Multimedia Analysis of Video Sources
Juan Arraiza Irujo
1
, Montse Cuadros
1
, Naiara Aginako
1
, Matteo Raffaelli
2
,
Olga Kaehm
3
, Naser Damer
3
and Joao P. Neto
4
1
Vicomtech-IK4, Paseo Mikeletegi 57, 20009 Donostia-San Sebastin, Spain
2
Synthema, 56121 Pisa, Italy
3
Fraunhofer Institute for Computer Graphics Research (IGD), 64283 Darmstadt, Germany
4
VoiceInteraction, 1000-029 Lisbon, Portugal
Keywords:
Multimedia Analysis, Video Processing, Multimodal Surveillance, Decision Support Systems.
Abstract:
Law Enforcement Agencies (LEAs) spend increasing efforts and resources on monitoring open sources,
searching for suspicious behaviours and crime clues. The task of efficiently and effectively monitoring open
sources is strongly linked to the capability of automatically retrieving and analyzing multimedia data. This
paper presents a multimodal analytics system, created in cooperation with European LEAs. In particular it
is described how the video analytics subsystem produces a workflow of multimedia data analysis processes.
After a first analysis of video files, images are extracted in order to perform image comparison, classification
and face recognition. In addition, audio content is extracted to perform speaker recognition and multilingual
analysis of text transcripts. The integration of multimedia analysis results allows LEAs to extract pertinent
knowledge from the gathered information.
1 INTRODUCTION
The digital age has produced unparalleled access to a
proliferation of multimedia data. A huge amount of
multimedia data (video, image, audio and text con-
tents) is now available in open sources. The digital
age has also facilitated the growth of organized crime
in the same way that it has reduced barriers for en-
terprises. The richness and quantity of information
available from open sources, if properly gathered and
processed, can provide valuable intelligence. That is
the reason why Law Enforcement Agencies (LEAs)
are becoming more inclined to using open source in-
telligence (OSINT) tools.
This paper presents an OSINT system, created in
cooperation with European LEAs, for the detection
and prevention of organized crime. The system is able
to gather and process raw open source data, heteroge-
neous both for source, format, protocol and language,
with the aim of retrieving social networks. These
are finally explored using Visual Analytics (VA) tech-
nologies. In particular, we focus here on the video
analytics subsystem. We describe the workflow of
multimedia data analysis processes that is generated
by the system once a video has been crawled.
The system crawler allows to retrieve multimedia
contents in four ways: (1) Looking for a parametric
number of documents on the web with a key-words
search; (2) Looking for documents in a given Uniform
Resource Locator (URL) until a parametric depth of
levels, based on specified key-words; (3) Crawling
Online Social Networks (OSNs), in particular Face-
book’s User Generated Contents (UGCs), focusing
on the information contained in posts (textual con-
tent, photos, videos, creation time, etc.); (4) captur-
ing all of the free-to-air signals (television and radio
programs) that are received in a specific location or
environment. Once a video has been crawled (either
from the web, from Facebook or from a tv program),
it is reduced to its base components and the following
analysis workflows are automatically launched:
1. Image Analysis: Images are extracted and com-
pared with a set of reference images, providing a
similarity score and a class belonging probability.
2. Face Recognition: Automatic face recognition
services are applied to the images of a face avail-
able in a video stream for person identification
and verification.
3. Audio Analysis: Audio content is extracted and
346
Arraiza Irujo J., Cuadros M., Aginako N., Raffaelli M., Kaehm O., Damer N. and P. Neto J..
Multimedia Analysis of Video Sources.
DOI: 10.5220/0005126903460352
In Proceedings of the 11th International Conference on Signal Processing and Multimedia Applications (MUSESUAN-2014), pages 346-352
ISBN: 978-989-758-046-8
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
processed for speaker recognition and tracking,
gender and age identification, etc.
4. Text Analysis: Multilingual text analysis is ap-
plied to the text transcripts of the extracted au-
dio content. Natural Language Processing (NLP)
techniques are used to retrieve entity relation-
ships.
In this paper, we will begin by giving a quick
overviewof currently available solutions for multime-
dia analysis. We will then outline the system archi-
tecture, focusing on the different analysis workflows
mentioned above.
2 RELATED WORK
Current solutions regarding the analysis of video, im-
age, audio and text contents from open sources are
not reported here in detail. Regarding complete video
analysis there are known projects such as Kinesense-
VCA
1
, where advanced video analysis technologies
are implemented in order to help investigators in de-
tecting suspicious behaviors in videos. Another rel-
evant solution for video surveillance in the security
domain is the one offered by IKUSI
2
. In addition,
current European projects such as SAVASA
3
or MO-
SAIC
4
are developing solutions for the creation of
video archive search and analysis platforms that also
include semantic tools.
