Dissertations / Theses on the topic 'Keyframes'

The aim of this doctoral thesis is to present a body of work aimed at reducing the time spent by animators manually constructing keyframed animation. To this end we present a number of state of the art machine learning techniques applied to the domain of character animation. Data-driven tools for the synthesis and production of character animation have a good track record of success. In particular, they have been adopted thoroughly in the games industry as they allow designers as well as animators to simply specify the high-level descriptions of the animations to be created, and the rest is produced automatically. Even so, these techniques have not been thoroughly adopted in the film industry in the production of keyframe based animation [Planet, 2012]. Due to this, the cost of producing high quality keyframed animation remains very high, and the time of professional animators is increasingly precious. We present our work in four main chapters. We first tackle the key problem in the adoption of data-driven tools for key framed animation - a problem called the inversion of the rig function. Secondly, we show the construction of a new tool for data-driven character animation called the motion manifold - a representation of motion constructed using deep learning that has a number of properties useful for animation research. Thirdly, we show how the motion manifold can be extended as a general tool for performing data-driven animation synthesis and editing. Finally, we show how these techniques developed for keyframed animation can also be adapted to advance the state of the art in the games industry.

This master thesis is about modifying a certain monocular visual SLAM algorithm to address some of its limitations. The SLAM algorithm is not robust to quick camera motions and input images in which there are few visible features. A second camera and an inertial measurement unit was added to the hardware. Then a method for selecting the appropriate camera for tracking depending on the estimated number of features was implemented to solve the issue of few features. Experiments and results show that this method works well for slow motions. A gyrometer threshold method along with a motion model to solve the issue of quick motions was implemented and reviewed in this thesis.<br>Detta examensarbete handlar om att ta itu med några begränsningar som en viss monokulär visuell SLAM-algoritm har. SLAM-algoritmen är inte robust mot snabba kamerarörelser och indatabilder som innehåller få karaktärsdrag. Genom att introducera en extra kamera, en accelerometer och en gyrometer, behandlas dessa problem i denna rapport. En metod för att välja kamera att hämta indatabilder från, baserat på det skattade antalet karaktärsdrag i respektive bild implementerades. Denna metod är tänkt att lösa problemet då indatabilder har få karaktärsdrag. Experiment visar att denna metod fungerar bra för långsamma rörelser. En metod som jämför gyrometerdata med ett tröskelvärde tillsammans med en rörelsemodell implementerades för att lösa problemen vid snabb rörelse. Dessa metoder undersöks och diskuteras i rapporten.

Made available in DSpace on 2014-06-11T19:24:06Z (GMT). No. of bitstreams: 0 Previous issue date: 2013-04-25Bitstream added on 2014-06-13T20:09:04Z : No. of bitstreams: 1 portarijunior_sc_me_bauru.pdf: 1375429 bytes, checksum: 55957c89fe49d551d28ecadf740b82ff (MD5)<br>Utilizar um cenário virtual em TV Digital tornou-se comum com o avanço das tecnologias de hardware e software. Mas para se obter um cenário virtual que convença o telespectador pode-se utilizar reconstruções tridimensionais foto-realísticas como uma possível alternativa. Este trabalho apresenta uma proposta de um método para o pré-processamento de um vídeo, capturado, no mundo real, onde são extraídos frames adequados à reconstrução 3D (keyframes) pelo método SFM (Structure From Motion). Desta forma o processamento para a reconstrução 3D utiliza apenas os frames considerados essenciais, diminuindo as redundâncias e falhas. Com a utilização deste método conseguiu-se reduzir o tempo de processamento durante a reconstrução 3D. Neste trabalho também comparou-se o método proposto com os métodos tradicionais, onde não existe uma prévia seleção de keyframes, utilizando-se diferentes ferramentas de reconstrução baseadas no método SFM<br>Using a virtual scenario on Digital TV became common with the advance of hardware and software technologies. However, to obtain a virtual scenario that persuades the viewer, photorealistic three-dimensional reconstructions can be used as a possible alternative. This work proposes a method for the pre-processing of a video, captured in the real world, where frames which are appropriate for 3D reconstruction (keyframes) by the method SFM (Structure From Motion) are extracted. Thus the processing for 3D reconstruction use only the frames considered essential, reducing redundancies and gaps. Using this method it was possible to reduce the processing time for the 3D reconstruction. In this work, the proposed method was compared to traditional methods, where there is no prior selection of keyframes, using different tools of reconstruction based on the SFM method

