**3.2 Case-based reasoning**

In this work package, we will develop novel case-based reasoning methods for a case-based reasoning system that can keep complex multimedia cases based on their different multimedia features and specific event features in a case base so that they can be easily retrieved and applied for new situations. Meta-learning methods based on CBR over the proposed multimedia processing chain will be developed in this WP in order to achieve the best processing results. The case-based reasoning system will consist of novel probabilistic and similarity-based methods. It will provide a wide range of novel similarity measures for the different feature types and representations for identification and similarity determination. Methods for hierarchical organization of the case base will allow very fast retrieval of similar cases. A special taxonomy for similarity determination and measures will be worked out and implemented in the CBR system. It will provide explanation capabilities for similarity, as those will help a forensic data analyst to identify the right reasoning method for his particular problem. This aspect goes along with the training and education aspect for forensic data analysis. Part of this will be self-contained in the chosen methods and realized by the system.

We will also develop learning methods to include new data into the existing cases and summarization of new and old cases into more general cases applicable to a wider range of tasks for further law purposes. The lifetime aspect of such a CBR system will be considered by special case base maintenance methods and modularity of the system architecture.

• Development of the system architecture

The main architecture of the CBR system will be developed, taking into account the different multimedia data types and data representations. The interface to the preprocessing units and the feature extraction units will be defined. The initial case description that can represent the different multimedia data types and data representations will be developed.

• Development and implementation of the case base for the different multimedia sources

The case base that is the heart of a CBR system will be developed and implemented. For the different multimedia-representation, the right database will be chosen as well as the right data structure. The interfaces and the data structure will be defined.

• Development and implementation of similarity measures for the different feature types and representation

An overview about different similarity measures for the different media typerepresentations will be developed. The pros and cons of the similarity measures will be worked out, and novel similarity measures for the respective data types will be developed. Aggregation of similarities of different types will be studied and evaluated.

*Novel Methods for Forensic Multimedia Data Analysis: Part II DOI: http://dx.doi.org/10.5772/intechopen.92548*

In this task we will utilize and develop blind methods for detecting image forgery, that is, methods that use the image function to perform the forgery detection task. The methods will try to identify various traces of tampering and detect them separately. The final decision about the forgery will be carried out by fusion of

In this work package, we will develop novel case-based reasoning methods for a case-based reasoning system that can keep complex multimedia cases based on their different multimedia features and specific event features in a case base so that they can be easily retrieved and applied for new situations. Meta-learning methods based on CBR over the proposed multimedia processing chain will be developed in this WP in order to achieve the best processing results. The case-based reasoning system will consist of novel probabilistic and similarity-based methods. It will provide a wide range of novel similarity measures for the different feature types and representations for identification and similarity determination. Methods for hierarchical organization of the case base will allow very fast retrieval of similar cases. A special taxonomy for similarity determination and measures will be worked out and implemented in the CBR system. It will provide explanation capabilities for similarity, as those will help a forensic data analyst to identify the right reasoning method for his particular problem. This aspect goes along with the training and education aspect for forensic data analysis. Part of this will be self-contained in the

We will also develop learning methods to include new data into the existing cases and summarization of new and old cases into more general cases applicable to a wider range of tasks for further law purposes. The lifetime aspect of such a CBR system will be considered by special case base maintenance methods and modular-

The main architecture of the CBR system will be developed, taking into account the different multimedia data types and data representations. The interface to the preprocessing units and the feature extraction units will be defined. The initial case

• Development and implementation of the case base for the different multimedia

description that can represent the different multimedia data types and data

The case base that is the heart of a CBR system will be developed and implemented. For the different multimedia-representation, the right database will be chosen as well as the right data structure. The interfaces and the data structure

• Development and implementation of similarity measures for the different

An overview about different similarity measures for the different media typerepresentations will be developed. The pros and cons of the similarity measures will be worked out, and novel similarity measures for the respective data types will be developed. Aggregation of similarities of different types will be studied and

results of separate detectors.

*Digital Forensic Science*

**3.2 Case-based reasoning**

chosen methods and realized by the system.

• Development of the system architecture

ity of the system architecture.

representations will be developed.

feature types and representation

sources

will be defined.

evaluated.

**140**

• Development and implementation of a taxonomy over the similarity measures

As an outcome of these tasks, a taxonomy over similarities will be developed. This taxonomy will be represented by a hierarchical concept, and the user will be guided through this taxonomy for his special needs. A conversational strategy will be developed that helps the user to figure out what his needs are.

