Dissertations / Theses on the topic 'Activity detection'

Práce pojednává o využití support vector machines v detekci řečové aktivity. V první části jsou zkoumány různé druhy příznaků, jejich extrakce a zpracování a je nalezena jejich optimální kombinace, která podává nejlepší výsledky. Druhá část představuje samotný systém pro detekci řečové aktivity a ladění jeho parametrů. Nakonec jsou výsledky porovnány s dvěma dalšími systémy, založenými na odlišných principech. Pro testování a ladění byla použita ERT broadcast news databáze. Porovnání mezi systémy bylo pak provedeno na databázi z NIST06 Rich Test Evaluations.

This Thesis presents a Voice Activity Detection (VAD) system, entirely implemented in the analog domain with a 180-nm CMOS technology. The circuit features a current consumption of 0.9 μA from a 1.8-V supply voltage. The VAD system is composed of three main blocks: a preamplifier, a signal energy computation block, and a VAD decision block. The audio signal coming from the microphone is amplified and filtered by a preamplifier that features a variable gain ranging from −12 dB to +12 dB with 6-dB steps and a bandpass transfer function with poles at 300 Hz and 6.8 kHz. The preamplifier has been implemented both with continuous-time resistors to allow large decoupling capacitors at the input, where the gain is set by the resistance ratio, and with switched resistors to reduce the chip area, where the gain is set by capacitance ratio. The second block of the circuit computes the audio signal energy in the analog domain, exploiting the transistor quadratic current-voltage relation to square the signal and integrating the resulting current with a resettable capacitance. The final block produces the VAD signal. In this block the computed signal energy is used for two different purposes: determine the background noise level and the energy average. The noise level is constantly updated and compared with the averaged energy to provide the VAD signal. The measurement results on an integrated prototype demonstrate that the analog VAD can achieve performances comparable with state-of-the-art digital implementations, but with much lower power consumption.
This Thesis presents a Voice Activity Detection (VAD) system, entirely implemented in the analog domain with a 180-nm CMOS technology. The circuit features a current consumption of 0.9 μA from a 1.8-V supply voltage. The VAD system is composed of three main blocks: a preamplifier, a signal energy computation block, and a VAD decision block. The audio signal coming from the microphone is amplified and filtered by a preamplifier that features a variable gain ranging from −12 dB to +12 dB with 6-dB steps and a bandpass transfer function with poles at 300 Hz and 6.8 kHz. The preamplifier has been implemented both with continuous-time resistors to allow large decoupling capacitors at the input, where the gain is set by the resistance ratio, and with switched resistors to reduce the chip area, where the gain is set by capacitance ratio. The second block of the circuit computes the audio signal energy in the analog domain, exploiting the transistor quadratic current-voltage relation to square the signal and integrating the resulting current with a resettable capacitance. The final block produces the VAD signal. In this block the computed signal energy is used for two different purposes: determine the background noise level and the energy average. The noise level is constantly updated and compared with the averaged energy to provide the VAD signal. The measurement results on an integrated prototype demonstrate that the analog VAD can achieve performances comparable with state-of-the-art digital implementations, but with much lower power consumption.

Activity Recognition is concerned with identifying the physical state of a user at a particular point in time. Activity recognition task requires the training of classification algorithm using the processed sensor data from the representative population of users. The accuracy of the generated model often reduces during classification of new instances due to the non-stationary sensor data and variations in user characteristics. Thus, there is a need to adapt the classification model to new user haracteristics. However, the existing approaches to model adaptation in activity recognition are blind. They continuously adapt a classification model at a regular interval without specific and precise detection of the indicator of the degrading performance of the model. This approach can lead to wastage of system resources dedicated to continuous adaptation. This thesis addresses the problem of detecting changes in the accuracy of activity recognition model. The thesis developed a classifier for activity recognition. The classifier uses three statistical summaries data that can be generated from any dataset for similarity based classification of new samples. The weighted ensemble combination of the classification decision from each statistical summary data results in a better performance than three existing benchmarked classification algorithms. The thesis also presents change detection approaches that can detect the changes in the accuracy of the underlying recognition model without having access to the ground truth label of each activity being recognised. The first approach called `UDetect' computes the change statistics from the window of classified data and employed statistical process control method to detect variations between the classified data and the reference data of a class. Evaluation of the approach indicates a consistent detection that correlates with the error rate of the model. The second approach is a distance based change detection technique that relies on the developed statistical summaries data for comparing new classified samples and detects any drift in the original class of the activity. The implemented approach uses distance function and a threshold parameter to detect the accuracy change in the classifier that is classifying new instances. Evaluation of the approach yields above 90% detection accuracy. Finally, a layered framework for activity recognition is proposed to make model adaptation in activity recognition informed using the developed techniques in this thesis.

Spam, the distribution of unsolicited bulk email, is a big security threat on the Internet. Recent studies show approximately 70-90% of the worldwide email traffic—about 70 billion messages a day—is spam. Spam consumes resources on the network and at mail servers, and it is also used to launch other attacks on users, such as distributing malware or phishing. Spammers have increased their virulence and resilience by sending spam from large collections of compromised machines (“botnets”). Spammers also make heavy use of URLs and domains to direct victims to point-of-sale Web sites, and miscreants register large number of domains to evade blacklisting efforts. To mitigate the threat of spam, users and network administrators need proactive techniques to distinguish spammers from legitimate senders and to take down online spam-advertised sites. In this dissertation, we focus on characterizing spam-related activities and developing systems to detect them early. Our work builds on the observation that spammers need to acquire attack agility to be profitable, which presents differences in how spammers and legitimate users interact with Internet services and exposes detectable during early period of attack. We examine several important components across the spam life cycle, including spam dissemination that aims to reach users' inboxes, the hosting process during which spammers set DNS servers and Web servers, and the naming process to acquire domain names via registration services. We first develop a new spam-detection system based on network-level features of spamming bots. These lightweight features allow the system to scale better and to be more robust. Next, we analyze DNS resource records and lookups from top-level domain servers during the initial stage after domain registrations, which provides a global view across the Internet to characterize spam hosting infrastructure. We further examine the domain registration process and present the unique registration behavior of spammers. Finally, we build an early-warning system to identify spammer domains at time-of-registration rather than later at time-of-use. We have demonstrated that our detection systems are effective by using real-world datasets. Our work has also had practical impact. Some of the network-level features that we identified have since been incorporated into spam filtering products at Yahoo! and McAfee, and our work on detecting spammer domains at time-of-registration has directly influenced new projects at Verisign to investigate domain registrations.

Sectra Communications AB has developed a terminal for encrypted communication called the Tiger platform. During voice communication delays have sometimes been experienced resulting in conversational complications.

A solution to this problem, as was proposed by Sectra, would be to introduce voice activity detection, which means a separation of speech parts and non-speech parts of the input signal, to the Tiger platform. By only transferring the speech parts to the receiver, the bandwidth needed should be dramatically decreased. A lower bandwidth needed implies that the delays slowly should disappear. The problem is then to come up with a method that manages to distinguish the speech parts from the input signal. Fortunately a lot of theory on the subject has been done and numerous voice activity methods exist today.

In this thesis the theory of voice activity detection has been studied. A review of voice activity detectors that exist on the market today followed by an evaluation of some of these was performed in order to select a suitable candidate for the Tiger platform. This evaluation would later become the foundation for the selection of a voice activity detector for implementation.

Finally, the implementation of the chosen voice activity detector, including a comfort noise generator, was done on the platform. This implementation was based on the special requirements of the platform. Tests of the implementation in office environments show that possible delays are steadily being reduced during periods of speech inactivity, while the active speech quality is preserved.

In this publication the results of Principal Component Analysis (PCA) of finger movements electrocorticography (ECoG) are presented. Eigenvalues configuration was analyzed for ECoG with and without any motor activity. PCA components of ECoG can be separated into motor activity and background parts, enabling spatial localization of motor activity areas in future.