Orientador: Antonio Carlos Sementille<br>Banca: Wilson Massashiro Yonezawa<br>Banca: Idelberto Aparecido Rodello<br>Resumo: Utilizar um cenário virtual em TV Digital tornou-se comum com o avanço das tecnologias de hardware e software. Mas para se obter um cenário virtual que convença o telespectador pode-se utilizar reconstruções tridimensionais foto-realísticas como uma possível alternativa. Este trabalho apresenta uma proposta de um método para o pré-processamento de um vídeo, capturado, no mundo real, onde são extraídos frames adequados à reconstrução 3D (keyframes) pelo método SFM (Structure From Motion). Desta forma o processamento para a reconstrução 3D utiliza apenas os frames considerados essenciais, diminuindo as redundâncias e falhas. Com a utilização deste método conseguiu-se reduzir o tempo de processamento durante a reconstrução 3D. Neste trabalho também comparou-se o método proposto com os métodos tradicionais, onde não existe uma prévia seleção de keyframes, utilizando-se diferentes ferramentas de reconstrução baseadas no método SFM<br>Abstract: Using a virtual scenario on Digital TV became common with the advance of hardware and software technologies. However, to obtain a virtual scenario that persuades the viewer, photorealistic three-dimensional reconstructions can be used as a possible alternative. This work proposes a method for the pre-processing of a video, captured in the real world, where frames which are appropriate for 3D reconstruction (keyframes) by the method SFM (Structure From Motion) are extracted. Thus the processing for 3D reconstruction use only the frames considered essential, reducing redundancies and gaps. Using this method it was possible to reduce the processing time for the 3D reconstruction. In this work, the proposed method was compared to traditional methods, where there is no prior selection of keyframes, using different tools of reconstruction based on the SFM method<br>Mestre

The goal of this work is to create an application, which will combine procedural and keyfram animations with subsequent visualization. Composition of this two different animations techniques is used to animate a virtual character. To combine this two techniques one starts with interpolations from keyframe animation and then enchance them by procedural animations to properly fit into the characters surroundings. This procedural part of animation is obtained by using forward and inverse kinematics. Whole application is written in C++, uses GLM math library for computations and OpenGL and GLUT for final visualization.

Approved for public release; distribution is unlimited<br>The explosive growth of user-generated video presents opportunities and challenges. The videos may possess valuable information that was once unavailable. On the other hand, the information may be buried or difficult to access with traditional methods. Automatic keyframe video summarization technologies exist that attempt to address this problem. A keyframe summary can often be viewed quicker than the underlying video. However, a theoretical framework for objectively assessing keyframe summary quality has been absent. The work presented here bridges this gap by presenting a semantically high-level, stakeholder-centered evaluation frame-work. The framework can capture common stakeholder concerns and quantitatively measure the extent to which they are satisfied by keyframe summaries.With this framework, keyframe summary stakeholders and algorithm developers can now identify when success has been achieved. This work also develops a number of novel keyframe summarization algorithms and shows, using the evaluation framework, that they outperform baseline methods.

Context: When considering the use of Augmented Reality to provide navigation cues in a completely unknown environment, the content must be delivered into the environment with a repeatable level of accuracy such that the navigation cues can be understood and interpreted correctly by the user. Aims: This thesis aims to investigate whether a still image based reconstruction of an Augmented Reality environment can be used to develop a content delivery system that providers a repeatable level of accuracy for content placement. It will also investigate whether manipulation of the properties of a Spatial Marker object is sufficient to reduce object selection ambiguity in an Augmented Reality environment. Methods: A series of experiments were conducted to test the separate aspects of these aims. Participants were required to use the developed Keyframe Tagging tool to introduce virtual navigation markers into an Augmented Reality environment, and also to identify objects within an Augmented Reality environment that was signposted using different Virtual Spatial Markers. This tested the accuracy and repeatability of content placement of the approach, while also testing participants’ ability to reliably interpret virtual signposts within an Augmented Reality environment. Finally the Keyframe Tagging tool was tested by an expert user against a pre-existing solution to evaluate the time savings offered by this approach against the overall accuracy of content placement. Results: The average accuracy score for content placement across 20 participants was 64%, categorised as “Good” when compared with an expert benchmark result, while no tags were considered “incorrect” and only 8 from 200 tags were considered to have “Poor” accuracy, supporting the Keyframe Tagging approach. In terms of object identification from virtual cues, some of the predicted cognitive links between virtual marker property and target object did not surface, though participants reliably identified the correct objects across several trials. Conclusions: This thesis has demonstrated that accurate content delivery can be achieved through the use of a still image based reconstruction of an Augmented Reality environment. By using the Keyframe Tagging approach, content can be placed quickly and with a sufficient level of accuracy to demonstrate its utility in the scenarios outlined within this thesis. There are some observable limitations to the approach, which are discussed with the proposals for further work in this area.

