Speech communication is gaining in popularity in many different contexts as technology evolves. With the introduction of mobile electronic devices such as cell phones and laptops, and fixed electronic devices such as video and teleconferencing systems, more people are communicating which leads to an increasing demand for new services and better speech quality. Methods to enhance speech recorded by microphones often operate blindly without prior knowledge of the signals. With the addition of multiple microphones to allow for spatial filtering, many blind speech enhancement methods have to operate blindly also in the spatial domain. When attempting to improve the quality of spoken communication it is often necessary to be able to reliably determine the location of the speakers. A dedicated source localization method on top of the speech enhancement methods can assist the speech enhancement method by providing the spatial information about the sources. This thesis addresses the problem of speech-source localization, with a focus on the problem of localization in the presence of multiple concurrent speech sources. The primary work consists of methods to estimate the direction of arrival of multiple concurrent speech sources from an array of sensors and a method to correct the ambiguities when estimating the spatial locations of multiple speech sources from multiple arrays of sensors. The thesis also improves the well-known SRP-based methods with higher-order statistics, and presents an analysis of how the SRP-PHAT performs when the sensor array geometry is not fully calibrated. The thesis is concluded by two envelope-domain-based methods for tonal pattern detection and tonal disturbance detection and cancelation which can be useful to further increase the usability of the proposed localization methods. The main contribution of the thesis is a complete methodology to spatially locate multiple speech sources in enclosed environments. New methods and improvements to the combined solution are presented for the direction-of-arrival estimation, the location estimation and the location ambiguity correction, as well as a sensor array calibration sensitivity analysis.

The research in this thesis is intended to aid caregivers’ supervision of toddlers to prevent accidental injuries, especially injuries due to falls in the home environment. There have been very few attempts to develop an automatic system to tackle young children’s accidents despite the fact that they are particularly vulnerable to home accidents and a caregiver cannot give continuous supervision. Vision-based analysis methods have been developed to recognise toddlers’ fall risk factors related to changes in their behaviour or environment. First of all, suggestions to prevent fall events of young children at home were collected from well-known organisations for child safety. A large number of fall records of toddlers who had sought treatment at a hospital were analysed to identify a toddler’s fall risk factors. The factors include clutter being a tripping or slipping hazard on the floor and a toddler moving around or climbing furniture or room structures. The major technical problem in detecting the risk factors is to classify foreground objects into human and non-human, and novel approaches have been proposed for the classification. Unlike most existing studies, which focus on human appearance such as skin colour for human detection, the approaches addressed in this thesis use cues related to dynamic motions. The first cue is based on the fact that there is relative motion between human body parts while typical indoor clutter does not have such parts with diverse motions. In addition, other motion cues are employed to differentiate a human from a pet since a pet also moves its parts diversely. They are angle changes of ellipse fitted to each object and history of its actual heights to capture the various posture changes and different body size of pets. The methods work well as long as foreground regions are correctly segmented.

To positively identify marine mammals using passive acoustics, large volumes of data are often collected that need to be processed by a trained analyst. To reduce acoustic analyst workload, an automatic detector can be implemented that produces many detections, which feed into an automatic classifier to significantly reduce the number of false detections. This requires the development of a robust classifier capable of performing inter-species classification as well as discriminating cetacean vocalizations from anthropogenic noise sources. A prototype aural classifier was developed at Defence Research and Development Canada that uses perceptual signal features which model the features employed by the human auditory system. The dataset included anthropogenic passive transients and vocalizations from five cetacean species: bowhead, humpback, North Atlantic right, minke and sperm whales. Discriminant analysis was implemented to replace principal component analysis; the projection obtained using discriminant analysis improved between-species discrimination during multiclass cetacean classification, compared to principal component analysis. The aural classifier was able to successfully identify the vocalizing cetacean species. The area under the receiver operating characteristic curve (AUC) is used to quantify the two-class classifier performance and the M-measure is used when there are three or more classes; the maximum possible value of both AUC and M is 1.00 – which is indicative of an ideal classifier model. Accurate classification results were obtained for multiclass classification of all species in the dataset (M = 0.99), and the challenging bowhead/ humpback (AUC = 0.97) and sperm whale click/anthropogenic transient (AUC = 1.00) two-class classifications.