The variety of image analysis platforms is huge.
The most famous and widely used one among them
is Google Image Search, that introduced in 2011 the
Search by Image functionality
5
. Other widely used
applications are TinEye
6
and CamFind
7
, that allow
to find different kinds of information about the input
image.
Regarding text analysis, many platforms are avail-
able. Most of these platforms perform textual analy-
sis in only a few languages, and only a few of these
platforms are endowed with crawling or visualization
capabilities. Additionally, they usually do not include
specific knowledge integrated such as domain ontolo-
gies. The most common platforms are AlchemyAPI
8
,
1
http://www.kinesense-vca.com/product/
2
http://www.ikusi.com/en/sport/solutions/security/
video-surveillance-and-advanced-video
3
http://www.savasa.eu/
4
http://www.mosaic-fp7.eu/
5
http://www.google.com/insidesearch/features/images/
searchbyimage.html
6
https://www.tineye.com/
7
http://camfindapp.com
8
http://www.alchemyapi.com/
GATE
9
, Lexalytics
10
or LanguageWare
11
. An inter-
esting approach is Attensity
12
, which relies mostly on
social media sources. The most similar text analysis
tool to our proposal is the one from BasisTech
13
cov-
ering up to 55 languages, using Google technology.
All these platforms tackle the analysis of diverse
multimedia contents as independent solutions, but the
real challenge appears to be the creation of a unique
platform that integrates the most important multime-
dia analysis technologies. As aforementioned, the
presented system encompasses all the analysis meth-
ods necessary for an integrate processing of the input
multimedia streams. The capability of extracting in-
terrelated knowledge from the input data allowsLEAs
to significantly reduce data collection, evaluation and
integration efforts.
3 METHODOLOGY
3.1 Architecture Description
The OSINT system described in this paper includes
several distributed components. The back-bone of the
system is an Enterpise Service Bus (ESB), which has
been configured to allow integration, communication,
and orchestration of all those independent but mutu-
ally interacting components. These components are
loosely coupled and they integrate using Service Ori-
ented Architecture (SOA) services. Each component
has published a description of the web service and op-
erations it offers using the Web Services Description
Language (WSDL).
It is worth mentioning that the components might
run in different operating systems, might have been
developed in different programming languages, and
might use internally as many repositories or other re-
sources as needed. The overall system relies on each
component to implement and offer what has been de-
fined in its WSDL. As long as those services and op-
erations respond well, the rest of the system does not
need to know how that work has been implemented.
One benefit of this approach is that it allows replac-
ing one component by another one as long as the new
one complies with the defined integration interface,
in other words, as long as it complies with the defined
WSDL.
9
http://gate.ac.uk/
10
http://www.lexalytics.com/
11
http://www-01.ibm.com/software/ebusiness/jstart/
downloads/LanguageWareOverview.pdf
12
http://www.attensity.com/home/
13
http://www.basistech.com/
MultimediaAnalysisofVideoSources
347
E
S
B
ESB
Shared
repository
Crawler
Image
analysis
Social
Media
analysis
Audio
analysis
Text
analysis
Biometric
analysis
Video
analysis
Configuration
Management
Application
End-User
Visual Analytics
Visual Analytics
Figure 1: High level view of the architecture.
Another important component is the common (or
shared) repository. This repository is a Not Only
Structured Query Language (NoSQL) repository and
has three logical views. In the first one the original
version of the crawled content is stored; this view
is known as the original repository. A different log-
ical view stores the normalized version of the crawled
content. The crawler converts each content type to
one of the accepted normalized formats that the anal-
ysis components can process; this second view is
known as the normalized repository. Finally, a third
view stores the output of the analysis modules; this
view is known as the knowledge repository. Besides
the analysis components, the visual analytics compo-
nent also accesses this repository third view.
3.2 Video Analysis
As aforementioned, the video analystics subsystem is
the starting point of the analysis process. In the last
years, the retrieval of highly representative informa-
tion from videos has become one of the main topics
of the research community. Moreover, automatic con-
tent analysis is very important for the efficient inter-
pretation of all the information encompassed in the
multimedia stream (Maybury, 2012). The considera-
tion within the project of video files as the highest rep-
resentative of the compilation of diverse contents en-
genders the necessity for the development of a video
analysis module.
The first step for the consumption of the video
content is the demultiplexation of the diverse nature
tracks and their transcodification to more appropriate
formats in order to ease the analysis. These consid-
ered tracks are classified as text, audio and images
which represent the input units for the other analy-
sis modules deployed in the project. While audio and
text contents are directly obtained after the demulti-
plexation of the video file, the extraction of the rel-
evant images for the analysis constitute a trade-off.