A conversational robot will in many cases have todeal with multi-party spoken interaction in which one or morepeople could be speaking simultaneously. To do this, the robotmust be able to identify the speakers in order to attend to them.Our project has approached this problem from a visual pointof view where a Convolutional Neural Network (CNN) wasimplemented and trained using video stream input containingone or more faces from an already existing data set (AVA-Speech). The goal for the network has then been to for eachface, and in each point in time, detect the probability of thatperson speaking. Our best result using an added Optical Flowfunction was 0.753 while we reached 0.781 using another pre-processing method of the data. These numbers correspondedsurprisingly well with the existing scientific literature in thearea where 0.77 proved to be an appropriate benchmark level.
En social robot kommer i många fall tvingasatt hantera konversationer med flera interlokutörer och därolika personer pratar samtidigt. För att uppnå detta är detviktigt att roboten kan identifiera talaren för att i nästa ledkunna bistå eller interagera med denna. Detta projekt harundersökt problemet med en visuell utgångspunkt där ettFaltningsnätverk (CNN) implementerades och tränades medvideo-input från ett redan befintligt dataset (AVA-Speech).Målet för nätverket har varit att för varje ansikte, och i varjetidpunkt, detektera sannolikheten att den personen talar. Vårtbästa resultat vid användning av Optical Flow var 0,753 medanvi lyckades nå 0,781 med en annan typ av förprocessering avdatan. Dessa resultat motsvarade den existerande vetenskapligalitteraturen på området förvånansvärt bra där 0,77 har visatsig vara ett lämpligt jämförelsevärde.
Kandidatexjobb i elektroteknik 2020, KTH, Stockholm

Thesis: M. Eng., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2018.
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Cataloged from student-submitted PDF version of thesis.
Includes bibliographical references (pages 87-90).
The task of producing a Voice Activity Detector (VAD) that is robust in the presence of non-stationary background noise has been an active area of research for several decades. Historically, many of the proposed VAD models have been highly heuristic in nature. More recently, however, statistical models, including Deep Neural Networks (DNNs) have been explored. In this thesis, I explore the use of a lightweight, deep, recurrent neural architecture for VAD. I also explore a variant that is fully end-to-end, learning features directly from raw waveform data. In obtaining data for these models, I introduce a data augmentation methodology that allows for the artificial generation of large amounts of noisy speech data from a clean speech source. I describe how these neural models, once trained, can be deployed in a live environment with a real-time audio stream. I find that while these models perform well in their closed-domain testing environment, the live deployment scenario presents challenges related to generalizability.
by Matthew McEachern.
M. Eng.

The work described in this thesis has the goal of detecting enzyme activity on a Self-Assembled Monolayer (SAM). This required the synthesis of novel chemical tools and biochemical evaluation of these tools on a surface. A solid phase synthesis of novel functionalised alkanethiol-oliogethyleneglycols (AT-OEGs) has been described. This methodology has allowed the synthesis of a variety of functionalised AT-OEGs, which can form SAMs which are functionalised with (bio)molecules. This synthesis was used to construct AT-OEGs conjugated to peptidic substrates of the enzymes thrombin and sortase and an AT-OEG which was conjugated to a redox probe, methylene blue. The AT-OEGs that are conjugated to sortase substrate peptides have been used to detect enzyme activity on a SAM. This system represents a model system, which can inform future studies on clinically relevant enzymes. The AT-OEGs that are conjugated to thrombin substrates have yet to be used to detect thrombin activity. The AT-OEG conjugated to methylene blue has been used in preliminary biosensing studies and could potentially be used as a novel mode of biosensing in the future.

L’activité cérébrale se traduit par des courants ioniques circulant dans le réseau neuronal.La compréhension des mécanismes cérébraux implique de sonder ces courants, via des mesures électriques ou magnétiques, couvrant différentes échelles spatiales. A l’échelle cellulaire, les techniques d’électrophysiologie sont maitrisées depuis plusieurs décennies, mais il n’existe pas actuellement d’outils de mesure locale des champs magnétiques engendrés par les courants ioniques au sein du réseau neuronal. La magnéto-encéphalographie(MEG) utilise des SQUIDs(Superconducting QUantum Interference Devices)fonctionnant à très basse température, placés en surface du crâne, qui fournissent une cartographie des champs magnétiques mais dont la résolution spatiale est limitée du fait de la distance séparant les capteurs des cellules actives. Le travail présenté dans cette thèse propose de développer des capteurs magnétiques à la fois suffisamment sensibles pour être capable de détecter le champ magnétique extrêmement faible générés par les courants neuronaux (de l’ordre de 10⁻⁹ T), et dont la géométrie est adaptable aux dimensions des cellules, tout en fonctionnant à température ambiante. Ces capteurs,basés sur l’effet quantique de magnétorésistance géante (GMR, sont suffisamment miniaturisables pour être déposés à l’extrémité de sondes d’une finesse de l’ordre de 100 μm. L’utilisation de capteurs GMR pour la mesure de signaux biomagnétiques fut d’abord testée lors d’expériences in-vitro, réalisées sur le muscle soléaire de souris. Ce système biologique a été choisi pour sa simplicité,rendant la modélisation accessible, ainsi que pour sa robustesse, permettant d’avoir des résultats fiables et reproductibles. Le parfait accord entre les prédictions théoriques et les signaux magnétiques mesurés valide cette technologie. Enfin, des expériences in vivo dans le cortex visuel du chat ont permis de réaliser la toute première mesure de la signature magnétique de potentiels d’action générés par des neurones corticaux, ouvrant la voie à la magnétophysiologie
Information transmission in the brain occurs through ionic currents flowing inside the neuronal network. Understanding how the brain operates requires probing this electrical activity by measuring the associated electric or magnetic field. At the cellular scale, electrophysiology techniques are well mastered, but there is no tool to perform magnetophysiology. Mapping brain activity through the magnetic field generated by neuronal communication is done via magnetoencephalography (MEG). This technique is based on SQUIDs (Superconducting Quantum Interference Devices) that operate at liquid Helium temperature. This parameter implies to avoid any contact with living tissue and a shielding system that increases the distance between the neurons and the sensors, limiting spatial resolution. This thesis work aims at providing a new tool to performmagnetic recordings at the neuronal scale. The sensors developed during this thesis are based on the Giant Magneto-Resistance (GMR) effect. Operating at room temperature, they can be miniaturize and shaped according to the experiment, while exhibiting a sensitivity that allows to measure amplitude of 10⁻⁹ T. Before targeting neurons, the use of GMR-based sensors for magnetic recordings of biological activity has been validated through invitro experiments on the mouse soleus muscle. This biological system has been chosen because of its simple organization, allowing for a realistic modelling, and for its robustness, in order to get reliable and replicable results. The perfect agreement between the measurements and the theoretical predictions represents a consistent validation of the GMR technology for biological applications. Then a specially adapted needle-shaped probe carrying micron-sized GMR sensors has been developed for in-vivo experiment in cat visual cortex. The very first magnetic signature of action potentials inside the neuropil has been measured, paving the way towards magnetophysiology

Modern automobiles have been proven vulnerable to hacking by security researchers. By exploiting vulnerabilities in the car's external interfaces, attackers can access a car's controller area network (CAN) bus and cause malicious effects. We seek to detect these attacks on the bus as a last line of defence against automotive cyber attacks. The CAN bus standard defines a low-level message structure, upon which manufacturers layer their own proprietary command protocols; attacks must similarly be tailored for their target. This variability makes intrusion detection methods difficult to apply to the automotive CAN bus. Nevertheless, the bus traffic is generated by machines; thus we hypothesize that it can be characterized with machine learning, and that attacks produce anomalous traffic. Our goals are to show that anomaly detection trained without understanding of the message contents can detect attacks, and to create a framework for understanding how the characteristics of a novel attack can be used to predict its detectability. We developed a model that describes attacks based on their effect on bus traffic, informed by a review of published material on car hacking in combination with analysis of CAN traffic from a 2012 Subaru Impreza. The model specifies three high-level categories of effects: attacks that insert foreign packets, attacks that affect packet timing, and attacks that only modify data within packets. Foreign packet attacks are trivially detectable. For timing-based anomalies, we developed features suitable for one-class classification methods. For packet stream data word anomalies, we adapted recurrent neural networks and multivariate Markov model methods to sequence anomaly detection and compared their performance. We conducted experiments to evaluate our detection methods with special attention to the trade-off between precision and recall, given that a practical system requires a very low false alarm rate. The methods were evaluated by synthesizing anomalies within each attack category, parameterized to adjust their covertness. We generalize from the results to enable prediction of detection rates for new attacks using these methods.