• Development and implementation of an indexing structure

The indexing structure over the different data types will be developed and implemented.

• Development and implementation of meta-learning methods over image and video processing chain

The preprocessing and feature extraction methods developed will be evaluated, and it will be decided where meta-learning has to be applied. The architecture and CBR methods for meta-learning will be defined. The meta-learning architecture that fits the CBR system will be developed. The meta-learning algorithm will be implemented and evaluated.

• Development and implementation of a generalization mechanism over cases, case-classes and higher-order classes

The methods for meta-learning over case, case-classes and higher-order classes will be defined for the different multimedia data sources. The interaction between these different multimedia-representations will be studied, and methods that can improve the performance will be developed.

• Development of a learning mechanism over similarities

A recent learning mechanism will be studied, and based on that a new learning mechanism will be developed for feature weighting of the different data types and deciding about the correct similarity measure. These strategies will be implemented into the CBR system and evaluated.

• Development and implementation of a life cycle and maintenance function for the case-based reasoning system

The life cycle of a multimedia CBR system will be studied based on the experience the partners have with different data sources. A life cycle and maintenance function that can take this into account will be developed and implemented.

### **3.3 Multimedia feature extraction**

This work package will investigate, define and evaluate feature extraction methods to detect, describe and relate the multimedia data content relevant to forensic activities. The activities will concentrate on different biometric parameters that characterize individuals in terms of appearance, behavior, voice and handwritings, so as to enable the process of detection and recognition.

Features pertaining to face characteristics, distinctive individual marks, morphometric measurements of the body and gait analysis will be the subject of investigation for videos and images. Similarly, the exploration will regard distinctive features from audio signals for speaker's identification and recognition as well as text analysis to extract information pertinent to the forensic investigations.

Particular emphasis will be put on texts whose language shows characteristics deviating from the standard written form: this will be the case of transcriptions of speech recognizers as well as of the language of social media.

• Feature extraction from text

*DOI: http://dx.doi.org/10.5772/intechopen.92548*

*Novel Methods for Forensic Multimedia Data Analysis: Part II*

extraction as far as texts are concerned.

**3.4 Text mining**

**143**

• Feature integration and ontology development

ontology, resulting in a toolbox for feature extraction.

will include Dutch, Italian, Hungarian, English and Bulgarian.

• Linguistic analysis of Twitter data

This task will focus on the definition and extraction of features from running texts carrying relevant content for forensic activities. This goal will be pursued by combining machine learning stochastic algorithms with advanced natural language

respectively, aimed at (a) extracting ontological knowledge from texts to be used in the framework of Task 3.4 for building the feature ontology and (b) recognizing and semantically classifying relevant information in running texts, to be used as features describing individual documents. We will refer to these lines of research as 'ontology learning' and 'feature extraction', respectively. The planned work will mainly consist in the customization and integration of pre-existing software components available from the consortium partners contributing to this task, which will be specialized to meet the specific needs of particularly text. In particular, the main customizations will be concerned with the typology of information to be extracted (also including relational information) as well as with more challenging research topics such as the automatic analysis of texts representative of so-called noncanonical languages or the development of sophisticated technologies devoted to false witness detection. The final result of this task will include automatically extracted ontological knowledge (to be used as input to Task 3.4) as well as tools for feature extraction from texts to be possibly included in the toolbox for feature

All the features developed and collected in the other tasks will be precisely defined and formalized following a coherent feature definition model. This will enable the development of an ontological model to accommodate the different classes of features and give them an easily sharable and reusable standard organization. The final aim is to (i) standardize and homogenize the feature terminology, (ii) collect and disseminate structured classes of features and (iii) support the choice and computation of features according to a method-oriented strategy. Moreover, a library of algorithms will be supplied and linked to such an

The overall goal of this task is the design and development of methods and techniques for supporting human experts in the analysis of social media textual data. In particular, novel methods will be developed to (a) monitor in real time Twitter and identify potential threats including individuals and communities of users who are planning illegal activities and (b) build a dynamic model on Twitter text to forecast the upcoming significant events and emotions of the crowd associated with these events. Different approaches to the analysis of this type of texts will be pursued, including natural language processing (NLP) techniques, whose results will be eventually compared and—possibly—combined. The languages dealt with

