A Framework for a Video Analysis Tool for Suspicious Event Detection

Download Report

Transcript A Framework for a Video Analysis Tool for Suspicious Event Detection

1
Data Mining for Surveillance
Applications
Suspicious Event Detection
Dr. Bhavani Thuraisingham
April 2006
Outline





Acknowledgements
Data Mining for Security Applications
Surveillance and Suspicious Event Detection
Directions for Surveillance
Data Mining, Security and Privacy
2
Acknowledgements





Prof. Latifur Khan
Prof. Murat Kantarcioglu
Gal Lavee
Ryan Layfield
Sai Chaitanya
3
Our Vision:
Assured Information Sharing
Data/Policy for Coalition
Publish
Data/Policy
Publish
Data/Policy
Publish
Data/Policy
Component
Data/Policy for
Agency A
Component
Data/Policy for
Agency C
Component
Data/Policy for
Agency B
1.
Friendly partners
2.
Semi-honest partners
3.
Untrustworthy partners
4
Data Mining for Security
Applications

Data Mining has many applications in Cyber
Security and National Security
Intrusion detection, worm detection, firewall policy
management
 Counter-terrorism applications and Surveillance
 Fraud detection, Insider threat analysis


Need to enforce security but at the same time
ensure privacy
5
Data Mining for Surveillance
Problems Addressed



Huge amounts of surveillance and
video data available in the security
domain
Analysis is being done off-line
usually using “Human Eyes”
Need for tools to aid human
analyst ( pointing out areas in video
where unusual activity occurs)
6
Example

Using our proposed system:
Video Data
User Defined
Annotated Video w/
events of interest
highlighted
Event of interest

Greatly Increase video analysis efficiency
7
The Semantic Gap

The disconnect between the low-level features a
machine sees when a video is input into it and the highlevel semantic concepts (or events) a human being sees
when looking at a video clip

Low-Level features: color, texture, shape
High-level semantic concepts: presentation,
newscast, boxing match

8
Our Approach

Event Representation


Event Comparison


Estimate distribution of pixel intensity change
Contrast the event representation of different video
sequences to determine if they contain similar
semantic event content.
Event Detection

Using manually labeled training video sequences to
classify unlabeled video sequences
9
Event Representation



Measures the quantity and type of changes occurring within a scene
A video event is represented as a set of x, y and t intensity gradient
histograms over several temporal scales.
Histograms are normalized and smoothed
10
Event Comparison

Determine if the two video sequences contain similar
high-level semantic concepts (events).
l
l
2
[
h
(
i
)

h
(
i
)]
1
1k
2k
D2 

3L k ,l ,i h1lk (i)  h2l k (i)


Produces a number that indicates how close the two
compared events are to one another.
The lower this number is the closer the two events are.
11
Event Detection

A robust event detection system should be able
to
Recognize an event with reduced sensitivity to actor
(e.g. clothing or skin tone) or background lighting
variation.
 Segment an unlabeled video containing multiple
events into event specific segments

12
Labeled Video Events

These events are manually labeled and used to
classify unknown events
Walking1 Running1 Waving2
13
14
Labeled Video Events
walking1
walking2
walking3
running1
running2
running3
running4
waving 2
walking1
0
0.27625
0.24508
1.2262
1.383
0.97472
1.3791
10.961
walking2
0.27625
0
0.17888
1.4757
1.5003
1.2908
1.541
10.581
walking3
0.24508
0.17888
0
1.1298
1.0933
0.88604
1.1221
10.231
running1
1.2262
1.4757
1.1298
0
0.43829
0.30451
0.39823
14.469
running2
1.383
1.5003
1.0933
0.43829
0
0.23804
0.10761
15.05
running3
0.97472
1.2908
0.88604
0.30451
0.23804
0
0.20489
14.2
running4
1.3791
1.541
1.1221
0.39823
0.10761
0.20489
0
15.607
waving2
10.961
10.581
10.231
14.469
15.05
14.2
15.607
0
Experiment #1

