Dissertations / Theses on the topic 'Abnormal Event Detection'

碩士<br>國立清華大學<br>電機工程學系<br>94<br>Imagine you are asked to find out an unusual event in a daily recorded surveillance video. Questions aroused, how to detect events in a variety scenes? We focus our attention on finding out events that difference most from others and report it for further examinations. First we divide a video into several overlapping clips. Then we use optical flow to find out motion vectors of each frame in each clip. Magnitudes histogram, direction histogram and color histogram are selected as features. We form a similarity matrix by using difference and chamfer difference as the similarity measure of features in different clips. Then, we apply n-cut clustering .A threshold is selected to balance FAR (false alarm rate) and THR (true hit rate) according to ROC curve (receiver operating characteristic) and once a threshold is selected , clusters correspond to low self-similarity value is reported as unusual events and for further examination. Finally, this mechanism is tested on 6 different views.

碩士<br>元智大學<br>電機工程學系<br>97<br>This thesis proposes a system which aims at developing methodologies and techniques for abnormal event detection and navigation of a surveillance mobile robot. In the system, this approach can be divided into two parts, i.e., abnormal event detection and mobile robot navigation, for scene representation and exceptional change detection of important like paintings or antiques when mobile robot navigating known environment. In abnormal event detection, the operator controls the mobile robot to collection different videos for scene representation of training phase. Then, we use a method to build the background scene that is a patch-based technique. For detecting the abnormal event, in order to detect the abnormal object quickly, we use a patch searching algorithm that is present for scene registration. Therefore, that all possible exceptional changed that can be very efficiently detected form the scene panorama. In mobile robot navigation, the database is built as like the abnormal event detection system that let registers the path information of the mobile robot navigates known environment. Then, we use scene matching algorithm to compute the mobile robot direction and then guiding to the correct path. Experimental results are conducted to illustrate the feasibility and efficacy of the mobile robot of surveillance system.

碩士<br>國立中正大學<br>資訊工程研究所<br>103<br>Existing methods have addressed the issue of detecting abnormal events at a smart home for medical care or security monitoring services extensively in the past decades. However, most of approaches use wearable sensors that require users to be equipped with the sensor devices at every moment. If the monitored users stop or pause the sensors, any abnormal events are not able to be detected. The use of non-wearable and non-intrusive sensors (e.g., IP cameras) is necessary for providing better user experiences and achieving sustainable and reliable detection model. However, it is still very challenging to analyze such non-wearable sensor data with a high accuracy. In this work, we propose an event detection model using a Bayesian Network. We first obtain the features by analyzing the daily videos and audios captured from different angles by multiple IP cameras at a smart home. These features are then used to construct a Bayesian network. We propose a probabilistic graph model where the dependence relations are defined in the graph as opposed to the naive Bayesian network. The experiments are presented to demonstrate the performance and utility of our model.

碩士<br>國立成功大學<br>資訊工程學系<br>107<br>Abnormal event detection in surveillance videos refers to the identification of events that deviate from the normal pattern. We can use autoencoder to learn the normal pattern from normal videos and use the reconstruction error to indicate the presence of abnormalities. As the surveillance cameras are usually static, the surveillance videos can be divided into two components: dynamic objects and a static background. Because of the nature of the static background, we can assume that the source of abnormality is from the objects. In this work, we propose to use a two-stream decoder model to tackle abnormal event detection problem in surveillance videos. The two-stream decoder consists of background stream to model the static background and foreground stream to model the dynamic objects. We also utilized two-stream encoder to learn from optical flow, which contains motion information, and skip connection to improve the detail of output frames. Different constraints and adversarial training are also applied to train a more robust model for abnormality event detection task. Several experiments on publicly available datasets validate the effectiveness of our model.