Prostate cancer has become one of the leading causes of death and the most frequently diagnosed cancer in men, worldwide. It is highly treatable if detected early, but the current screening methods suffer from poor clinical specificity for prostate cancer, leading to unavoidable misdiagnosis and overtreatment. Elevated telomerase activity has been discovered as an indicator of a variety of cancers, including over 92 % of prostate cancers. High clinical specificity of telomerase activity for prostate cancer compared to other prostate anomalies is an important concept promoted in this thesis. To overcome the quantification complexity of current telomerase activity detection techniques, we have designed and demonstrated the TAME assay (Telomerase Activity Measured Electrochemically): a microfluidic biosensor to detect and measure telomerase activity using an electrochemical technique known as E-DNA. Telomerase activity has been successfully correlated to a TAME parameter, TRAP (Telomeric Repeat Amplification Protocol) product concentration, using two E-DNA schemes: ‘signal-off’ and ‘signal-on’. In terms of the E-DNA signal change, telomerase activity and TRAP product concentrations we investigated are linearly correlated, which is promising for prostate cancer screening and detection. The signal-off scheme exhibits electrochemical signal suppression if telomerase activity is present with alternating cyclic voltammetry (ACV) at 50 Hz. The signal-on scheme shows the reverse effect with square wave voltammetry (SWV) at 150 Hz. ‘Signal-on’ and ‘signal-off’ are transferable by altering SWV’s frequency. The limit of detection of ‘signal-off’ and ‘signal-on’ on tested E-DNA chips using un-purified TRAP samples, originated from un-purified prostate cell extracts, are 55 nM and 10 nM. The TAME assay well-differentiates prostate cancer cells from healthy prostate epithelial cells based on telomerase activity expressions.Due to the trend in medical devices towards miniaturization, portability, low power, and high integration capability, in addition to a microfluidic E-DNA chip, a polymerase chain reaction (PCR) microfluidic chip is under development. The PCR chamber is 2.5 mm X 2.5 mm with the integration of micro-heaters and temperature sensors. PCR chips are designed to achieve heating uniformity, which has been evaluated by thermal imagining. In future, a fully integrated TAME chip including both E-DNA and PCR is anticipated.

Antepartum detection of the fetus at risk of death or damage in-utero remains a major challenge in modern obstetrics. Ultrasonic monitoring of individual fetal biophysical activities (such as fetal heart rate, fetal breathing or fetal body movements) has become widely applied as a method for evaluating fetal well-being. The combined assessment of these fetal activities and the relations between them, however, may well be more useful both in predicting imminent fetal death and in preventing it. The results of ultrasound studies, however, are hampered (e.g. from a safety viewpoint) by the natural periodicity of fetal activities. Fetal phonocardiographic techniques, on the other hand, can be easily used over long time periods and there can be no doubt of their safety. Furthermore, the fetal phonocardiogram may contain more information that fetal heart rate alone. This thesis describes the design and development of a new high-fidelity fetal phono-sensor, based on a piezo-electric PVDF transducer, which offers a completely non-invasive and reliable method of assessing the fetus over the long term. This sensor has been optimised to record faithfully the acoustic output of the fetus and to maximise the signal energy transfer across the maternal abdominal wall, where hitherto this has not been achieved. This is done by matching the compliance of the sensor to that of the maternal abdominal wall. Using a purpose built measuring device, abdominal wall compliance was measured clinically as 3.5 mm/N (averaged over 76 patients). Theoretical and experimental techniques were used to adjust the sensor's compliance to match that of the maternal abdominal wall [to within 4:1], as well as to minimise noise and maximise signal capture. The sensor's force and displacement senstivities were measured as 2183 V/N and 2480 mm/N (much greater than for any present or past phono-sensors). Using a new experimental rig developed to simulate the transmission of the fetal phono-signals through the maternal abdominal wall, the dynamic performance and frequency response of the sensor were also optimised. Clinical studies on 18 patients from 28-41 weeks gestation, showed that the fetal phonocardiogram contains not only fetal heart rate information, but also information about fetal breathing movements (FBM) and fetal body movements (FM), as well as detailed beat-to-beat heart sound interval information. By comparison with real-time ultrasound, various phono-signal patterns were shown to be caused by the fetal activities: regular and cyclic for FBM; and intermittent and noise-like for FM. By using new computerised signal processing techniques in the time domain (such as template-matching and zero-crossing functions), these recorded fetal activities were detected automatically (in over 80% of the time) and their timing periodicities analysed. Frequency domain analysis (using techniques such as the Hilbert transform and cepstral analysis) of the timing periodicities of the fetal heart sounds, showed that diastolic and beat-to-beat time intervals (as wll as their variabilities) are significantly increased during FBM. Fetal heart rate is also decreased [from 147 to 141 bpm] during FBM episodes. Fetal body movements, on the other hand, are associated with significant decreases in systolic, diastolic and beat-to-beat time intervals, whilst their variabilities are increased. Fetal heart rate, in this case, is found to increase [from 144 to 157 bpm]. Although these techniques do not run in real-time at present, they would be capable of doing so if transferred onto fast computers (or transputers). As a result of this work, one is now perhaps in a better position to envisage an automatic real-time fetal phono-based monitoring system for the routine clinical assessment of fetal well-being.

If communicating via a teleconference phone the desired transmitted signal (speech) needs to be crystal clear so that all participants experience a good communication ability. However, there are many environmental conditions that contaminates the signal with background noise, i.e sounds not of interest for communication purposes, which impedes the ability to communicate due to interfering sounds. Noise can be removed from the signal if it is known and so this work has evaluated different ways of estimating the characteristics of the background noise. Focus was put on using speech detection to define the noise, i.e. the non-speech part of the signal, but other methods not solely reliant on speech detection but rather on characteristics of the noisy speech signal were included. The implemented techniques were compared and evaluated to the current solution utilized by the teleconference phone in two ways, firstly for their speech detection ability and secondly for their ability to correctly estimate the noise characteristics. The evaluation process was based on simulations of the methods' performance in various noise conditions, ranging from harsh to mild environments. It was shown that the proposed method showed improvement over the existing solution, as implemented in this study, in terms of speech detection ability and for the noise estimate it showed improvement in certain conditions. It was also concluded that using the proposed method would enable two sources of noise estimation compared to the current single estimation source and it was suggested to investigate how utilizing two noise estimators could affect the performance.

Computer scientists have made ceaseless efforts to replicate cognitive video understanding abilities of human brains onto autonomous vision systems. As video surveillance cameras become ubiquitous, there is a surge in studies on automated activity understanding and unusual event detection in surveillance videos. Nevertheless, video content analysis in public scenes remained a formidable challenge due to intrinsic difficulties such as severe inter-object occlusion in crowded scene and poor quality of recorded surveillance footage. Moreover, it is nontrivial to achieve robust detection of unusual events, which are rare, ambiguous, and easily confused with noise. This thesis proposes solutions for resolving ambiguous visual observations and overcoming unreliability of conventional activity analysis methods by exploiting multi-camera visual context and human feedback. The thesis first demonstrates the importance of learning visual context for establishing reliable reasoning on observed activity in a camera network. In the proposed approach, a new Cross Canonical Correlation Analysis (xCCA) is formulated to discover and quantify time delayed pairwise correlations of regional activities observed within and across multiple camera views. This thesis shows that learning time delayed pairwise activity correlations offers valuable contextual information for (1) spatial and temporal topology inference of a camera network, (2) robust person re-identification, and (3) accurate activity-based video temporal segmentation. Crucially, in contrast to conventional methods, the proposed approach does not rely on either intra-camera or inter-camera object tracking; it can thus be applied to low-quality surveillance videos featuring severe inter-object occlusions. Second, to detect global unusual event across multiple disjoint cameras, this thesis extends visual context learning from pairwise relationship to global time delayed dependency between regional activities. Specifically, a Time Delayed Probabilistic Graphical Model (TD-PGM) is proposed to model the multi-camera activities and their dependencies. Subtle global unusual events are detected and localised using the model as context-incoherent patterns across multiple camera views. In the model, different nodes represent activities in different decomposed re3 gions from different camera views, and the directed links between nodes encoding time delayed dependencies between activities observed within and across camera views. In order to learn optimised time delayed dependencies in a TD-PGM, a novel two-stage structure learning approach is formulated by combining both constraint-based and scored-searching based structure learning methods. Third, to cope with visual context changes over time, this two-stage structure learning approach is extended to permit tractable incremental update of both TD-PGM parameters and its structure. As opposed to most existing studies that assume static model once learned, the proposed incremental learning allows a model to adapt itself to reflect the changes in the current visual context, such as subtle behaviour drift over time or removal/addition of cameras. Importantly, the incremental structure learning is achieved without either exhaustive search in a large graph structure space or storing all past observations in memory, making the proposed solution memory and time efficient. Forth, an active learning approach is presented to incorporate human feedback for on-line unusual event detection. Contrary to most existing unsupervised methods that perform passive mining for unusual events, the proposed approach automatically requests supervision for critical points to resolve ambiguities of interest, leading to more robust detection of subtle unusual events. The active learning strategy is formulated as a stream-based solution, i.e. it makes decision on-the-fly on whether to request label for each unlabelled sample observed in sequence. It selects adaptively two active learning criteria, namely likelihood criterion and uncertainty criterion to achieve (1) discovery of unknown event classes and (2) refinement of classification boundary. The effectiveness of the proposed approaches is validated using videos captured from busy public scenes such as underground stations and traffic intersections.