Simulating realistic human behavior on a virtual avatar presents a difficult task. Because the simulated environment does not adhere to the same scientific principles that we do in the existent world, the avatar becomes capable of achieving infeasible postures. In an attempt to obtain realistic human simulation, real world constraints are imposed onto the non-sentient being. One such constraint, and the topic of this thesis, is self-collision avoidance. For the purposes of this topic, a posture will be defined solely as a collection of angles formed by each joint on the avatar. The goal of self-collision avoidance is to eliminate the formation of any posture where multiple body parts are attempting to occupy the exact same space. My work necessitates an extension of this definition to also include collision avoidance with objects attached to the body, such as a backpack or armor. In order to prevent these collisions from occurring, I have implemented an effort-based approach for correcting afflicted postures. This technique specifically pertains to postures that are sequenced together with the objective of animating the avatar. As such, the animation's coherence and defining characteristics must be preserved. My approach to this problem is unique in that it strategically blends the concept of keyframe interpolation with an optimization-based strategy for posture prediction. Although there has been considerable work done with methods for keyframe interpolation, there has been minimal progress towards integrating a realistic collision response strategy. Additionally, I will test this optimization-based approach through the use of a complex kinematic human model and investigate the use of the results as input to an existing dynamic motion prediction system.

SLAM (do inglês Simultaneous Localization and Mapping) Monocular baseado em Keyframes é uma das principais abordagens de SLAM Visuais, usado para estimar o movimento da câmera juntamente com a reconstrução do mapa sobre frames selecionados. Estas técnicas representam o ambiente por pontos no mapa localizados em um espaço tri-dimensional, que podem ser reconhecidos e localizados no frame. Contudo, estas técnicas não podem decidir quando um ponto do mapa se torna um outlier ou uma informação obsoleta e que pode ser descartada, ou combinar pontos do mapa que correspondem ao mesmo ponto tri-dimensional. Neste trabalho, apresentamos um método robusto para manter um mapa refinado. Esta abordagem usa o grafo de covisibilidade e um algoritmo baseado na fusão de informações para construir um mapa probabilístico, que explicitamente modela medidas de outlier. Além disso, incorporamos um mecanismo de poda para reduzir informações redundantes e remover outliers. Desta forma, nossa abordagem gerencia a redução do tamanho do mapa, mantendo informações essenciais do ambiente. Finalmente, a fim de avaliar a performance do nosso método, ele foi incorporado ao sistema do ORB-SLAM e foi medido a acurácia alcançada em datasets publicamente disponíveis que contêm sequências de imagens de ambientes internos gravados com uma câmera monocular de mão.<br>Keyframe-based monocular SLAM (Simultaneous Localization and Mapping) is one of the main visual SLAM approaches, used to estimate the camera motion together with the map reconstruction over selected frames. These techniques based on keyframes represent the environment by map points located in the three-dimensional space that can be recognized and located in the frames. However, many of these techniques cannot combine map points corresponding to the same three-dimensional point or detect when a map point becomes outlier and an obsolete information. In this work, we present a robust method to maintain a refined map that uses the covisibility graph and an algorithm based on information fusion to build a probabilistic map, which explicitly models outlier measurements. In addition, we incorporate a pruning mechanism to reduce redundant information and remove outliers. In this way our approach manages the map size maintaining essential information of the environment. Finally, in order to evaluate the performance of our method, we incorporate it into an ORB-SLAM system and measure the accuracy achieved on publicly available benchmark datasets which contain indoor images sequences recorded with a hand-held monocular camera.