碩士<br>元智大學<br>電機工程學系<br>99<br>Modeling human crowds is an important issue for video surveillance and is a challenging task due to their nature of non-rigid shapes. In this paper, for real time constraint, Haar-like features are first employed to approximately locate the position of an isolated region that comprise an individual person or a set of occluded persons. Each isolated region is considered a vertex and a human crowd is thus modeled by a graph. To regularly construct a graph, Delaunay triangulation is used to systematically connect vertices and therefore the problem of event detection of human crowds is formulated as measuring the topology variation of consecutive graphs in temporal order. To effectively model the topology variation, local characteristics such as triangle deformations and eigenvalue-based subgraph analysis, and global features such as moments are all computed and finally combined as an indicator to detect if any anomalies of human crowd(s) present in the scene. Experimental results obtained by using extensive dataset show that our system is effective in detecting anomalous events for uncontrolled environment of surveillance videos.

碩士<br>國立臺灣科技大學<br>資訊工程系<br>103<br>In order to maintain social security, everywhere the camera has become an integral part of the street scenery. On the one hand it deters crime, but it is also to enhance the people&apos;s sense of security, and to provide a method to remedy the situation have been a rainy day. However, those who have been staring at the monitor screen all day by the monitor not only time-consuming, often inefficient. So have real time anomaly detection of intelligent surveillance system is an important issue. Although intelligent surveillance system is not a new technology, but the previous systems usually need to pre-defined the abnormal behavior or require complex object features and lots of calculations to achieve high accuracy. Supervised learning monitoring system needs to define the abnormal event through human involvement, time-consuming and cannot achieve fully automatic. Unsupervised learning surveillance system which using machine learning, while avoid human intervention, but could not contain all the unusual events that missed the focus screen. If we can combine the advantages of these two systems and show the abnormal reason when anomaly happened will be the most critical success for the intelligent surveillance system. In this thesis, we presents a hybrid supervised learning real time automatic intelligent surveillance system. Capturing local features of moving objects’ trajectory. Using common clustering algorithms combined with a special structure which is the unsupervised part to train the supervised learning classifier for automatic abnormal motion detection. This kind of structure which combining the supervised learning part and unsupervised learning part is called hybrid supervised learning. When anomaly detected, the system can show the abnormal reason. If there are some error results, the supervisor can also remark the results and feedback information to the system as a basis for correction. In this way the supervisor can not only save energy, but also can improve the accuracy of abnormal detection.

碩士<br>銘傳大學<br>金融研究所<br>89<br>Event studies offer useful evidence on how stock prices respond to information. Long-run abnormal returns research focus on delayed stock price reaction and abnormal performances, which are persisting for years following the specific events. To understand how information transmits to stock prices, it must observe long-run stock performance. This paper use firms listed on Taiwan Stock Market with available data on the monthly return, different pricing model, testing statistics, and estimating benchmark, to test which methods are better on long-run abnormal returns estimation. 1. Using Cumulating Abnormal Return or Buy-and-Hold Abnormal Return to estimate long-run abnormal returns on Taiwan stock market is misspecified. 2. Estimating long-run abnormal returns by reference portfolios can improve skewness bias in random sample. Due to different industry characters, it still is misspecified in nonrandom sample, and become worst when events are clustering than events are not clustering. 3. Time series statistics in estimating event month abnormal return perform well in random sample when it uses test period standard deviation, but is misspecified in any other situation.

碩士<br>淡江大學<br>資訊工程學系碩士班<br>95<br>In this paper, we describe a video surveillance system that automatically detects abnormal events arisen from static foreground objects, which includes both abandoned and removed ones in the monitored scene. While a abnormal event arisen from static foreground objects is detected, the operator is given notice to take care of this event, and the system provides appropriate key frames for interpreting this abnormal event. 　　The method of our proposed system consists of three major phases. First, we utilize a scheme of background initialization and updating, which incorporates a background model of histogram estimation and the concept named “Sliding Buffer”, to be the foundation of the two-layer background model. This two-layer background model is used to generate current background image and reference background image. The difference image that might contain static foreground objects is then obtained by subtracting current background image and reference background image. Second, the system applies a sequence of image processing techniques, including gray-level transformation, Otsu automatic thresholding method, morphological erosion and dilation operations and connected component labeling, on the difference image acquired in previous phase to obtain the area and position information of each connected component. Finally, the system determines if a static foreground object exists in the monitored scene according to area information of each connected component. If a static foreground object is detected, the system issues an alarm of abnormal event. 　　Our system is tested under Pentium-M 1.60GHz CPU and 1GB RAM; the format of input image is 320 x 240 true color (24 bits) Bitmap. The performance of our system can reach 10 frames per second (about 0.09 ~ 0.18 seconds to process a frame according to different environment), and the average accuracy of system is higher than 85%. Comparing with other detection methods, our proposed method is relative simpler in theorem, improves the efficiency, and is capable of detecting static foreground object precisely in time while its stagnant time reaches a given threshold.