Thesis (Ph. D.)--University of California, San Diego, 2009.
Title from first page of PDF file (viewed June 16, 2009). Available via ProQuest Digital Dissertations. Vita. Includes bibliographical references (p. 134-141).

Action detection is an attractive area for researchers in computer vision, healthcare, physiotherapy, psychology, and others. Intensive work has been done in this area due to its wide range of applications such as security surveillance, video tagging, Human-Computer Interaction (HCI), robotics, medical diagnosis, sports analysis, interactive gaming, and many others. After the deep learning booming results in computer vision tasks like image classification, many researchers have tried to extend the success of deep learning models to video classification and activity recognition. The research question of this thesis is to study the use of the 2D human poses extracted by a DNN-based model from RGB frames only, for the online activity detection task and comparing it with the state of the art solutions that utilize the human 3D skeletal data extracted by a depth sensor as an input. At the same time, this work showed the importance of input pre-processing and filtering on improving the performance of the online human activity detector. Detecting gym exercises and counting the repetitions in real-time using the human skeletal data versus the 2D poses have been studied in-depth in this work. The contributions of this work are as follows: 1) generating RGB-D dataset for a set of gym exercises, 2) proposing a novel real-time skeleton-based Double Representational RNN (DR-RNN) network architecture for the online action detection, 3) Demonstrating the ability of the proposed model to achieve satisfiable results using pose estimation models applied on RGB frames, 4) introducing a novel learnable exponential filter for the online low latency filtering applications.

Thesis (Ph. D.) -- University of Maryland, College Park, 2004.
Thesis research directed by: Electrical Engineering. Title from t.p. of PDF. Includes bibliographical references. Published by UMI Dissertation Services, Ann Arbor, Mich. Also available in paper.

Lo sviluppo di strumenti robusti ed affidabili capaci di rilevare l’attività elettromiografica muscolare e valutare la forza richiesta al muscolo porterebbe notevoli avanzamenti nelle interfacce uomo-macchina. Le attuali soluzioni basate su elettromiografia non sono pratiche e sono utilizzabili solo da esperti. In primo luogo, è stato sviluppato un algoritmo off-line capace di rilevare in maniera affidabile l’attività elettromiografica attraverso misure contact-less, ed è stato testato con dati sperimentali. Le performance ottenute sono paragonabili a quelle di algoritmi basati su acquisizioni elettromiografiche a contatto diretto. Successivamente, è stato progettato e realizzato un sistema di acquisizione elettromiografica capacitivo basato su un vettore di sensori. Il sistema è composto dai sensori, una parte hardware ed una parte software. Le performance sono state indagate attraverso due diversi setup sperimentali. Nel primo caso, il sistema ha operato in linea con la teoria e con i sistemi di acquisizione elettromiografici facenti parte lo stato dell’arte. Nel secondo caso, fattori di influenza non controllati durante l’esperimento hanno intaccato le misure, i risultati non possono essere considerati statisticamente significativi. Le performance ottenute sono fortemente incoraggianti e ulteriori test sono necessari per provare l’efficacia e la robustezza del sistema progettato. Si è concluso che la rilevazione di attività elettromiografica e la valutazione della forza richiesta al muscolo siano grandezze possibili da stimare attraverso acquisizioni elettromiografiche contact-less, senza alcuna preparazione della cute, né calibrazioni manuali, né un accurato posizionamento degli elettrodi ed attraverso uno strato di tessuto.

Detecting suspicious or unauthorized activities is an important concern for High-Performance Computing (HPC) systems administrators. Automatic classification of programs running on these systems could be a valuable aid towards this goal. This thesis proposes a machine learning model capable of classifying programs running on a HPC system into various types by monitoring metrics associated with different physical and architectural system components. As a specific case study, we consider the problem of detecting password-cracking programs that may have been introduced into the normal workload of a HPC system through clandestine means. Our study is based on data collected from a HPC system called DAVIDE installed at Cineca. These data correspond to hundreds of physical and architectural metrics that are defined for this system. We rely on Principal Component Analysis (PCA) as well as our personal knowledge of the system to select a subset of metrics to be used for the analysis. A time series oversampling technique is also proposed in order to increase the available data related to password-cracking activities. Finally, a deep learning model based on Temporal Convolutional Networks (TCNs) is presented, with the goal of distinguishing between anomalous and normal activities. Our results show that the proposed model has excellent performance in terms of classification accuracy both with balanced (95%) and imbalanced (98%) datasets. The proposed network achieves an F score of 95.5% when training on a balanced dataset, and an AUC-ROC of 0.99 for both balanced and imbalanced data.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2006.
Includes bibliographical references (p. 115-122).
This dissertation develops approaches to rapidly detect malicious network traffic including packets sent by portscanners and network worms. The main hypothesis is that stochastic models capturing a host's particular connection-level behavior provide a good foundation for identifying malicious network activity in real-time. Using the models, the dissertation shows that a detection problem can be formulated as one of observing a particular "trajectory" of arriving packets and inferring from it the most likely classification for the given host's behavior. This stochastic approach enables us not only to estimate an algorithm's performance based on the measurable statistics of a host's traffic but also to balance the goals of promptness and accuracy in detecting malicious network activity. This dissertation presents three detection algorithms based on Wald's mathematical framework of sequential analysis. First, Threshold Random Walk (TRW) rapidly detects remote hosts performing a portscan to a target network. TRW is motivated by the empirically observed disparity between the frequency with which connections to newly visited local addresses are successful for benign hosts vs. for portscanners. Second, it presents a hybrid approach that accurately detects scanning worm infections quickly after the infected local host begins to engage in worm propagation.
(cont.) Finally, it presents a targeting worm detection algorithm, Rate-Based Sequential Hypothesis Testing (RBS), that promptly identifies high-fan-out behavior by hosts (e.g., targeting worms) based on the rate at which the hosts initiate connections to new destinations. RBS is built on an empirically-driven probability model that captures benign network characteristics. It then presents RBS+TRW, a unified framework for detecting fast-propagating worms independently of their target discovery strategy. All these schemes have been implemented and evaluated using real packet traces collected from multiple network vantage points.
by Jaeyeon Jung.
Ph.D.

The threat of data theft posed by self-propagating, remotely controlled bot malware is increasing. Cyber criminals are motivated to steal sensitive data, such as user names, passwords, account numbers, and credit card numbers, because these items can be parlayed into cash. For anonymity and economy of scale, bot networks have become the cyber criminal’s weapon of choice. In 2010 a single botnet included over one million compromised host computers, and one of the largest botnets in 2011 was specifically designed to harvest financial data from its victims. Unfortunately, current intrusion detection methods are unable to effectively detect data extraction techniques employed by bot malware. The research described in this Dissertation Report addresses that problem. This work builds on a foundation of research regarding artificial immune systems (AIS) and botnet activity detection. This work is the first to isolate and assess features derived from human computer interaction in the detection of data theft by bot malware and is the first to report on a novel use of the HTTP protocol by a contemporary variant of the Zeus bot.

Em vista da tentência de se criarem intefaces entre humanos e máquinas que cada vez mais permitam meios simples de interação, é natural que sejam realizadas pesquisas em técnicas que procuram simular o meio mais convencional de comunicação que os humanos usam: a fala. No sistema auditivo humano, a voz é automaticamente processada pelo cérebro de modo efetivo e fácil, também comumente auxiliada por informações visuais, como movimentação labial e localizacão dos locutores. Este processamento realizado pelo cérebro inclui dois componentes importantes que a comunicação baseada em fala requere: Detecção de Atividade de Voz (Voice Activity Detection - VAD) e Localização de Fontes Sonoras (Sound Source Localization - SSL). Consequentemente, VAD e SSL também servem como ferramentas mandatórias de pré-processamento em aplicações de Interfaces Humano-Computador (Human Computer Interface - HCI), como no caso de reconhecimento automático de voz e identificação de locutor. Entretanto, VAD e SSL ainda são problemas desafiadores quando se lidando com cenários acústicos realísticos, particularmente na presença de ruído, reverberação e locutores simultâneos. Neste trabalho, são propostas abordagens para tratar tais problemas, para os casos de uma e múltiplas fontes sonoras, através do uso de informações audiovisuais, explorando-se variadas maneiras de se fundir as modalidades de áudio e vídeo. Este trabalho também emprega um arranjo de microfones para o processamento de som, o qual permite que as informações espaciais dos sinais acústicos sejam exploradas através do algoritmo estado-da-arte SRP (Steered Response Power). Por consequência adicional, uma eficiente implementação em GPU do SRP foi desenvolvida, possibilitando processamento em tempo real do algoritmo. Os experimentos realizados mostram uma acurácia média de 95% ao se efetuar VAD de até três locutores simultâneos, e um erro médio de 10cm ao se localizar tais locutores.
Given the tendency of creating interfaces between human and machines that increasingly allow simple ways of interaction, it is only natural that research effort is put into techniques that seek to simulate the most conventional mean of communication humans use: the speech. In the human auditory system, voice is automatically processed by the brain in an effortless and effective way, also commonly aided by visual cues, such as mouth movement and location of the speakers. This processing done by the brain includes two important components that speech-based communication require: Voice Activity Detection (VAD) and Sound Source Localization (SSL). Consequently, VAD and SSL also serve as mandatory preprocessing tools for high-end Human Computer Interface (HCI) applications in a computing environment, as the case of automatic speech recognition and speaker identification. However, VAD and SSL are still challenging problems when dealing with realistic acoustic scenarios, particularly in the presence of noise, reverberation and multiple simultaneous speakers. In this work we propose some approaches for tackling these problems using audiovisual information, both for the single source and the competing sources scenario, exploiting distinct ways of fusing the audio and video modalities. Our work also employs a microphone array for the audio processing, which allows the spatial information of the acoustic signals to be explored through the stateof- the art method Steered Response Power (SRP). As an additional consequence, a very fast GPU version of the SRP is developed, so that real-time processing is achieved. Our experiments show an average accuracy of 95% when performing VAD of up to three simultaneous speakers and an average error of 10cm when locating such speakers.