Modern technology and computer science is spreading and developing very fast. The use and facilities evolve too. One of the fields where computer science is widely used – computer games. Computer animation makes games more interactive and attractive to the user. The purpose of this project was to review and analyze the animation techniques used in modern games development. Review file formats witch are used in computer games to store game objects and animation data. Try to point out the strengths and weaknesses of animation techniques. Main attention was committed to skeletal and key frame animation. Implement skeletal and key frame animation techniques.

The main goal of this project was to familiarize readers with the techniques used in real-time animation of 3D characters. This work is focused on two types of animation: keyframe animation and skeletal animation. There are described algorithms for software and hardware accelerated model deformations, keyframe interpolations, animation blending, inverse kinematics and ragdoll. The result of this project is a framework, which consists of an animation library, examples demonstrating library functions and tools for export animations from 3D Studio Max and MilkShape 3D.

[EN] This Thesis presents a set of tools that allows the improvement of the digital video coding efficiency by exploiting the fundamentals of the state of the art video coding standards. This work has been focused both on research and on the application of the results to the encoding of digital video in real time mobile environments. The first contribution is an automatic shot change detection algorithm integrated in the encoding process. This algorithm is based on the monitoring of the coding mode of the macroblocks of the sequence, and the proper definition of a set of parameters provides excellent detection rates, precision and recall. The results also indicate an improvement on the encoded video quality when these detection techniques are used, which triggers the definition of a content-based keyframe selection algorithm. With this method, the optimal position of reference pictures can be determined. These keyframes are then used by the encoder to perform temporal prediction of the subsequent frames, which improves the compression rate and the encoded video quality (both objective and subjective). This quality improvement is the main objective of this Thesis. In the last part of this work, a rate control algorithm for variable bitrate and frame rate environments has been defined, being able to generate a bitstream that quickly follows the varying conditions of the mobile channel. In parallel to all this work, a set of training and test sequences has been obtained, providing an optimal environment for the design, development, configuration, optimization and test of the algorithms described here.<br>[ES] Esta tesis presenta un conjunto de herramientas que permiten mejorar la eficiencia de codificación de vídeo mediante la explotación de los fundamentos en los que se basan los principales estándares de codificación actuales. El trabajo se ha orientado tanto a la investigación como a la aplicación de los resultados a la codificación de vídeo en tiempo real en entornos móviles. En primer lugar se ha definido un algoritmo de detección automática de cambios de plano para entornos de tiempo real integrado en el proceso de codificación. Este algoritmo está basado en la monitorización del modo de codificación de los macrobloques de la secuencia y la correcta definición de un conjunto de parámetros consigue unas tasas de detección, una precisión y una eficacia superiores a otros métodos similares existentes en la literatura. Los resultados muestran también una mejora en la calidad del vídeo codificado al aplicar estas técnicas de detección, lo que lleva a la definición de un algoritmo de selección de imágenes de referencia (keyframes) basado en el contenido. Así se pueden obtener las posiciones óptimas para las imágenes de referencia utilizadas por el codificador para realizar predicciones temporales que aumentan la calidad tanto objetiva como subjetiva del vídeo codificado, lo que constituye a su vez el objetivo principal de esta tesis. Por último, se ha diseñado un algoritmo de control de tasa capaz de obtener un bitstream que se adapta rápidamente a los cambios tanto de bitrate como de tasa de imágenes por segundo producidos en el canal móvil. Paralelamente, se ha obtenido un conjunto de secuencias de entrenamiento y test que proporcionan un entorno óptimo para el diseño, desarrollo, configuración, optimización y prueba de los algoritmos aquí descritos.<br>[CAT] Aquesta tesi presenta un conjunt de ferramentes que permeten millorar la eficiència de codificació de vídeo digital mitjançant l'explotació dels fonaments en els que es basen els principals estàndards de codificació actuals. El treball ha estat orientat tant a la investigació com a l'aplicació dels resultats a la codificació de vídeo en temps real en entorns mòbils. En primer lloc s'ha definit un algoritme de detecció automàtica de canvis de plànol integrat en el propi procés de codificació. Aquest algoritme s'ha basat en la monitorització del mode de codificació dels macroblocs de la seqüència, i la correcta definició d'un conjunt de paràmetres de configuració permet aconseguir unes taxes de detecció, una precisió i una eficàcia superiors a altres mètodes similars presents a la literatura. Aquests resultats també indiquen una millora en la qualitat del vídeo codificat al aplicar aquestes tècniques de detecció la qual ens porta a la definició d'un algoritme de selecció d'imatges de referència (keyframes) basada en el contingut. Amb aquest algoritme es poden obtenir les posicions òptimes per a les imatges de referència utilitzades pel codificador per a realitzar prediccions temporals òptimes que augmenten la qualitat tant objectiva com subjectiva del vídeo codificat. Amb esta millora s'assoleix l'objectiu principal d'aquesta tesi. Per últim, s'ha dissenyat un algoritme de control de taxa capaç d'obtenir un bitstream que s'adapta ràpidament als canvis tant de bitrate com de taxa d'imatges per segon requerits per les condicions canviants del canal mòbil. Paral·lelament s'ha obtingut un conjunt de seqüències d'entrenament i test que permet disposar d'un entorn òptim per al disseny, desenvolupament, configuració, optimització i prova dels algoritmes descrits en aquestes fulles.<br>Usach Molina, P. (2015). TÉCNICAS DE MEJORA DE LA EFICIENCIA DE CODIFICACIÓN DE VÍDEO [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/59446<br>TESIS