Problem: Recognize and classify events irrespective of
direction (right-to-left, left-to-right) and with reduced
sensitivity to spatial variations (Clothing)

“Disguised Events”- Events similar to testing data
except subject is dressed differently
Compare Classification to “Truth” (Manual Labeling)

15
16
Experiment #1
Disguised Walking 1
walking1
0.97653
walking2
0.45154
walking3
0.59608
running1
1.5476
running2
1.4633
running3
1.5724
Classification: Walking
running4
1.5406
waving2
12.225
17
Experiment #1
Disguised Running 1
walking1
1.411
walking2
1.3841
walking3
1.0637
running1
0.56724
running2
0.97417
running3
0.93587
Classification: Running
running4
1.0957
waving2
11.629
18
Experiment #1
Disguised Running 3
walking1
1.3049
walking2
1.0021
walking3
0.88092
running1
0.8114
running2
1.1042
running3
1.1189
Classification: Running
running4
1.0902
waving2
12.801
19
Experiment #1
Disguised Waving 1
walking1
13.646
walking2
13.113
walking3
13.452
running1
18.615
running2
19.592
running3
18.621
Classification: Waving
running4
20.239
waving2
2.2451
20
Classifying Disguised Events
Disguise
walking1
Disguise
walking2
Disguise
running1
Disguise
running2
Disguise
running3
Disguise
waving1
Disguise
waving2
Disguise
walking1
0
0.19339
1.2159
0.85938
0.67577
14.471
13.429
Disguise
walking2
0.19339
0
1.4317
1.1824
0.95582
12.295
11.29
Disguise
running1
1.2159
1.4317
0
0.37592
0.45187
15.266
15.007
Disguise
Running2
0.85938
1.1824
0.37592
0
0.13346
16.76
16.247
Disguise
Running3
0.67577
0.95582
0.45187
0.13346
0
16.252
15.621
Disguise
waving1
14.471
12.295
15.266
16.76
16.252
0
0.45816
Disguise
waving2
13.429
11.29
15.007
16.247
15.621
0.45816
0
Experiment #1



This method yielded 100% Precision (i.e. all
disguised events were classified correctly).
Not necessarily representative of the general
event detection problem.
Future evaluation with more event types, more
varied data and a larger set of training and
testing data is needed
21
Experiment #2


Problem: Given an unlabeled video sequence describe the high-level events
within the video
Capture events using a sliding window of a fixed width (25 frames in
example)
22
Experiment #2

Running Similarity Graph
23
Experiment #2

Walking Similarity Graph
24
Experiment #2

Waving Similarity Graph
25
Experiment #2
Minimum Similarity Graph
Walking
Running
Waving Running
26
XML Video Annotation


Using the event detection scheme we generate a video
description document detailing the event composition of a
specific video sequence
This XML document annotation may be replaced by a more
robust computer-understandable format (e.g. the VEML video
event ontology language).
<?xml version="1.0" encoding="UTF-8"?>
<videoclip>
<Filename>H:\Research\MainEvent\
Movies\test_runningandwaving.AVI</Filename>
<Length>600</Length>
<Event>
<Name>unknown</Name>
<Start>1</Start>
<Duration>106</Duration>
</Event>
<Event>
<Name>walking</Name>
<Start>107</Start>
<Duration>6</Duration>
</Event>
</videoclip>
27
Video Analysis Tool



Takes annotation document as input and organizes the corresponding video
segment accordingly.
Functions as an aid to a surveillance analyst searching for “Suspicious” events
within a stream of video data.
Activity of interest may be defined dynamically by the analyst during the
running of the utility and flagged for analysis.
28
Directions


Enhancements to the work
 Working toward bridging the semantic gap and enabling more efficient
video analysis
 More rigorous experimental testing of concepts
 Refine event classification through use of multiple machine learning
algorithm (e.g. neural networks, decision trees, etc…). Experimentally
determine optimal algorithm.
Develop a model for the following