Thesis (Ph. D.) -- University of Maryland, College Park, 2006.
Thesis research directed by: Geography. Title from t.p. of PDF. Includes bibliographical references. Published by UMI Dissertation Services, Ann Arbor, Mich. Also available in paper.

Functional Magnetic Resonance Imaging (fMRI) is a neuroimaging technique used to study brain functionality to enhance our understanding of the brain. This technique is based on MRI, a painless, noninvasive image acquisition method without harmful radiation. Small local blood oxygenation changes which are reflected as small intensity changes in the MR images are utilized to locate the active brain areas. Radio frequency pulses and a strong static magnetic field are used to measure the correlation between the physical changes in the brain and the mental functioning during the performance of cognitive tasks.

This master thesis presents approaches for the analysis of fMRI data. The constrained Canonical Correlation Analysis (CCA) which is able to exploit the spatio-temporal nature of an active area is presented and tested on real human fMRI data. The actual distribution of active brain voxels is not known in the case of real human data. To evaluate the performance of the diagnostic algorithms applied to real human data, a modified Receiver Operating Characteristics (modified ROC) which deals with this lack of knowledge is presented. The tests on real human data reveal the better detection efficiency with the constrained CCA algorithm.

A second aim of this thesis was to implement the promising technique of constrained CCA into the software environment SPM. To implement the constrained CCA algorithms into the fMRI part of SPM2, a toolbox containing Matlab functions has been programmed for the further use by neurological scientists. The new SPM functionalities to exploit the spatial extent of the active regions with CCA are presented and tested.

The possibility to detect reliably human brain signals by small sensors can have substantial impact in healthcare, training, and rehabilitation. This Master the- sis studies Electroencephalography (EEG) wireless sensors, and the properties of their signals. The main goal is to investigate the problem of data interpre- tation accuracy. The measurements provided by small wireless EEG sensors show high variability and high noises, which makes it dicult to interpret the brain signals. The analysis is further exacerbated by the diculty in statistical modeling of these signals. This work presents an attempt to a simple statistical modeling of brain signals. Then, based on such a modeling, an optimal data fusion rule of sensors readings is proposed so to reach a high accuracy in the signal's interpretation. An experimental implementation of the data fusion by real EEG wireless sensors is developed. The experimental results show that the fusion rule provides an error probability of nearly 25% in detecting correctly brain signals. It is concluded that substantial improvements have still to be done to understand the statistical properties of signals and develop optimal decision rules for the detection.

A system for automated detection of epileptiform activity in the electroencephalogram (EEG) has been developed and tested on prerecorded data from a range of patients. Epileptiform activity is manifest as spikes in the EEG and consequently the automated detection of spikes in the EEG is an important tool in the diagnosis of epilepsy and is a goal sought by many researchers. The system presented herein is centred around artificial neural networks (ANNs), in particular the multi-layer perceptron (MLP) and the self-organising feature map (SOFM). The MLP is used in the form of an adaptive filter to enhance the presence of epileptiform transients in the EEG while the SOFM is used to form a novel pattern classifier. A modification to the 'standard' calibration technique for the SOFM is proposed based on a method involving Bayesian probabilities. The SOFM allows a large quantity of EEG data to be used to form a pattern classifier in an unsupervised manner. Fuzzy logic is introduced in order to incorporate spatial contextual information in the spike detection process. By using fuzzy logic it has been possible to develop an approximate model of the spatial reasoning performed by an electroencephalographer (EEGer) as opposed to a precise biological model. The human brain is overviewed in terms of its structure, organisation and function. Simplistic mathematical modelling of the neural network of the brain is discussed and ANNs are introduced. After reviewing ANNs in general the perceptron based network is introduced and discussed. The SOFM is introduced and through a number of computer simulations several suggestions are put forward regarding the choice of parameters for training the SOFM. After a review of the literature on spike detection systems, in particular ANN based systems, a multi-stage spike detection system is proposed. There are four stages to the system: spike enhancer, mimetic stage, SOFM and fuzzy logic stage. Each stage of the system is discussed at length and measures of performance are indicated at each stage. The importance of spatial and temporal contextual information is discussed and a method using fuzzy logic is proposed to model the spatial reasoning of an EEGer. The system was trained on 35 epileptiform EEGs containing in excess of 3000 epileptiform events and was tested on a different set of 7 EEGs (6 containing epileptiform activity and 1 'normal') containing 133 epileptiform events. The EEGs consisted of standard clinical recordings with an average length of 22.9 minutes. Preliminary results show that the system has a sensitivity of 59% and a selectivity of 31% with an average false detection rate of 61 per hour. The performance compares well with other leading systems to be found in the literature once the measures of performance obtained in each are case placed in context. Several aspects in the system have been identified for modification which should lead to considerable improvements in performance (e.g., temporal context, improved mimetic stage). The new approach to the spike detection problem presented in this thesis shows that it is possible to form an accurate classifier in a self-organised fashion, thus eliminating the need to accurately label large quantities of data - a weak point in many spike detection systems. Furthermore, the importance of spatial contextual analysis is highlighted showing that it is possible to model the spatial reasoning of an EEGer with a fuzzy logic system, thus eliminating the need to produce accurate models of the process.

This thesis is partly a theoretical introduction to some basic concepts of signal processing such as the Fourier transform, linear time invariant systems and spectral analysis of random signals, both in the continuous and discrete setting. A second part is devoted to theory and applications of echo cancellation and voice activity detection in so called social robotics. Existing methods are presented along with new specialized methods and both are later evaluated.
Vi ger en teoretisk introduktion till grundläggande koncept inom kontinuerlig och diskret signalhantering som Fourier-transformen, linjära tidsinvarianta system och spektralanalys av slumpsignaler. En andra del behandla teori och tillämpningar för ekokansellering och röstdetektion i så kallad social robotik. Existerande metoder presenteras tillsammans med nya specialiserade metoder och samtliga utvärderas

The abrupt expansion of the Internet use over the last decade led to an uncontrollable amount of media stored in the Web. Image, video and news information has ooded the pool of data that is at our disposal and advanced data mining techniques need to be developed in order to take full advantage of them. The focus of this thesis is mainly on developing robust video analysis technologies concerned with detecting and recognizing activities in video. The work aims at developing a compact activity descriptor with low computational cost, which will be robust enough to discriminate easily among diverse activity classes. Additionally, we introduce a motion compensation algorithm which alleviates any issues introduced by moving camera and is used to create motion binary masks, referred to as compensated Activity Areas (cAA), where dense interest points are sampled. Motion and appearance descriptors invariant to scale and illumination changes are then computed around them and a thorough evaluation of their merit is carried out. The notion of Motion Boundaries Activity Areas (MBAA) is then introduced. The concept differs from cAA in terms of the area they focus on (ie human boundaries), reducing even more the computational cost of the activity descriptor. A novel algorithm that computes human trajectories, referred to as 'optimal trajectories', with variable temporal scale is introduced. It is based on the Statistical Sequential Change Detection (SSCD) algorithm, which allows dynamic segmentation of trajectories based on their motion pattern and facilitates their classification with better accuracy. Finally, we introduce an activity detection algorithm, which segments long duration videos in an accurate but computationally efficient manner. We advocate Statistical Sequential Boundary Detection (SSBD) method as a means of analysing motion patterns and report improvement over the State-of-the-Art.