In this phase the tweets will be linguistically analyzed: the text will be segmented

in sentences, tokenized, morphologically analyzed and lemmatized. Linguistic preprocessing is needed to extract information from text: however, the linguistic

processing techniques in line with the state of the art in the computational linguistics field. Two main lines of research can be envisaged for this task,

Aiming at recognition in the wild, focus will be given to the definition and verification of features that enable detection and recognition in unconstrained conditions and environments. This means that feature invariance to different condition's changes and robustness to noise will be two fundamental issues that will be tackled. A systematic approach will be used to feature organization. This means that, starting from existing metadata standards for image, video, audio as well as textual data, a formal ontological model will be defined to organize and categorize all the features collected from pertinent literature as well as the ones newly defined. The resulting feature ontology will standardize feature definition and computation, catalog features and model the multimedia data analysis domain. Such an ontology will be integrated with a library of algorithms for the computation of the features considered, resulting in a toolbox for feature extraction.

• Feature extraction from images and videos

This task will focus on the definition and extraction of features to be used in detection, recognition, authentication and tracking of individuals, event analysis and anomaly/novelty detection.

The work will concentrate on biometric and general appearance features. The field of biologically inspired features will be investigated as well, to verify whether methods that try to mimic the human visual capabilities are able to ensure a better performance. This will be particularly addressed to tackle the difficulties of forensic scenarios where it is not possible to make restrictive assumptions about ambient illumination, subject pose, sensor resolution and compression. In particular, for face recognition in the wild, three-dimensional features will also be explored.

The approach based on bags of features will be studied to carry out a first 'skimming off the top' in large repositories to find out only relevant objects pertinent to the case at hand. Local descriptors will be then defined to handle intensity, rotation, scale and affine variations. Improvements based on the inclusion of spatial information will be investigated.

• Feature extraction from audio streams

This task will focus on the definition and extraction of features for audio streams for the purpose of (i) identification/authentication of individuals and (ii) audio event detection.

The following two main lines of research will be followed: (i) Audio diarization where first the audio stream will be first segmented in speech, nonspeech audio (incl. music) and background noise. Then the speech portions of the audio streams will be segmented by speaker turn and identity. Finally the speaker identity will be verified or identified (depending on the scenario). (ii) Computation of the saliency of the audio stream using low-level features to identify surprising audio events. The audio events will be then classified into semantic (ontological) categories that will be used in Task 3.4.

The audio feature extraction package will include generic short-time envelope features (e.g., mel frequency cepstrum coefficients (MFCC), perceptual linear prediction coefficients (PLP), spectral envelope coefficients (SMAC)) as well as time-domain features. In addition, for speaker identification/verification, modulation spectrum and micro-modulation (AM-FM) features will be employed. *Novel Methods for Forensic Multimedia Data Analysis: Part II DOI: http://dx.doi.org/10.5772/intechopen.92548*

### • Feature extraction from text

Particular emphasis will be put on texts whose language shows characteristics deviating from the standard written form: this will be the case of transcriptions of

Aiming at recognition in the wild, focus will be given to the definition and verification of features that enable detection and recognition in unconstrained conditions and environments. This means that feature invariance to different condition's changes and robustness to noise will be two fundamental issues that will be tackled. A systematic approach will be used to feature organization. This means that, starting from existing metadata standards for image, video, audio as well as textual data, a formal ontological model will be defined to organize and categorize all the features collected from pertinent literature as well as the ones newly defined. The resulting feature ontology will standardize feature definition and computation, catalog features and model the multimedia data analysis domain. Such an ontology will be integrated with a library of algorithms for the computation of the features

This task will focus on the definition and extraction of features to be used in detection, recognition, authentication and tracking of individuals, event analysis

The work will concentrate on biometric and general appearance features. The field of biologically inspired features will be investigated as well, to verify whether methods that try to mimic the human visual capabilities are able to ensure a better performance. This will be particularly addressed to tackle the difficulties of forensic scenarios where it is not possible to make restrictive assumptions about ambient illumination, subject pose, sensor resolution and compression. In particular, for face

recognition in the wild, three-dimensional features will also be explored.