Dissertação de mestrado em Engenharia e Sistemas de Informação<br>A fraude no retalho é um dos graves problemas com que os retalhistas se deparam actualmente. Os valores dessa fraude podem ser bastantes elevados. Numa era de grande concorrência entre retalhistas, esses valores podem significar a sobrevivência da empresa. Os pontos mais sensíveis, da ocorrência de fraude, são os POS (Point Of Sale) e o armazém. Devido a esse facto surgiu a necessidade de adoptar novas abordagens para a prevenção da fraude. Com esse problema em mente surgiu esta dissertação que tem como propósito a identificação de possíveis operações fraudulentas no POS com base no tratamento das imagens das operações. Utilizando o processamento de imagem, são sugeridas duas abordagens de prevenção de fraude em POS: - A detecção de fraude através de eventos anómalos. Esta foca-se em detectar os eventos que fogem à normalidade. -A detecção de fraude através de detecção de padrões foca-se em detectar determinados padrões realizados pelos funcionários de POS. Para atingir esse fim são analisadas algumas abordagens existentes de detecção e seguimento de objectos, e de detecção anómala de eventos. Com esse estudo pretendeuse adaptar algoritmos já existentes à área de detecção de fraude em POS. Um ponto importante para a realização desta dissertação é o estado da fraude actualmente. Devido a esse facto é estudado o problema da fraude na área do retalho. Com este estudo pode-se perceber que os valores envolvidos justificam a necessidade de uma solução para o problema. Contudo este não é um tema novo, e existem soluções disponíveis no mercado, porém estas são pobres e não conseguem responder às necessidades dos retalhistas. A solução final, constituído pela integração de múltiplos componentes, foi testada num ambiente controlado, com vários objectos. Com base nos resultados dos testes, verificou-se que a solução encontrada realiza de forma eficaz a detecção de eventos anómalos. Para realização desses testes foi utilizada uma câmara de vídeo IP, a cores e fixa.<br>Fraud in Retail is one of the serious problems that retailers face today. The values of this fraud can reach quite high values. In a era of great competition between retailers, these values can mean the survival of the company. The most sensitive points, of the occurrence of fraud, are the POS (Point Of Sale) and the store. Because that fact, it’s became necessary to adopt new approaches to prevention. With this problem in mind this thesis has the purpose to identify possible fraudulent operations at the POS, based on processing of the images operations. Using image processing, are suggested two approaches for the loss prevention in POS: - Fraud detection by anomalous events focuses on detecting events that are not normal. - Fraud detection by pattern recognition focuses on detecting patterns made by employees of POS. To achieve this goal are analyzed some existing approaches for detecting and tracking moving objects, and detection of anomalous events. With this study was intended to adapt existing algorithms to the area of fraud detection POS. An important point to realize this work is the actual state of fraud. Because of this fact it is studied the problem in the area of retail. With this study can be seen that the values involved justify the need for a solution to the problem. Yet this is not a new theme, and there are solutions available, but these are poor and cannot meet the needs of retailers. The final solution, formed by the integration of multiple components, was tested in a controlled environment with multiple objects. Based on tests results, the solution detect anomalous events. For these tests we used a video camera IP, and fixed color.