The implemented Shot Boundary Detection (SBD)
(Zhang et al., 2012)(Zhang et al., 1993) and Best
Frame Extraction (BFE) (Rui et al., 1998)(Ejaz et al.,
2012)(Sun and Fu, 2003) approaches tackle this prob-
lem.
Video content segmentation or structuring has
been defined as the hierarchical decomposition of
videos into units. In this solution, shots are consid-
ered as the representative units of the video content,
so SBD detectors have been implemented to extract
these basic units. Two different detectors are available
in the platform, the first based on contiguous frames
histogram similarity analysis and the second based on
Discrete Cosine Transform (DCT) components calcu-
lated for different compression and codification pur-
poses. Once the shots are identified, one frame for
each shot is selected for the further image analysis
to dismiss the high data redundancy in videos, which
slows down the analysis process. The frame is se-
lected considering the information contained within
each shot. To that end, a feature vector is calculated
for all the frames and the one with the minimum dis-
tance to the rest of the frames is extracted.
Once the information of the video content has
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
348
Figure 2: High level view of the video analytics subsystem architecture.
been separated, more specific analysis for each type
of content is invoked to give the platform the most
concise information about the multimedia content.
3.3 Image Analysis on Video Frames
Once a video has been processed, the most significant
videoimages (best frames) are extracted and saved in
the normalized repository of the system, becoming
available for the image analysis module.
There are two functionalities related with image
analysis, image comparison and image classification.
As mentioned above in 3.1, before launching a Re-
search Line (RL), the user configures the system and
can introduce among other parameters different sets
of images. One kind of set represents images that the
user wants to detect in collected data by image com-
parison, and the other one is used as the training set
for the classification process. All the images belong-
ing to the same classification set contain objects of the
same class.
Once video analysis has been performed and the
best frames have been uploaded into the normal-
ized repository, the orchestrator calls the image anal-
ysis module to process them. First of all, feature
points and their descriptors are extracted for each im-
age (Tuytelaars and Mikolajczyk, 2008)(Bay et al.,
2006)(Lowe, 2003). Depending on the process func-
tion, the module loads the RefModel from the Knowl-
edge Repository which contains the descriptors or the
class model of the already preprocessed Reference
Sets. Afterwards, a comparison between the input im-
age and the RefSet images or the classification of the
input image is performed (Chen et al., 2001). Based
on the calculated descriptor and feature extraction al-
gorithm implemented, a matching algorithm calcu-
lates a similarity score that indicates the visual sim-
ilarity level between images, and the classification al-
gorithm decides if the input image belongs to the re-
quired reference classes or not. The resulted scores
along with the IDs of the compared images or, in
the case of the classification function, the decision
whether the input image belongs to the inquired class
or not, are sent to the Knowledge Repository as an
Extensible Markup Language (XML) file that will be
available to the system.
In this way users can detect required images not
only among other images but also broaden the results
looking inside the videos crawled by the system as
additional information.
3.4 Face Recognition on Video Frames
In the context of the discussed solution, biometrics
is used to search for predefined subjects of interest
within the processed videos. Face biometrics tech-
nology is utilized in this task as it represents the most
clearly available biometric characteristic in such a
scenario, as well as being one of the most thoroughly
studied and widely accepted biometric solutions. A
biometric solution consists of two processes, the en-
rolment and the recognition. Enrolment defines a ref-
erence model for the subject of interest where the
recognition process identifies or verifies the identity
of the captured subject. In the following, both pro-
cesses are discussed.
For enrolment, a number of face images are used
to create a biometric reference for the subject of inter-
est. Those images can be different frames of a video
sequence or static images of the same subject. The
main face is detected in each image using multi-scale
sliding window detection as proposed by Viola and
Jones (Viola and Jones, 2001). The detected faces
are normalized then passed on to the pose alignment
module. Pose alignment here is based on the unsuper-
vised joint alignment of complex images presented by
Huang et. al. (Huang et al., 2007). A binary feature
vector based on the Local Binary Linear Discriminant
MultimediaAnalysisofVideoSources
349
Figure 3: Overview of the proposed face recognition in video solution.
Analysis (LBLDA) (Fratric and Ribaric, 2011) is ex-
tracted from each face image. Those feature vectors
are then fused to create a robust feature vector that
represents the subject of interest.