The approach based on bags of features will be studied to carry out a first 'skimming off the top' in large repositories to find out only relevant objects pertinent to the case at hand. Local descriptors will be then defined to handle intensity, rotation, scale and affine variations. Improvements based on the inclusion

This task will focus on the definition and extraction of features for audio streams for the purpose of (i) identification/authentication of individuals and (ii) audio

The following two main lines of research will be followed: (i) Audio diarization where first the audio stream will be first segmented in speech, nonspeech audio (incl. music) and background noise. Then the speech portions of the audio streams will be segmented by speaker turn and identity. Finally the speaker identity will be verified or identified (depending on the scenario). (ii) Computation of the saliency of the audio stream using low-level features to identify surprising audio events. The audio events will be then classified into semantic (ontological) categories that will

The audio feature extraction package will include generic short-time envelope features (e.g., mel frequency cepstrum coefficients (MFCC), perceptual linear prediction coefficients (PLP), spectral envelope coefficients (SMAC)) as well as time-domain features. In addition, for speaker identification/verification,

modulation spectrum and micro-modulation (AM-FM) features will be employed.

speech recognizers as well as of the language of social media.

considered, resulting in a toolbox for feature extraction.

• Feature extraction from images and videos

and anomaly/novelty detection.

*Digital Forensic Science*

of spatial information will be investigated.

event detection.

be used in Task 3.4.

**142**

• Feature extraction from audio streams

This task will focus on the definition and extraction of features from running texts carrying relevant content for forensic activities. This goal will be pursued by combining machine learning stochastic algorithms with advanced natural language processing techniques in line with the state of the art in the computational linguistics field. Two main lines of research can be envisaged for this task, respectively, aimed at (a) extracting ontological knowledge from texts to be used in the framework of Task 3.4 for building the feature ontology and (b) recognizing and semantically classifying relevant information in running texts, to be used as features describing individual documents. We will refer to these lines of research as 'ontology learning' and 'feature extraction', respectively. The planned work will mainly consist in the customization and integration of pre-existing software components available from the consortium partners contributing to this task, which will be specialized to meet the specific needs of particularly text. In particular, the main customizations will be concerned with the typology of information to be extracted (also including relational information) as well as with more challenging research topics such as the automatic analysis of texts representative of so-called noncanonical languages or the development of sophisticated technologies devoted to false witness detection. The final result of this task will include automatically extracted ontological knowledge (to be used as input to Task 3.4) as well as tools for feature extraction from texts to be possibly included in the toolbox for feature extraction as far as texts are concerned.

• Feature integration and ontology development

All the features developed and collected in the other tasks will be precisely defined and formalized following a coherent feature definition model. This will enable the development of an ontological model to accommodate the different classes of features and give them an easily sharable and reusable standard organization. The final aim is to (i) standardize and homogenize the feature terminology, (ii) collect and disseminate structured classes of features and (iii) support the choice and computation of features according to a method-oriented strategy. Moreover, a library of algorithms will be supplied and linked to such an ontology, resulting in a toolbox for feature extraction.

### **3.4 Text mining**

The overall goal of this task is the design and development of methods and techniques for supporting human experts in the analysis of social media textual data. In particular, novel methods will be developed to (a) monitor in real time Twitter and identify potential threats including individuals and communities of users who are planning illegal activities and (b) build a dynamic model on Twitter text to forecast the upcoming significant events and emotions of the crowd associated with these events. Different approaches to the analysis of this type of texts will be pursued, including natural language processing (NLP) techniques, whose results will be eventually compared and—possibly—combined. The languages dealt with will include Dutch, Italian, Hungarian, English and Bulgarian.

• Linguistic analysis of Twitter data

In this phase the tweets will be linguistically analyzed: the text will be segmented in sentences, tokenized, morphologically analyzed and lemmatized. Linguistic preprocessing is needed to extract information from text: however, the linguistic

analysis of small-sized texts, such as tweets, is nowadays a challenging task. It is a widely acknowledged fact that NLP systems, typically trained on newswire texts, have a drop of accuracy when tested against these kinds of texts. In tweets, punctuation and capitalization are often inconsistent, slang and technical jargon are widely exploited, and no-canonical syntactic structures frequently occur. This task is aimed at devising and testing domain adaptation methods to allow the NLP tools to achieve reliable results on these types of texts.

manual labor should be used to extract keywords, identify priorities among attributes, etc. In the later phases we aim at automating this process and the previously discussed visualization methods stay useful for explaining the

To provide more effective results, automatic identification of upcoming threatening events is essential. We will develop methods to identify events unknown beforehand, especially with potential interest to legal forces. We plan to build a dynamic model on Twitter text to forecast the upcoming significant events and emotions of the crowd associated with these events. While there can be many events with a strong presence in social media, some of them would have stronger negative emotions associated with them. These events are candidates that may have