Dissertação de mestrado integrado em Engenharia Biomédica (área de especialização em Eletrónica Médica)<br>In order to actively aid or restore legged locomotion to individuals suffering from muscular impairments, weakness or neurologic injury, rehabilitation is recommended as a more appropriate way to achieve the ultimate goal of a continuous ambulatory monitoring. Also, the assistance with wearable robots (WRs) during daily living activities provides a more intensive and purposeful targeted therapeutic training, and also reduces the treatment cost and the number of health care personnel. Thus, it is crucial the development of locomotion strategies that recognize in real-time the locomotion mode of human-robot interaction in overground daily living activities. Thus, this thesis intends to develop two locomotion strategies which will be integrated in high level control of exoskeleton H2 (Exo-H2), the WR developed under the scope of BioMot project. The first locomotion strategy proposed and validated addresses online detection of events and gait phases uniquely through information from embedded sensors. This knowledge will allow determining in real-time the biomechanical parameters of assisted walking, and consequently to assess the progress of rehabilitation process by means of WR. The solution validation in different locomotion conditions (assisted walking by WR, walking of humanoid robot and walking of healthy subject) shows up that the proposed solution led to a robust and general tool for gait detection, which is also capable to detect more events and gait phases comparatively to the works presented in literature. Locomotion mode recognition is the second locomotion strategy developed in this thesis, which allows the recognition of different locomotion modes. Based on an exhaustive state of the art survey, a more robust and accurate procedure that leads to a more robust and accurate tool was delineated. According to the results achieved for offline scenario it was verified that the performance of the locomotion strategy increases by using different types of biomechanical parameters, which should be previously selected by means of multivariate statistic methods. Both binary and multiclass classification were addressed through support vector machine (SVM). The implementation of these methods led to a powerful and accurate tool of offline recognition of locomotion modes. Additionally, a strategy for online recognition was proposed. Further work will consist on the application of these locomotion strategies in real-time environment of gait rehabilitation.<br>De forma a apoiar ou a restaurar a locomoção de indivíduos que apresentam fraqueza muscular ou doenças neurológicas, a reabilitação é recomendada como a forma mais apropriada para alcançar uma monitorização ambulatória contínua. Além disso, a assistência com robots ambulatório (RA) durante as atividades diárias promove um treino terapêutico mais intensivo e direcional, assim como também reduz os custos de tratamento e o número de profissionais de saúde. Como tal, é crucial o desenvolvimento de estratégias que reconheçam, em tempo real, o modo de locomoção da interação sujeito-robot durante as atividades quotidianas. Assim, esta tese visa desenvolver duas estratégias de locomoção, as quais serão integradas no controlo de alto nível do Exo-H2, que corresponde ao RA desenvolvido no âmbito do projeto BioMot. A primeira estratégia de locomoção proposta e validada consiste na deteção online dos eventos e fases da marcha, usando exclusivamente a informação fornecida pelos sensores embebidos. Este conhecimento permitirá a determinação em tempo real dos parâmetros biomecânicos da marcha assistida, e por conseguinte permitirá avaliar o progresso da reabilitação. A validação da solução proposta em diferentes contextos de locomoção (marcha assistida por RA, marcha de um robot humanoide e a marcha de um sujeito saudável) revelou que esta constitui uma ferramenta robusta e geral para a deteção da marcha, sendo capaz de detetar mais eventos e fases da marcha comparativamente aos estudos apresentados na literatura. O reconhecimento do modo de locomoção é a segunda estratégia desenvolvida nesta tese, a qual permite o reconhecimento de diferentes modos de locomoção. Com base no exaustivo levantamento do estado da arte, foi delineado um procedimento robusto, que conduziu a uma ferramenta mais robusta e precisa. De acordo com os resultados alcançados para o cenário offline verificou- se que o desempenho desta estratégia de locomoção aumenta com a utilização de diferentes tipos de parâmetros biomecânicos, os quais devem ser previamente selecionados por meio de métodos estatísticos. Tanto a classificação binária, como a classificação de multi-classes, foram implementadas através do support vector machine (SVM). A implementação destes métodos conduziu a uma ferramenta precisa de reconhecimento dos modos de locomoção em offline. Além disso, também foi proposta a estratégia para o reconhecimento em tempo real. Como trabalho futuro propõe-se a aplicação destas estratégias de locomoção no ambiente em tempo real de reabilitação da marcha.