This research addressed one of the hard problems still plaguing the information security profession; detection of network activity deviations from authorized accounts when the deviations are similar to normal network activity. Specifically, when user and administrator type accounts are used for malicious activity, harm can come to the organization. Accurately modeling normal user network activity is hard to accomplish and detecting misuse is a complex problem. Much work has been done in the past with intrusion detection systems, but being able to detect masquerade events with high accuracy and low false alarm rates continues to be an issue. Bayesian networks have been successfully used in the past to reason under certainty by combining prior knowledge with observed data. The use of dynamic Bayesian Networks, such as multi-entity Bayesian network, extends the capability and can address complex problems. The goal of the research was to extend previous research with multi-entity Bayesian networks along with discretization methods to improve the effectiveness of the detection rate while maintaining an acceptable level of false positives. Preprocessing continuous variables has proven effective in prior research but has not been applied to multi-entity Bayesian networks in the past. Five different discretization methods were used in this research. Analysis using receiver operating characteristic curves, confusion matrix, and other comparison methods were completed as part of this research. The results of the research demonstrated that a multi-entity Bayesian network model based on multiple data sources and the relationship between the user attributes could be used to detect unauthorized access to data. The supervised top down discretization methods had better performance related to the overall classification accuracy. Specifically, the class-attribute interdependence maximization discretization method outperformed the other four discretization methods. When compared to previous masquerade detection methods, the class-attribute interdependence maximization discretization method had a comparable true positive rate with a lower false positive rate.

Dissertação apresentada à Universidade Fernando Pessoa como partes dos requisitos para a obtenção do grau de Mestre em Engenharia Informática, ramo de Computação Móvel
Vivemos numa era em que a esperança média de vida tem aumentado, traduzindo-se diretamente num crescimento cada vez mais significativo da população na faixa etária da terceira idade. Este fenómeno faz com que cada vez mais pessoas idosas vivam sozinhas em casa e tenham dificuldade em encontrar pessoas que as possam acompanhar e ajudar no seu dia-a-dia. Reconhece-se ainda que a atividade e o exercício físico são fundamentais para manter e promover a saúde, em particular nesta faixa etária, permitindo em geral melhorar a qualidade de vida e segurança das pessoas. Neste sentido têm vindo a ser desenvolvidos vários sistemas comerciais para monitorizar a atividade de pessoas, em particular idosos, por forma a acompanhar o que estes fazem durante o dia, recolhendo dados estatísticos de atividade bem como permitindo identificar potenciais situações de perigo (e.g., quedas). Contudo, as soluções existentes envolvem normalmente a aquisição de equipamentos com custos geralmente elevados e que baseiam o seu funcionamento na análise de apenas um tipo de informação ou fenómeno físico (e.g., acelerómetro, análise de cenas, etc.), limitando desta forma o resultado e a fidelidade da monitorização. Neste contexto, o trabalho aqui apresentado propõe a utilização de um sistema de monitorização económico, baseado na fusão de vários tipos de informação (e.g., atividade física, localização do utilizador, etc.) recolhida num espaço inteligente preparado para o efeito. Este sistema combina informação recolhida e processada por vários componentes. Utiliza, por exemplo, dispositivos móveis hoje em dia vulgarizados e disponíveis por custos razoáveis (cf. smartphones) e que vêm equipados com um conjunto de sensores e capacidades de processamento adequados; utiliza ainda outros componentes vulgarmente existentes em contextos residenciais (cf. computador e câmaras de baixa resolução apontadas aos locais a monitorizar) e que podem ser integrados e reutilizados na solução proposta. Esta dissertação propõe e descreve a estrutura física e lógica do sistema PADS (Personal Activity Detection System). O protótipo desenvolvido organiza-se em vários componentes talhados para a recolha e fusão de diferentes tipos de informação: i) uma aplicação Android que identifica e regista cinco atividades físicas distintas (cf. parado, andar, correr, deitado e queda); ii) uma aplicação desenvolvida em C++ para deteção da presença do utilizador com base no reconhecimento de faces; iii) uma aplicação que utiliza serviços da nuvem da Google para consultar as tomas de medicamentos do utilizador que se encontram agendadas; iv) uma aplicação em C++ que faz a fusão de toda a informação obtida pelos módulos anteriores e permite determinar a atividade do utilizador. Apresenta-se a arquitetura global do sistema e os vários componentes envolvidos, descrevendo-se com pormenor todos os algoritmos e os detalhes de implementação. Procedeu-se ainda à avaliação do módulo de deteção de atividade bem como da aplicação de fusão. Procurou-se analisar e comparar diferentes técnicas de determinação de atividade que podem ser consideradas como alternativa à opção aplicada neste projeto; analisaram-se ainda outros trabalhos relacionados na deteção de atividade recorrendo a várias fontes de informação, procurando comparar as diferentes vertentes e mais-valias de cada alternativa.
We live in an era in which the average life expectancy has increased, resulting directly in a growth each time more significant of senior population. This phenomenon makes that more and more elderly people live at home alone and have difficulty finding people who can attend and help in their daily lives. It is also recognized that activity and physical exercise are essential to maintain and promote health, in particular in this age group, allowing to, in general, improve the quality of life and people’s safety. In this way, many commercial systems have been developed in order to monitor people’s activity, in particular elderly, to follow what they do during the day, collecting statistical data of activity, as well as identifying potential dangerous situations (e.g. falls). However, the existing solutions involve, usually, the acquisition of equipment with generally high cost and they base their operation in the analysis of just one kind of information or physical phenomenon (e.g. accelerometer, scene analysis, etc.), which limits, in this way, the results and the monitoring fidelity. In this context, the work presented here proposes the use of an economic monitoring system, based in the fusion of several kinds of information (e.g. physical activity, user’s location, etc.) collected in a smart place set for this purpose. This system combines information collected and processed by several components. It uses, for example, mobile devices that are ordinary nowadays and that are available for reasonable costs (cf. smartphones) and that are equipped with a set of sensors and adequate processing abilities; it also uses other components commonly existents in residential contexts (cf. computer and low resolution cameras oriented to the places to monitor) and that can be integrated and reused in the proposed solution. This dissertation proposes and describes the physical and logical structure of the PADS (Personal Activity Detection System) system. The developed prototype is organized in several components that are able to collect and fuse different kinds of information: i) an Android application that identifies and records 5 distinctive physical activities (cf. standing, walking, running, lying down and fall); ii) an application developed in C++ to detect the user’s presence, based in the face recognition; iii) an application that uses Google Cloud services to consult the user’s drug doses that are scheduled; iv) an application in C++ that fuses all the information obtained by the previous modules and allows to determinate the user’s activity. It is presented the global architecture of the system and the various involved components, describing with detail all of the algorithms and the details of implementation. We also proceeded to the evaluation of the detection of activities module, as well as the fusion application. We sought to analyze and compare different techniques of activities determination that could be considered as an alternative to the applied option in this project. We also analyzed other works related to activity detection, recurring to several information sources, trying to compare several sides and strengths of each alternative.
Nous vivons une époque où l'espérance de vie augmente considérablement, tout en ayant comme conséquence directe une croissance importante de la population de personne âgé. Les raisons de ce phénomène font que, de plus en plus de personne âgées vivent seules ayant du mal à trouver des personnes qui puisse surveiller et aider leur vie au quotidien. Aussi, il est reconnu que les activités et l’exercice physique sont essentiels pour maintenir et promouvoir la santé, en particulier dans ce groupe d'âge, ce qui permet en général d'améliorer la qualité de vie et la sécurité des personnes. C’est pour ce fait que l’apparition et le développement de divers systèmes commercial pour la surveillance des personnes sont conçus, en particulier pour les personnes âgées, afin de garder une “empreinte“ de leurs quotidien, recueillant les données des activités, tout en permettent d'identifier des situations potentiellement dangereuses (p. ex., les chutes). Cependant, les équipements existantes sont très couteux à l’achat et fonctionne ayant comme base l’analyse d’un seul et unique type d’information ou phénomène physiques, (p.ex. accéléromètre, analyse de scène, etc.), limitant ainsi le résultat et la précision des même. Dans ce contexte, le travail présenté propose l'utilisation d'un système de surveillance économique, ayant comme base la fusion de différents types d'informations (par exemple, localisation de l'utilisateur, l'activité physique, etc.) recueillie dans un espace intelligent préparé pour cette effet. Ce système assimile les informations réunies et par moyen de plusieurs composants, traites ces mêmes informations. Ce système utilise des appareils mobiles disponibles de nos jours et à des prix accessible au grand public (p.ex. Smartphones, etc.) qui sont équipées d'un ensemble de capteurs et des capacités de traitement d’informations appropriées. Aussi, ce système utilise aussi d’autres composants qui se trouvent couramment dans les maisons de nos jours (tels qu’ordinateur et caméras basse résolution qui sont dirigée vers les locaux a surveillé) et qui peuvent être intégrées et réutilisées dans la solution proposée. Cette dissertation propose et décrit la structure physique et logique du système PADS (Personal Activity Detection System). Le prototype développé est conçus avec différents composants scrupuleusement choisi pour la recueille et fusion des différents types d'informations: i) Une application Android qui identifie et enregistre 5 activités physiques distinctes (debout, marche, courir, couchée et tomber); ii) Une application développée en C++ pour la détection de présence de l'utilisateur, basée sur la reconnaissance de visage; iii) Une application qui utilise les services “Cloud“ de Google pour consulter quelle sont les médicaments, après une correct calendarisation, à prendre par l’utilisateur; IV) Une application en C++ qui permet la fusion de toutes les informations obtenues par les modules précédents et qui permet de déterminer l'activité de l'utilisateur. Il est présenté l'architecture globale du système et les différents composants impliqués, décrivant avec détail tous les algorithmes et les détails de mise en oeuvre. De plus, il fut effectué une évaluation du module de détection de l'activité, tout comme pour l’application de fusion. Le but de cette même évaluation fut analyser et comparer les différentes techniques de détermination de l'activité qui peut être considérée comme une alternative à l'option appliquée dans ce projet. Nous avons également examiné d’autres travaux liés à la détection de mouvement, à l'aide de plusieurs sources d'information, cherchent à comparer les différents aspects et les plus-values de chaque solution.

