analysis algorithms and/or involve human (e.g.,
subject matter expert) to collect needed information
and decision making.
The CBRE combines the coordination
capabilities of workflow systems (Georgakopoulos,
et al., 2000; BEA; TIBCO; Vitria; Georgakopoulos,
2004) with the content routing and syndication
capabilities of existing content management systems
(EMC; FatWire; FileNet; Georgakopoulos, 2004).
In addition, CBRE provides a novel drill down
capability for determining the evidence of and
pedigree of each alert in support of situation
understanding by end-users, as shown in
Figure 4.
8 CONCLUSION
VEAS helps automate surveillance by analyzing
surveillance video from potentially thousands of
cameras and other non-video sensors and
automatically detecting complex events that indicate
situations of interest, alerting humans about them,
providing the evidence that led to the alerts, and do
this in near real-time—at pace with their input video
streams. When information is missing or uncertain,
VEAS has the capability to gather additional
information proactively to make the appropriate
determinations. In this paper, we focused in the
presentation of VEAS’s novel runtime architecture
and event processing capabilities, and described the
application of these in the video surveillance
domain. The novel capabilities and key benefits of
SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA
APPLICATIONS
250
VEAS are increased alert accuracy (by narrowing
the video analysis problem via installation-specific
customization), proactive gathering of additional
information, and the ability to monitor any facility of
interest, and to do so inexpensively and in near real-
time. VEAS is scalable and does not require adding
more humans to monitor the additional security
cameras when a facility is expanded.
Although in this paper we focused on video
surveillance, VEAS can effectively utilize non-video
sensors and/or be applied in other domains. Our
directions for future work include applying VEAS to
support complex event detection in domains related
to meetings and conferences. We are also working to
extend VEAS capabilities for managing non-video
sensors, such as RFID readers, badge scanners, and
intelligent locks aimed at providing Critical
Infrastructure Protection (CIP) without relying on
video. We are also experimenting with algorithms
for learning of normal event patterns to enable
VEAS to automatically detect anomalous events that
have not been anticipated.
REFERENCES
Abadi, D. J., D. Carney, U. Çetintemel, M. Cherniack, C.
Convey, S. Lee, M. Stonebraker, N. Tatbul, S. Zdonik
(2003). Aurora: a new model and architecture for data
stream management. VLDB Journal, 12(2).
Abadi, D. J., Y. Ahmad, M. Balazinska, U. Çetintemel, M.
Cherniack, J.-H. Hwang, W. Lindner, A. Maskey, A.
Rasin, E. Ryvkina, N. Tatbul, Y. Xing, and S. Zdonik
(2005). The Design of the Borealis Stream Processing
Engine. Second Biennial Conference on Innovative
Data Systems Research (DR 2005).
Aref, W., A. C. Catlin, A. Elmagarmid, J. Fan, J. Guo, M.
Hammad, I. Ilyas, M. Marzouk, S. Prabhakar, A.
Rezgui, S. Teoh, E. Terzi, Y. Tu, A. Vakali and X.
Zhu (2002). A distributed server for continuous
media. 18th International Conference on Data
Engineering (ICDE’02).
VACE (2006). VACE Phase III BAA Discussion.
Retrieved 2006 from http://www.nbc.gov/fort_h
/vaceiii/BidderBrief-final.ppt.
Baker, D., D. Georgakopoulos, H. Schuster, and A.
Cichocki (2002). Customized Process and Situation
Awareness. International Journal of Cooperative
Information Systems, 11(3 & 4), World Scientific.
Baker D., A. Cassandra, and M. Rashid (1999).
CEDMOS: Complex Event Detection and Monitoring
System. MCC Technical Report CEDMOS-002-99.
Retrieved 2005 from http://citeseer.ist.psu.edu/
baker99cedmos.html
BEA (nd). WebLogic. Retrieved 2005 from
http://www.bea.com/.
Borealis Project. Borealis Second Generation Stream
Processing Engine. Retrieved 2005 from
http://nms.lcs.mit.edu/projects/borealis/.
Chakravarthy, S. (1994). Snoop: An Expressive Event
Specification Language for Active Databases. IEEE
Data and Knowledge Engineering, 14(10).
EMC (nd). Documentum. Retrieved 2005 from
http://software.emc.com/.
FatWire (nd). Content Server. Retrieved 2005 from
http://www.fatwire.com/.
FileNet (nd). Content and Workflow Managers.
Retrieved 2005 from http://www.filenet.com/.
Georgakopoulos, D., H. Schuster, D. Baker, and A.
Cichocki (2000). Managing Escalation of
Collaboration Processes in Crisis Mitigation
Situations. 16th International Conference on Data
Engineering (ICDE’00).
Georgakopoulos, D. (2004). Teamware: An Evaluation of
Key Technologies and Open Problems. Distributed
and Parallel Databases, 15(1).
GVI (nd). GVI TRAK. Retrieved 2005 from
http://www.samsung-security.com/.
Hongeng, S., R. Nevatia, and F. Bremond (2003). Video-
based Event Recognition: Activity Representation and
Probabilistic Recognition Methods. In Computer
Vision and Image Understanding, 96.
Moser, F., A. Kraiss, and W. L. Klas (1995). A buffer
management strategy for interactive continuous data
flows in a multimedia DBMS. 21st International
Conference on Very Large Data Bases (VLDB’95).
Nodine, M., J. Fowler, T. Ksiezyk, B. Perry, M. Taylor
and A. Unruh (2000). Active Information Gathering
in InfoSleuth. International Journal of Cooperative
Information Systems, 9(1/2).
Stanford University (nd). STREAM: The Stanford Stream
Data Manager. Retrieved 2005 from http://www-
db.stanford.edu/stream/.
Streambase (nd). Streambase System. Retrieved 2005
from http://www.streambase.com/.
TIBCO (nd). TIBCO Staffware Process Suite. Retrieved
2006 from http://www.tibco.com.
UC Berkeley (nd). TinyDB: A Declarative Database for
Sensor Networks. Retrieved 2005 from
http://telegraph.cs.berkeley.edu/tinydb/.
Vitria (nd). BusinessWare. Retrieved 2005 from
http://www.vitria.com/.
PROVIDING PHYSICAL SECURITY VIA VIDEO EVENT AWARENESS
251