To search for the subject of interest in a video
sequence, faces are detected and tracked across the
frames (when possible). Those faces are normalized,
the face pose is aligned and then passed on to feature
extraction. Each feature vector (resulted from one
face image) is compared to the reference face model
resulting in a similarity score value. The score val-
ues related to the tracked face in different frames (if
available) are fused using simple combination score-
level fusion rules (Damer et al., 2013). The result-
ing fused score indicates the probability that the pro-
cessed video (or image) contains the face of the sub-
ject of interest. An overview on the proposed biomet-
ric solution is presented in Figure 3.
3.5 Speaker Identification
In the workflow of the proposed solution, speaker
identification is performed on the audio crawled or
extracted from a video and saved in the shared repos-
itory.
The algorithm creates models for a set of speakers
associated to the specific RL and searches for those
speakers in all audio files processed by the system.
It is responsibility of the system operator to define
the speakers of interest and to request the creation of
speaker models to the system. When an audio is pro-
cessed, the algorithm loads the current speaker mod-
els, provides audio analysis and outputs an XML file
per audio file with the speaker segments belonging to
the speaker models.
The algorithm works in two steps: creating mod-
els for specific speakers and audio analysis to identify
the speakers previously defined by models. The al-
gorithm is based on a Bayesian Information Criterion
(BIC) segmentation followed by a BIC clustering. It
starts by detecting speaker turns using BIC, where
change points are detected through generalized like-
lihood ratio (GLR), using Gaussians with full covari-
ance matrices. Speech-Non-Speech (SNS) segments
are also modeled with Full Gaussian and compared
with the current speaker Full Gaussian. If the BIC
score is lower than 0, then a merge is performed be-
tween the current SNS Full Gaussian and the current
speaker Full Gaussian. If the BIC score is higher than
0, then a new speaker Full Gaussian is created and a
hierarchical clustering is performed for the previous
speaker cluster. In our hierarchical clustering algo-
rithm, the current speaker cluster, provided by turn
detection and modeled with Full Gaussian, is com-
pared with the clusters obtained so far. This method
allows on-line processing of the clusters.
The speaker identification component works af-
ter the speaker clustering. Our speaker identification
component uses the low-dimensionality total variabil-
ity factors (i-vector) produced by the Total Variability
technique to model known speaker identities (Dehak
et al., 2011). Since this component works on-line,
every time an unseen speaker starts talking, the com-
ponent is incapable of knowing the speaker identity
immediately. To overcome this problem, a first es-
timate for his/her identity is produced(if he/she is a
known speaker) after 10 seconds of speech and a final
identity estimation after 30 seconds. Since the zero
and first-order sufficient statistics (and the respective
i-vectors) from Total Variability are associated with
the cluster, the speaker information is immediately
available whenevera cluster with a known identity ap-
pears.
As output of the algorithm an XML file is gener-
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
350
Figure 4: Overview of text analysis architecture.
ated with the timing information of all speaker clus-
ters and the IDs associated to the segments whose
speakers are registered and of which the system has
models.
3.6 Text Analysis: Entity and Entity
Relationship Recognition
In the workflow of the proposed solution, textanalysis
is performed on the audio extracted and transcribed
from videos and saved in the shared repository.
The text analysis module is the responsible for the
analysis in different languages of any input given in
the project. The languages considered are Catalan,
Spanish, English, French, Italian, Portugues, German,
Romanian, Basque, Chinese, Japanese, Hebrew, Ara-
bic and Russian.
The workflow used in particular for the last part of
the workflow chain focused on the text analysis con-
tains 4 different core components. The first compo-
nent named language detection detects the language
of the document. The second component named Part-
of-Speech (POS) tagger analyzes the morphosynthac-
tic category of each word in a document. The third
component named Named Entity Recognition and
Classification (NERC) (Nadeau and Sekine, 2007)
detects and categorizes the Named Entities (NE) con-
tained in a document. The main clustered categories
are Person, Location, and Organization. Finally, the
fourth and last component uses the extracted NEs to
give an output based on the relationship (weighted-
distance based) between the different entities in the
whole document. This last component is used in
the Visual Analytics component for visualization pur-
poses.
Text analysis is performed in a webservice based
architecture where the different language processors
are host in different servers.
Figure 4 shows the general schema of the Text
Analysis workflow. The language identification com-
ponent detects the document language and sends it to
the language processor for that language.
NE networks as the result of text analysis are en-
coded in a layered XML-based format named Knowl-
edge Annotation Format (KAF) (Bosma et al., 2009),
which is saved in the shared repository once the pro-
cessing has finished.