Passive acoustic mapping (PAM) is a novel method for monitoring ultrasound therapies by mapping sources of acoustic activity, and in most cases cavitation activity, using an array of detectors. Although the range of its applications is indicative of its great potential, clinical adoption is currently hindered by its limited spatial resolution and the inherent difficulty of distinguishing, at depth, between nonlinear signals arising from nonlinear propagation and those arising from processes such as cavitation. The objective of this thesis is to address this limitation, by improving both the detection of the signal-of-interest and the source localisation. An optimum data-adaptive array beamforming algorithm is proposed, Robust Beamforming by Linear Programming (RLPB), which exploits the higher-orderstatistics of the recorded signals, aiming at improving PAM source localisation. Both simulations and in vitro experimentation demonstrated improvement in PAM spatial resolution compared to a previously introduced algorithm, Robust Capon Beamformer. More specifically, under the in vitro conditions examined here, a 22% and 14% increase in the axial and transverse PAM resolution is respectively achieved. In terms of reliable signal-of-interest detection amongst interfering signals, a time-domain data-adaptive parametric model, Sum-of-Harmonics (SOH) model, is developed. This model enables accurate estimation of time-varying-amplitude narrowband components in the presence of broadband signals. Respectively, it can recover a weak broadband signal in the presence of a dominant narrowband component. Compared to conventional comb filtering, SOH model enables PAM of cavitation sources that better reflect their physical location and extent. PAM performance enhancement achieved by combining the proposed beamforming and filtering approaches is assessed in a context where spatial resolution really matters, namely for distinguishing between cavitation activity occurring inside a channel and perivascularly following cavitation-mediated extravasation. Adoption of the proposed method results in more accurate isolation of the broadband emissions from inertially cavitating sources, and more reliable localisation of these sources despite the long source-to-array distance has been observed. Such an improvement to the spatial accuracy of PAM paves the way towards its clinical translation, and in vivo experimentation is the next step for further validation of PAM in conjunction with the proposed methods under clinically relevant conditions.

Recent studies pinpoint visually cued networks of avalanches with MEG/EEG data, and hemodynamic fingerprints of neuronal avalanches as sudden high signal activity peaks in classical fMRI data. However, detection of neural avalanches faces the problem that the data contain a lot of physiological noise making the automatic analysis difficult. The aim of this study was to detect dynamic patterns of brain activity spreads with the use of ultrafast magnetic resonance encephalography (MREG). MREG achieves 10 Hz whole brain sampling, allowing the estimation of the spatial spread of an avalanche. A computational method was developed to separate neuronal avalanches from motion and physiological pulsations, and detect activity avalanches in human brain default mode network (DMN). Brain activity peaks could be identified from parts of the DMN, and normalized MREG data around each peak was extracted individually in order to show dynamic avalanche spreads as videos within the DMN. Individual avalanche videos of specific parts of the DMN were then averaged over a group of subjects. The results indicate that the detected peaks must be parts of activity avalanches, starting from (or crossing) the DMN. To support analyses on large fMRI data, like MREG recordings, also a method and implementation are presented to achieve a thousand fold speed-up for seeking in large compressed NIfTI neuroimaging data files. The method includes the creation of a novel index structure for the compressed data in order to achieve a speed-up of over hundred up to even five thousand, compared to the currently available implementations. By configuring the index structure, one can set an operating point which optimizes the efficiency as speed-up versus index size according to the requirements by the user
Uusimmat tutkimukset osoittavat, että MEG/EEG-datasta on visuaalisesti havaittavissa neuraalisia verkkoja joissa tapahtuu avalanssi-ilmiöitä. Lisäksi klassisessa fMRI-datassa on havaittu neuronaalisiin avalansseihin liittyviä hemodynaamisia jälkiä, jotka ilmenevät äkillisinä voimakkaina piikkeinä datassa. Neuraalisen avalanssin havaitsemisen automatisointi on kuitenkin hyvin haastavaa, koska data sisältää myös merkittäviä fysiologista kohinakomponentteja. Tämän tutkimuksen tavoitteena oli kehittää laskennallinen menetelmä havaita aivojen aktiviteetin leviämisen dynaamisia rakenteita hyödyntäen ultranopeaa magneettisen resonanssin enkefalografiaa (MREG). MREG kykenee saavuttamaan aivojen näytteistyksen 10 Hz taajuudella, mikä mahdollistaa neuraalisen avalanssin spatiaalisen leviämisen havaitsemisen. Työssä kehitettiin menetelmä erottaa neuraalinen avalanssi liikkeen ja fysiologisten pulsaatioiden tuottamista signaalikomponenteista, sekä havaita aktiviteettiavalanssi ihmisaivojen lepotilan aikaisessa neuraalisessa verkossa (default mode network, DMN). Menetelmä identifioi aivojen aktiviteettipiikkejä DMN-verkosta, normalisoi piikkien ympärillä olevan aktiviteettidatan yksilöllisesti ja lopulta esittää avalanssin leviämisen videona. Verkon toiminnan tutkimiseksi yksilölliset avalanssivideot määrätyistä DMN-verkon osista keskiarvoistettiin koehenkilöryhmän ylitse, jolloin ryhmäkäyttäytymisestä pääteltiin identifioitujen piikkien todella liittyvän DMN-verkosta alkaneisiin tai sen ylittäviin avalansseihin. Lisäksi työssä kehitettiin menetelmä nopeuttaa fMRI/MREG-datan käsittelyaikoja merkittävästi, mistä on suurta etua käsiteltäessä kompressoituja NIfTI-muodossa tallennettuja suuria neurokuvantamisen aineistoja. Menetelmä perustuu uudenlaiseen indeksointimenetelmään, jolla kompressoitua aineistoa voidaan selata nopeudella, joka ylittää monisatakertaisesti tai jopa monituhatkertaisesti perinteellisen menetelmän nopeuden. Konfiguroimalla indeksirakenne sopivasti voidaan asettaa toimintapiste menetelmälle siten, että haluttu kompromissi nopeuden ja indeksirakenteen viemän muistitilan kesken saavutetaan