Orientador: Carlos Roberto Minussi / Banca: Anna Diva Plasencia Lotufo / Banca: Sandra Cristina Marchiori de Brito / Resumo: Este trabalho tem por objetivo apresentar uma metodologia, para detecção e classificação de faltas de curto-circuito, direcionada para os eventos de alta impedância em subestações de energia elétrica. A detecção é executada por um cálculo algébrico simples via observação da diferença absoluta entre o valor eficaz médio estimado e o valor eficaz medido a partir da oscilografia das correntes trifásicas. Caso seja detectada a anomalia de corrente, procede-se a sua classificação baseada na lógica fuzzy. Trata-se de um problema que apresenta maiores dificuldades de identificação e classificação, tendo em vista que tais perturbações são sutis, o que se diferenciam em relação aos eventos caracterizados como de baixa e ou de média impedância. Os distúrbios de alta impedância, muitas vezes, não são detectados pelo sistema proteção. Neste caso, o sistema de proteção consideraos como uma operação em regime permanente, não distinguindo entre uma falta de alta impedância e um aumento/diminuição da corrente elétrica em decorrência das variações da demanda solicitada pelos consumidores. Ressalta-se que a metodologia, aqui proposta, segue os princípios formulados na referência Decanini (2008), com as devidas adaptações ao problema associado às faltas de alta impedância. A metodologia, proposta neste trabalho, usa dados oscilográficos que são processados de modo que a detecção e classificação das faltas possam ser estimadas através de um conjunto de características extraídas dos sinais de correntes. Este conjunto de característica é classificado pela lógica nebulosa e sua saída resulta na indicação do tipo da falta. Deve-se ressaltar que este algoritmo além de ser eficiente na detecção e classificação de faltas de alta impedância, sua eficiência destacase também na localização e detecção de faltas de baixa impedância... (Resumo completo, clicar acesso eletrônico abaixo) / Abstract: This work presents a methodology to detect and classify short circuit faults principally for high impedance occurrences in electrical power systems. The detection is executed by a simple algebraic calculus observing the absolute difference with the efficient medium estimated value and the efficient value measured from the oscillogram of the three-phase currents. When a current anomaly is detected the classification is based on the fuzzy logic. It is a problem that presents some difficulties in identification and classification, considering that these perturbations are little which are different from the events characterized as low or medium impedance. The high impedance perturbations sometimes are not detected by the protection system. In this case, the protection system considers as one operation in steady state, and do not distinguish with a high impedance fault or an increasing/ decreasing of the electrical current due to the demand variation requested by the users. It is emphasized that the proposed methodology follows the fundamentals according the reference Decanini (2008), with the adaptations to the associated problem to the high impedance faults. The proposed methodology use oscillogram data that are processed in a way that the detection and classification of the faults are estimated by a set of characteristics from the current signals. This set is classified by logic fuzzy and the output results on the type of the fault. It is emphasized that this algorithm besides being efficient is detecting and classifying high impedance faults, is also efficient in localizing and detecting low impedance faults. The necessary data for the diagnosis of the faults were obtained by simulation of a radial feeder model with the ATP software. Results are satisfactory and show the viability of the proposed methodology that is faster for obtaining the solutions and is able to detect... (Complete abstract click electronic access below) / Mestre