The huge amount of CCTV systems has increased the importance of video and image evidence in forensic labs. In order to retrieve the frames of interests, human experts usually spend a lot of time to examine hours and hours of video sequences. The low quality of the images due to high compression algorithm or bad light conditions or video coming from different record source multiplexed in a unique streaming makes the problem quite difficult to tackle. Another important aspect is the great number of native file formats of CCTV video and the difficulty to create a working copy of the data from a format conversion. This WP will develop novel automatic video processing tools aimed at supporting the expert in selecting the portions of videos that might contain the interesting facts he/she is looking for. In particular, techniques for people identification based on face recognition will be devised. As people may appear in nonfrontal poses, we will exploit the presence of multiple cameras in a given area to track a person and make the identification when

To reconstruct the dynamic of the event of interest, the expert may require a lot of time spent to examine hours and hours of video sequences. A lot of factors can make the operation quite difficult in the absence of a suitable automatic software: low quality of the images due to high compression algorithm, bad light conditions or video coming from different record source multiplexed in an unique streaming. In addition, as multiple native file formats for CCTV video have been produced, the expert should also produce working copies of the data by normalizing the video

This task will be aimed to develop a tool that automatically separates video sequences coming from different cameras and converts the video sequence in a common format that can be used for further processing steps, e.g., feature

This task aims at producing a semi-automatic tool to assist the human expert in selecting the most meaningful frames. Due to the huge quantity of manufacturers operating in the CCTV marketplace, there are a broad range of system for retrieve and export images from compressed video. This task will analyze the most common

automatically made decisions as well as validate them.

*Novel Methods for Forensic Multimedia Data Analysis: Part II*

*DOI: http://dx.doi.org/10.5772/intechopen.92548*

• Forecasting events relevant to legal forces

criminal nature or significant social consequences.

**3.5 Video analysis**

the face is in frontal position.

format.

extraction.

**145**

• Frame selection

• Video sequence analysis

• Keyword extraction

In the first phase, keywords for early warning indicators on certain selected types of crimes will be developed. Traditional fully automated keyword extraction techniques have shown to perform poorly on small-sized texts. We will consider semi-automated keyword extraction methods. In particular, we will start from a domain specific set of keywords for football hooliganism provided by experienced police officers. This collection will unavoidably be incomplete; however, generic keywords in this set can help us zoom in on a subset of the entire dataset. Starting from this manually built set of keywords, we will develop methods to extract other relevant keywords. This will be done by exploring different strategies, also based on NLP techniques. The approaches that generate more reliable results will be used for continuously extending the set of keywords. Some of these methods will also be exploited in the feature extraction process from text carried out in the framework of WP3.

### • Visualization

In the next phase, the Twitter feeds which remained after applying the early warning indicators should be visualized in an intuitive to interpret manner such that police officers can swiftly zoom in on potentially dangerous conversations. We plan to first use self-organizing maps which are built using a training algorithm that allows for incorporating user-defined and automatically inferred attribute priorities. The SOM will partition the collection of tweets into risk areas. The user can choose to group tweets based on the person who wrote them, and the map will show in this case the distribution of persons. If the user would like to analyze in detail a person, a Twitter conversation or a group of persons, we intend to provide functionality such that he or she can select an object or a collection of objects of interest and analyze it with a formal concept analysis and its temporal variant.

• Twitter data analysis

An important step is to identify communities of users in these large amounts of Twitter data. We hereby extend our analysis methods from object-attribute data to object-object data. Fuzzy or probabilistic measures depicting the strength of the relation between individual objects should be used to preprocess the data in order to cope with scalability issues. Based on these distance-related measures, the data is segmented, and strongly related subcommunities are extracted. These subcommunities can be investigated with traditional social network analysis methods, complemented by temporal concept analysis, temporal relational semantic systems, etc. in order to infer its threat level and the role of the actors in the network. To facilitate this phase, we need linguistic methods which will be used to extract relational attributes from Twitter data and can complement the keyword attributes.

• Development of text classification and regression techniques

Finally, automated classification and regression techniques should be developed to partially automate the suspect identification process. Initially a large amount of

### *Novel Methods for Forensic Multimedia Data Analysis: Part II DOI: http://dx.doi.org/10.5772/intechopen.92548*

manual labor should be used to extract keywords, identify priorities among attributes, etc. In the later phases we aim at automating this process and the previously discussed visualization methods stay useful for explaining the automatically made decisions as well as validate them.