29
simultaneous events within the same video sequence
Face detection, Gait detection
Security and Privacy
 Define an access control model that will allow access to surveillance video
data to be restricted based on semantic content of video objects
 Biometrics applications
 Privacy preserving surveillance
Access Control and Biometrics

Access Control



RBAC and UCON-based models for surveillance data
Initial work to appear in ACM SACMAT Conference 2006
Biometrics



Restrict access based on semantic content of video rather
then low-level features
Behavioral type access instead of “fingerprint”
Used in combination with other biometric methods
30
Privacy Preserving
Surveillance - Introduction
•A recent survey at Times Square found 500 visible
surveillance cameras in the area and a total of 2500 in New
York City.
•What this essentially means is that, we have scores of
surveillance video to be inspected manually by security
personnel
•We need to carry out surveillance but at the same time ensure
the privacy of individuals who are good citizens
32
System Use
Raw video surveillance data
Faces of trusted
people derecognized
to preserve privacy
Face Detection and
Face
Derecognizing
system
Suspicious Event
Detection System
Manual Inspection
of video data
Suspicious people
found
Suspicious events
found
Report of security personnel
Comprehensive
security report
listing suspicious
events and people
detected
System Architecture
Input Video
Finding location of
the face in the image
Breakdown input
video into sequence
of images
Raise an alarm
that a potential
intruder was
detected
Perform
Segmentation
Potential
intruder
found
Compare face to
trusted and
untrusted
individuals
Trusted face found
Derecognize the face
in the image
Other Applications of
Data Mining in Security











Intrusion detection and worm detection
Firewall policy management
Insider Threat Analysis – both network/host and physical
Fraud Detection
Protecting children from inappropriate content on the Internet
Digital Identity Management and Detecting Identity Theft
Steganalysis and digital watermarking
Biometrics identification and verification
Digital Forensics
Source Code Analysis
National Security / Counter-terrorism
34
Data Mining Needs for Counterterrorism:
Non-real-time Data Mining
35







Gather data from multiple sources
 Information on terrorist attacks: who, what, where, when, how
 Personal and business data: place of birth, ethnic origin, religion,
education, work history, finances, criminal record, relatives, friends and
associates, travel history, . . .
 Unstructured data: newspaper articles, video clips, speeches, emails,
phone records, . . .
Integrate the data, build warehouses and federations
Develop profiles of terrorists, activities/threats
Mine the data to extract patterns of potential terrorists and predict future
activities and targets
Find the “needle in the haystack” - suspicious needles?
Data integrity is important
Techniques have to SCALE
Data Mining Needs for Counterterrorism:
Real-time Data Mining
36





Nature of data
 Data arriving from sensors and other devices
 Breaking news, video releases, satellite images, surveillance data
 Continuous data streams
 Some critical data may also reside in caches
Rapidly sift through the data and discard unwanted data for later use and
analysis (non-real-time data mining)
Data mining techniques need to meet timing constraints
Quality of service (QoS) tradeoffs among timeliness, precision and accuracy
Presentation of results, visualization, real-time alerts and triggers
Origins of
Privacy Preserving Data Mining



Prevent useful results from mining
 Introduce “cover stories” to give “false” results
 Only make a sample of data available so that an adversary is
unable to come up with useful rules and predictive functions
Randomization/Perturbation
 Introduce random values into the data and/or results
 Challenge is to introduce random values without significantly
affecting the data mining results
 Give range of values for results instead of exact values
Secure Multi-party Computation
 Each party knows its own inputs; encryption techniques used
to compute final results
37
Data Mining and Privacy:
Friends or Foes?





They are neither friends nor foes
Need advances in both data mining and privacy
Data mining is a tool to be used by analysis and decision makers
 Due to also positives and false negatives, need human in the loop
Need to design flexible systems
 Data mining has numerous applications including in security
 For some applications one may have to focus entirely on “pure” data
mining while for some others there may be a need for “privacypreserving” data mining
 Need flexible data mining techniques that can adapt to the changing
environments
Technologists, legal specialists, social scientists, policy makers and privacy
advocates MUST work together
38