Some fish produce sounds for a variety of reasons, such as to find mates, defend their territory, or maintain cohesion within their group. These sounds could be used to non-intrusively detect the presence of fish and potentially to estimate their number (or density) over large areas and long time periods. However, many fish sounds have not yet been associated to specific species, which limits the usefulness of this approach. While recording fish sounds in tanks is reasonably straightforward, it presents several problems: many fish do not produce sounds in captivity or their behavior and sound production is altered significantly, and the complex acoustic propagation conditions in tanks often leads to distorted measurements. The work presented in this thesis aims to address these issues by providing methodologies to record, detect, and identify species-specific fish sounds in the wild. A set of hardware and software solutions are developed to simultaneously record fish sounds, acoustically localize the fish in three-dimensions, and record video to identify the fish and observe their behavior. Three platforms have been developed and tested in the field. The first platform, referred to as the large array, is composed of six hydrophones connected to an AMAR acoustic recorder and two open-source autonomous video cameras (FishCams) that were developed during this thesis. These instruments are secured to a PVC frame of dimension 2 m x 2 m x 3 m that can be transported and assembled in the field. The hydrophone configuration for this array was defined using a simulated annealing optimization approach that minimized localization uncertainties. This array provides the largest field of view and most accurate acoustic localization, and is well suited to long-term deployments (weeks). The second platform, referred to as the mini array, uses a single FishCam and four hydrophones connected to a SoundTrap acoustic recorder on a one cubic meter PVC frame; this array can be deployed more easily in constrained locations or on rough/uneven seabeds. The third platform, referred to as the mobile array, consists of four hydrophones connected to a SoundTrap recorder and mounted on a tethered Trident underwater drone with built-in video, allowing remote control and real-time positioning in response to observed fish presence, rather than long-term deployments as for the large and mini arrays. For each array, acoustic localization is performed by measuring time-difference of arrivals between hydrophones and estimating the sound-source location using linearized (for the large array) or non-linear (for the mini and mobile arrays) inversion. Fish sounds are automatically detected and localized in three dimensions, and sounds localized within the field of view of the camera(s) are assigned to a fish species by manually reviewing the video recordings. The three platforms were deployed at four locations off the East coast of Vancouver Island, British Columbia, Canada, and allowed the identification of sounds from quillback rockfish (Sebastes maliger), copper rockfish (Sebastes caurinus), and lingcod (Ophiodon elongatus), species that had not been documented previously to produce sounds. While each platform developed during this thesis has its own set of advantages and limitations, using them in coordination helps identify fish sounds over different habitats and with various budget and logistical constraints. In an effort to make passive acoustics a more viable way to monitor fish in the wild, this thesis also investigates the use of automatic detection and classification algorithms to efficiently find fish sounds in large passive acoustic datasets. The proposed approach detects acoustic transients using a measure of spectrogram variance and classifies them as “noise” or “fish sounds” using a binary classifier. Five different classification algorithms were trained and evaluated on a dataset of more than 96,000 manually annotated examples of fish sounds and noise from five locations off Vancouver Island. The classification algorithm that performed best (random forest) has an Fscore of 0.84 (Precision = 0.82,Recall = 0.86) on the test dataset. The analysis of 2.5 months of acoustic data collected in a rockfish conservation area off Vancouver Island shows that the proposed detector can be used to efficiently explore large datasets, formulate hypotheses, and help answer practical conservation questions. / Graduate

No description available.

No description available.

Orientador: Fábio Roberto Chavarette / Resumo: Redes Neurais Artificiais (RNAs) são algoritmos de aprendizado, geralmente estruturados em torno de categorização de dados de entrada e/ou seu agrupamento por similaridade. Tendo em vista características desejáveis como aprendizado rápido e estabilidade frente a vetores de entrada altamente mutáveis, adotou-se uma RNA do tipo Fuzzy ARTMAP como mecanismo central de um método de monitoramento de saúde estrutural para detectar e categorizar falhas em dados experimentais provenientes de um sistema mecânico similar a um pequeno prédio de dois andares. Mais especificamente, com o objetivo de detectar alterações das frequências naturais da estrutura, fenômeno ligado à deterioração da mesma, e determinar qual(is) andar(es) está(ão) ligado(s) ao comportamento anômalo, se detectado. A acurácia da rede foi avaliada, sendo realizado um estudo da quantidade de dados necessárias para o desempenho satisfatório da rede. Observou-se desempenho satisfatório, a acurácia do método tendendo a aproximadamente 94% a partir de certas quantidades de dados. / Abstract: Artificial Neural Networks (ANNs) are learning algorithms, largely revolving around categorizing data sets based on measures of similarity between its members. Due to desirable characteristics such as fast learning and stability when dealing with highly mutable input vectors, a Fuzzy ARTMAP ANN was selected as the core mechanism of a structural health monitoring method. Its goal was to detect and categorize faults in experimental data collected from a mechanical system akin to a small two-story building. More specifically, to detect disturbances on the structure's natural frequencies, phenomenon linked to its deterioration, and to determine which story or stories are linked to anomalous behavior, if any. The accuracy of the method was evaluated, and the amount of data needed for optimal operation was determined. Satisfactory performance was observed; the method's accuracy tended towards 94% with enough training samples. / Mestre