With the ever growing media and music catalogs, tools that search and navigate this data are important. For more complex search queries, meta-data is needed, but to manually label the vast amounts of new content is impossible. In this thesis, automatic labeling of musical instrument activities in song mixes is investigated, with a focus on ways to alleviate the lack of annotated data for instrument activity detection models. Two methods for alleviating the problem of small amounts of data are proposed and evaluated. Firstly, a self-supervised approach based on automatic labeling and mixing of randomized instrument stems is investigated. Secondly, a domain-adaptation approach that trains models on sampled MIDI files for instrument activity detection on recorded music is explored. The self-supervised approach yields better results compared to the baseline and points to the fact that deep learning models can learn instrument activity detection without an intrinsic musical structure in the audio mix. The domain-adaptation models trained solely on sampled MIDI files performed worse than the baseline, however using MIDI data in conjunction with recorded music boosted the performance. A hybrid model combining both self-supervised learning and domain adaptation by using both sampled MIDI data and recorded music produced the best results overall.
I och med de ständigt växande media- och musikkatalogerna krävs verktyg för att söka och navigera i dessa. För mer komplexa sökförfrågningar så behövs det metadata, men att manuellt annotera de enorma mängderna av ny data är omöjligt. I denna uppsats undersöks automatisk annotering utav instrumentsaktivitet inom musik, med ett fokus på bristen av annoterad data för modellerna för instrumentaktivitetsigenkänning. Två metoder för att komma runt bristen på data föreslås och undersöks. Den första metoden bygger på självövervakad inlärning baserad på automatisk annotering och slumpartad mixning av olika instrumentspår. Den andra metoden använder domänadaption genom att träna modeller på samplade MIDI-filer för detektering av instrument i inspelad musik. Metoden med självövervakning gav bättre resultat än baseline och pekar på att djupinlärningsmodeller kan lära sig instrumentigenkänning trots att ljudmixarna saknar musikalisk struktur. Domänadaptionsmodellerna som endast var tränade på samplad MIDI-data presterade sämre än baseline, men att använda MIDI-data tillsammans med data från inspelad musik gav förbättrade resultat. En hybridmodell som kombinerade både självövervakad inlärning och domänadaption genom att använda både samplad MIDI-data och inspelad musik gav de bästa resultaten totalt.

De nos jours, il est important de trouver des solutions pour détecter et prévenir les chutes des personnes âgées. Nous avons proposé un dispositif bas coût à base d’une paire de capteurs thermiques. La contrepartie de ces capteurs bas-coût est leur faible résolution (80x60 pixels), la faible fréquence de rafraîchissement, le bruit et des effets de halo. Nous avons donc proposé quelques approches pour contourner ces inconvénients. Tout d’abord, nous avons proposé une nouvelle méthode de calibration avec une grille adaptée à l’image thermique et une méthodologie assurant la robustesse de l’estimation des paramètres malgré la faible résolution. Ensuite, pour la vision 3D, nous avons proposé une méthode de mise en correspondance stéréo avec une précision sous-pixels (appelée ST pour Subpixel Thermal) composée : 1) d’une méthode robuste d’extraction des caractéristiques basée sur la congruence de phase, 2) d’une mise en correspondance de ces caractéristiques au pixel près, et 3) d’une mise correspondance raffinée en précision sous-pixel basée sur la corrélation de phase locale. Nous avons également proposé une méthode de super-résolution appelée Edge Focused Thermal Super-Resolution (EFTS) qui contient un module d’extraction de contours amenant le réseau de neurones artificiels de se concentrer sur les contours des objets dans les images. Par la suite, pour la détection des chutes, nous avons proposé une nouvelle méthode (TSFD pour Thermal Stereo Fall Détection) basée sur les correspondances stéréo mais sans calibration et un apprentissage de points au sol. Enfin, pour la surveillance des activités des personnes âgées, nous avons exploré de nombreuses approches basées sur l’apprentissage profond pour classer des activités avec une quantité limitée de données d’apprentissage
Nowadays, it is essential to find solutions to detect and prevent the falls of seniors. We proposed a low-cost device based on a pair of thermal sensors. The counterpart of these low-cost sensors is their low resolution (80x60 pixels), low refresh rate, noise, and halo effects. We proposed some approaches to bypass these drawbacks. First, we proposed a calibration method with a grid adapted to the thermal image and a framework ensuring the robustness of the parameters estimation despite the low resolution. Then, for 3D vision, we proposed a threefold sub-pixel stereo matching framework (called ST for Subpixel Thermal): 1) robust features extraction method based on phase congruency, 2) matching of these features in pixel precision, and 3) refined matching in sub-pixel accuracy based on local phase correlation. We also proposed a super-resolution method called Edge Focused Thermal Super-resolution (EFTS), which includes an edge extraction module enforcing the neural networks to focus on the edge in images. After that, for fall detection, we proposed a new method (called TSFD for Thermal Stereo Fall Detection) based on stereo point matching but without calibration and the classification of matches as on the ground or not on the ground. Finally, we explored many approaches to learn activities from a limited amount of data for seniors activity monitoring

De nos jours, les mesures sans contact du signal cardiaque du patient en utilisant le radar Doppler a suscité un intérêt considérable chez les chercheurs, surtout que les électrocardiographes traditionnels avec des électrodes fixes ne sont pas pratiques dans certains cas comme les nourrissons ou les victimes de brûlure. En raison de la sensibilité des micro­ondes à de petits mouvements, le radar a été utilisé comme système de surveillance de l'activité cardio-pulmonaire humaine. Selon l'effet Doppler, un signal de fréquence constante est transmis vers la cible ayant un déplacement variable puis réfléchi. Le signal réfléchit possède une variation de phase par rapport au temps. Dans notre cas, la cible est la poitrine du patient; Le signal réfléchi de la poitrine de la personne contient le signal cardiorespiratoire. Le système est basé sur un analyseur de réseau vectoriel et deux antennes cornet. Le S21 est calculé en utilisant un analyseur de réseau. La variation de phase de S21 contient des informations de l'activité cardio-pulmonaire. Des techniques de traitement sont utilisées pour extraire le signal cardiaque de la variation de la phase de S21 . Cette thèse présente une étude comparative dans la détection des signaux de battements cardiaques au niveau de la puissance rayonnée et de la fréquence opérationnelle. Les puissances rayonnées sont comprises entre 3 et -17 dBm et les fréquences opérationnelles utilisées sont 2.4, 5.8, 1 0 et 20 GHz. Cela permet de spécifier la fréquence opérationnelle optimale, qui donne un compromis entre la puissance minimale émise ainsi que la complexité du système de mesure. De plus, une étude comparative entre plusieurs méthodes de traitement de signal est proposée pour extraire la meilleure méthode qui permet de mesurer le signal cardiaque et par suite extraire ses paramètres. Des techniques de traitement basées sur des transformées en ondelettes ou le filtrage classique sont présentées et utilisées afin de faire une comparaison entre elles. Le paramètre extrait dans cette thèse est le taux des battements cardiaques. Les mesures ont été effectuées simultanément avec un électrocardiographe afin de valider les mesures du signal cardiaque. Puisque la personne peut se déplacer d'une pièce à une autre à l'intérieur de son domicile, des mesures des quatre côtés de la personne et derrière un mur sont réalisées. Ajoutons une approche de modélisation fondée sur la mesure cardio-respiratoire pour une personne qui exerce une marche en avant. De plus, une comparaison entre un système à micro-ondes à simple et deux antennes pour une personne qui prend son souffle est effectuée afin de tester la précision du système à antenne unique par rapport au a la deuxième. Par suite, des mesures sont effectuées pour une personne qui respire en utilisant un système à une seule antenne
Nowadays, contact-less monitoring patient's heartbeat using Doppler radar has attracted considerable interest of researchers, especially when the traditional electrocardiogram (ECG) measurements with fixed electrodes is not practical in some cases like infants at risk or sudden infant syndrome or burn victims. Due to the microwave sensitivity toward tiny movements, radar has been employed as a noninvasive monitoring system of human cardiopulmonary activity. According to Doppler effect, a constant frequency signal reflected off an object having a varying displacement will result in a reflected signal, but with a time varying phase. In our case, the object is the patient's chest; the reflected signal of the person's chest contains information about the heartbeat and respiration. The system is based on a vector network analyzer and 2 horn antennas. The S21 is computed using a vector network analyzer. The phase variation of S21 contains information about cardiopulmonary activity. Processing techniques are used to extract the heartbeat signal from the S21 phase. This thesis presents a comparative study in heartbeat detection, considering different radiated powers and frequencies. The radiated powers used are between 3 and -17 dBm and the operational frequencies used are 2.4, 5.8, 10 and 20 GHz. This helps to make a compromise between the minimum power emitted and the complexity of the measurement system. In addition, a comparative study of several signal processing methods is proposed to extract the best technique for heartbeat measurement and thus to extract its parameters. Processing techniques are based on wavelet transforms and conventional filtering in order to make a comparison between them. The parameter extracted in this thesis is the heartbeat rate HR. Measurements were performed simultaneously with a PC-based electrocardiograph to validate the heartbeat rate measurement. Since the person can move from a room to another inside his home, measurements from the four sides of the person and behind a wall are performed. In addition, a modeling approach based on cardio-respiratory measurement for a person who is walking forward is presented. Furthermore, a comparison between single and two-antenna microwave systems for a non-breathing person is carried out to test the accuracy of the single-antenna system relative to the two ­antenna microwave system. After that, measurements are performed using one antenna microwave system for a person who breathes normally