• Forecasting events relevant to legal forces

To provide more effective results, automatic identification of upcoming threatening events is essential. We will develop methods to identify events unknown beforehand, especially with potential interest to legal forces. We plan to build a dynamic model on Twitter text to forecast the upcoming significant events and emotions of the crowd associated with these events. While there can be many events with a strong presence in social media, some of them would have stronger negative emotions associated with them. These events are candidates that may have criminal nature or significant social consequences.

## **3.5 Video analysis**

analysis of small-sized texts, such as tweets, is nowadays a challenging task. It is a widely acknowledged fact that NLP systems, typically trained on newswire texts, have a drop of accuracy when tested against these kinds of texts. In tweets,

punctuation and capitalization are often inconsistent, slang and technical jargon are widely exploited, and no-canonical syntactic structures frequently occur. This task is aimed at devising and testing domain adaptation methods to allow the NLP tools

In the first phase, keywords for early warning indicators on certain selected types of crimes will be developed. Traditional fully automated keyword extraction techniques have shown to perform poorly on small-sized texts. We will consider semi-automated keyword extraction methods. In particular, we will start from a domain specific set of keywords for football hooliganism provided by experienced police officers. This collection will unavoidably be incomplete; however, generic keywords in this set can help us zoom in on a subset of the entire dataset. Starting from this manually built set of keywords, we will develop methods to extract other relevant keywords. This will be done by exploring different strategies, also based on NLP techniques. The approaches that generate more reliable results will be used for continuously extending the set of keywords. Some of these methods will also be exploited in the feature extraction

In the next phase, the Twitter feeds which remained after applying the early warning indicators should be visualized in an intuitive to interpret manner such that police officers can swiftly zoom in on potentially dangerous conversations. We plan to first use self-organizing maps which are built using a training algorithm that

priorities. The SOM will partition the collection of tweets into risk areas. The user can choose to group tweets based on the person who wrote them, and the map will show in this case the distribution of persons. If the user would like to analyze in detail a person, a Twitter conversation or a group of persons, we intend to provide functionality such that he or she can select an object or a collection of objects of interest and analyze it with a formal concept analysis and its temporal variant.

An important step is to identify communities of users in these large amounts of Twitter data. We hereby extend our analysis methods from object-attribute data to object-object data. Fuzzy or probabilistic measures depicting the strength of the relation between individual objects should be used to preprocess the data in order to cope with scalability issues. Based on these distance-related measures, the data is

subcommunities can be investigated with traditional social network analysis methods, complemented by temporal concept analysis, temporal relational semantic systems, etc. in order to infer its threat level and the role of the actors in the network. To facilitate this phase, we need linguistic methods which will be used to extract relational attributes from Twitter data and can complement the keyword attributes.

Finally, automated classification and regression techniques should be developed to partially automate the suspect identification process. Initially a large amount of

segmented, and strongly related subcommunities are extracted. These

• Development of text classification and regression techniques

allows for incorporating user-defined and automatically inferred attribute

to achieve reliable results on these types of texts.

process from text carried out in the framework of WP3.

• Keyword extraction

*Digital Forensic Science*

• Visualization

• Twitter data analysis

**144**

The huge amount of CCTV systems has increased the importance of video and image evidence in forensic labs. In order to retrieve the frames of interests, human experts usually spend a lot of time to examine hours and hours of video sequences. The low quality of the images due to high compression algorithm or bad light conditions or video coming from different record source multiplexed in a unique streaming makes the problem quite difficult to tackle. Another important aspect is the great number of native file formats of CCTV video and the difficulty to create a working copy of the data from a format conversion. This WP will develop novel automatic video processing tools aimed at supporting the expert in selecting the portions of videos that might contain the interesting facts he/she is looking for. In particular, techniques for people identification based on face recognition will be devised. As people may appear in nonfrontal poses, we will exploit the presence of multiple cameras in a given area to track a person and make the identification when the face is in frontal position.

• Video sequence analysis

To reconstruct the dynamic of the event of interest, the expert may require a lot of time spent to examine hours and hours of video sequences. A lot of factors can make the operation quite difficult in the absence of a suitable automatic software: low quality of the images due to high compression algorithm, bad light conditions or video coming from different record source multiplexed in an unique streaming. In addition, as multiple native file formats for CCTV video have been produced, the expert should also produce working copies of the data by normalizing the video format.

This task will be aimed to develop a tool that automatically separates video sequences coming from different cameras and converts the video sequence in a common format that can be used for further processing steps, e.g., feature extraction.