Todos os dias a quantidade de veículos nas estradas em todo o mundo está aumentando. Este crescimento combinado com a negligência dos motoristas e alguns fatores externos, tais como estradas mal conservadas e condições climáticas adversas resultaram em um enorme aumento na quantidade de acidentes e, conseqüentemente, de mortes. Atualmente muitos grupos de pesquisa e empresas automotivas estão desenvolvendo e adaptando tecnologias que podem ser incorporadas nos veículos para reduzir esses números. Um exemplo interessante dessas tecnologias é a detecção e classificação de obstáculos móveis (veículos, pessoas, etc.) em ambientes urbanos. Este trabalho apresenta o desenvolvimento de algoritmos para identificação, rastreamento e previsão de obstáculos móveis, determinação de direções proibidas para tráfego do veículo e cálculo de trajetórias livres de colisões. Para isso, foram utilizados dados do sistema de medidas de distância, SICK LMS 291-S05, para monitorar o ambiente a frente do veículo de teste (um automóvel de passeio modificado). Com base nesses dados foi realizado um tratamento computacional através da técnica de Trackers para classificar todos os obstáculos detectados em duas classes principais: os obstáculos estáticos e móveis. Uma vez identificado o obstáculo, este será acompanhado mesmo no caso em que saia do campo de visão do sensor. Após a classificação dos obstáculos presentes no ambiente, suas posições são analisadas e direções proibidas para tráfego são determinadas peloalgoritmo Velocity Obstacle Approach. Finalmente é aplicada a técnica de cálculo de trajetórias E* que gera um caminho suave e livre de colisões. No caso de algum obstáculo obstruir ou gerar risco de colisão com o caminho gerado é possível recalcular a rota sem que o mapa do ambiente seja novamente completamente analisado. Os resultados obtidos demonstraram a aplicabilidade da metodologia utilizada. O algoritmo de Trackers detectou pedestres e veículos e determinou suas características dinâmicas. O algoritmo Velocity Obstacle Approach conseguiu acompanhar os obstáculos e foi capaz de determinar as direções proibidas e, finalmente, o algoritmo E* foi capaz de gerar trajetórias livre de obstáculos em ambientes desconhecidos. / Every day the number of vehicles on the roads around the world is increasing. This growth combined with the negligence of drivers and some external factors such as poorly maintained roads and adverse weather conditions resulted in a huge increase in the number of accidents and hence casualties. Currently many research groups and automotive companies are developing and adapting technologies that can be incorporated into vehicles to reduce these numbers. An interesting example of these technologies is the detection and classification of moving obstacles (vehicles, people, etc.) in urban environments. This dissertation presents the development of algorithms which main objective are identify, track and predict moving obstacles, determine prohibited directions of traffic and calculate collision free trajectories. In order to accomplish with such task, data from the laser sensor SICK LMS 291-S05 later treated using computational resources such as the Trackers technique was used to monitor the environment ahead of the test vehicle (a modified passenger car). The Trackers technique was used to classify all the hurdles identified in two main classes: static and mobile obstacles. Once the obstacle was identified, this still been followed even if they leave the field of vision sensor. After classification of obstacles in the environment, their positions are analyzed and prohibited for traffic directions are determined by the algorithm Velocity Obstacle Approach. Finally the technique is applied to calculate trajectories of E* that generates a smooth path and free of collisions. If any obstacle block, or create a risk of collision through the generated path, the trajectory can be recalculated without the need to fully re-analyze de environment map. The results demonstrated the applicability of the methodology used. The Trackers algorithm has detected pedestrians and vehicles determining their dynamic characteristics. The algorithm Velocity Obstacle Approach keep up with the obstacles and was able to determine the prohibited directions and, finally, E* the algorithm was able to generate obstacle-free paths in unknown environments.

We propose a method to detect events and event boundaries in soccer videos by using web-casting texts and audio-visual features. The events and their inaccurate time information given in web-casting texts need to be aligned with the visual content of the video. Most match reports presented by popular organizations such as uefa.com (the official site of Union of European Football Associations) provide the time information in minutes rather than seconds. We propose a robust method which is able to handle uncertainties in the time points of the events. As a result of our experiments, we claim that our method detects event boundaries satisfactorily for uncertain web-casting texts, and that the use of audio-visual features improves the performance of event boundary detection.