4 CONCLUSIONS AND FUTURE
WORK
The increasing need of LEAs for the monitoring and
analysis of open sources boosts the implementation
of systems capable of extracting relevant information
from multimodal data. This work presents an effec-
tive platform for the analysis of all these crawled data
and an intuitive results presentation that eases the cues
searching process. The high implication of European
LEAs has bolstered the development of the required
solutions leading the presented system to the achieve-
ment of its goals.
Validation of this OSINT system has not been un-
dertaken yet. The system final version’s integration
is currently ongoing and its validation by LEA end-
users will take place by the end of the project.
In order to validate the results of the system a test
scenario has been defined. LEA end users will run the
test in two parallel researches, one using their exist-
ing tools and another one using the new OSINT sys-
tem. Several metrics have been defined to assess not
only the quantity and quality of the results obtained
by each method, but also the time spent (human ef-
fort) and the total duration of the process.
As aforementioned, the validation process is still
pending, but the undertaken tests have so far high-
lighted the necessity of approaches that nourish the
system with deeper knowledge about the gathered in-
formation. This knowledge is foreseen to be extracted
by the implementation of new analysis modules con-
nected to the CMA. Still more, the integration of
advanced visual analytics algorithms for the results
presentation interface will tackle the current situation
where a huge amount of data does not directly imply
a huge amount of insight.
MultimediaAnalysisofVideoSources
351
REFERENCES
Bay, H., Tuytelaars, T., and Gool, L. V. (2006). Surf:
Speeded up robust features. In In ECCV, pages 404–
417.
Bosma, W., Vossen, P., Soroa, A., Rigau, G., Tesconi,
M., Marchetti, A., Monachini, M., and Aliprandi, C.
(2009). Kaf: a generic semantic annotation format.
In Proceedings of the GL2009 Workshop on Semantic
Annotation.
Chen, Y., Zhou, X., and Huang, T. S. (2001). One-class svm
for learning in image retrieval. pages 34–37.
Damer, N., Opel, A., and Shahverdyan, A. (2013). An
overview on multi-biometric score-level fusion - ver-
ification and identification. In Marsico, M. D. and
Fred, A. L. N., editors, ICPRAM, pages 647–653.
SciTePress.
Dehak, N., Kenny, P., Dehak, R., Dumouchel, P., and Ouel-
let, P. (2011). Front-end factor analysis for speaker
verification. Audio, Speech, and Language Process-
ing, IEEE Transactions on, 19(4):788–798.
Ejaz, N., Tariq, T. B., and Baik, S. W. (2012). Adaptive key
frame extraction for video summarization using an ag-
gregation mechanism. Journal of Visual Communica-
tion and Image Representation, 23(7):1031–1040.
Fratric, I. and Ribaric, S. (2011). Local binary lda for
face recognition. In Proceedings of the COST 2101
European conference on Biometrics and ID manage-
ment, BioID’11, pages 144–155, Berlin, Heidelberg.
Springer-Verlag.
Huang, G. B., Jain, V., and Learned-Miller, E. G. (2007).
Unsupervised joint alignment of complex images. In
ICCV, pages 1–8. IEEE.
Lowe, D. G. (2003). Distinctive image features from scale-
invariant keypoints.
Maybury, M. T. (2012). Multimedia Information Extrac-
tion: Advances in Video, Audio, and Imagery Analysis
for Search, Data Mining, Surveillance and Authoring.
Wiley.
Nadeau, D. and Sekine, S. (2007). A survey of named entity
recognition and classification. Lingvisticae Investiga-
tiones, 30(1):3–26.
Rui, Y., Huang, T., and Mehrotra, S. (1998). Exploring
video structure beyond the shots. In Multimedia Com-
puting and Systems, 1998. Proceedings. IEEE Inter-
national Conference on, pages 237–240.
Sun, Z. and Fu, P. (2003). Combination of color- and object-
outline-based method in video segmentation.
Tuytelaars, T. and Mikolajczyk, K. (2008). K.: Local invari-
ant feature detectors: A survey. FnT Comp. Graphics
and Vision, pages 177–280.
Viola, P. and Jones, M. (2001). Rapid object detection using
a boosted cascade of simple features. In Computer Vi-
sion and Pattern Recognition, 2001. CVPR 2001. Pro-
ceedings of the 2001 IEEE Computer Society Confer-
ence on, volume 1, pages I–511–I–518 vol.1.
Zhang, H., Kankanhalli, A., and Smoliar, S. W. (1993). Au-
tomatic partitioning of full-motion video. Multimedia
systems, 1(1):10–28.
Zhang, J. F., Wei, Z. Q., Jiang, S. M., Li, J., Xu, S. J., and
Wang, S. (2012). An improved algorithm of video shot
boundary detection. Advanced Materials Research,
403:1258–1261.
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
352