Dans ce travail, nous proposons une représentation efficace de l’environnement adaptée à la problématique de la navigation autonome. Cette représentation topométrique est constituée d’un graphe de sphères de vision augmentées d’informations de profondeur. Localement la sphère de vision augmentée constitue une représentation égocentrée complète de l’environnement proche. Le graphe de sphères permet de couvrir un environnement de grande taille et d’en assurer la représentation. Les "poses" à 6 degrés de liberté calculées entre sphères sont facilement exploitables par des tâches de navigation en temps réel. Dans cette thèse, les problématiques suivantes ont été considérées : Comment intégrer des informations géométriques et photométriques dans une approche d’odométrie visuelle robuste ; comment déterminer le nombre et le placement des sphères augmentées pour représenter un environnement de façon complète ; comment modéliser les incertitudes pour fusionner les observations dans le but d’augmenter la précision de la représentation ; comment utiliser des cartes de saillances pour augmenter la précision et la stabilité du processus d’odométrie visuelle<br>Our aim is concentrated around building ego-centric topometric maps represented as a graph of keyframe nodes which can be efficiently used by autonomous agents. The keyframe nodes which combines a spherical image and a depth map (augmented visual sphere) synthesises information collected in a local area of space by an embedded acquisition system. The representation of the global environment consists of a collection of augmented visual spheres that provide the necessary coverage of an operational area. A "pose" graph that links these spheres together in six degrees of freedom, also defines the domain potentially exploitable for navigation tasks in real time. As part of this research, an approach to map-based representation has been proposed by considering the following issues : how to robustly apply visual odometry by making the most of both photometric and ; geometric information available from our augmented spherical database ; how to determine the quantity and optimal placement of these augmented spheres to cover an environment completely ; how tomodel sensor uncertainties and update the dense infomation of the augmented spheres ; how to compactly represent the information contained in the augmented sphere to ensure robustness, accuracy and stability along an explored trajectory by making use of saliency maps

This thesis proposes an architecture of animation engine flexible enough to support large scale of animation algorithms with unified approach to each one of them. One of the major goals of the engine is to support creation of very complex animation seqences with high degree of animation execution control. The main animation technique used in the engine is skeletal animation and some variants of this technique are already implemented within, but it was never meant to be skeletal animation only engine and is designed that way.

Advances in computer hardware and software have made it possible to use physical simulation methods to add unprecedented degrees of realism to computer animations. Unfortunately, the reliance on such methods makes it difficult or even impossible to control the results in order to achieve artistic goals. We present a method for allowing artistic control of physical realism by interpolating between given keyframes. The method is based on the discrete Lagrange-d'Alembert principle; we introduce additional "ghost forces" that are calculated to bring the system into the configuration requested by the artist. We derive a cost function that when minimized ensures the corresponding motion to be smooth and physical looking. We describe the implementation and show an example of animating a multi-particle mass-spring system in 3D using our method.

碩士<br>國立成功大學<br>電腦與通信工程研究所<br>98<br>With the recent 3D display technology improvements, people now can perceive 3D effects in 3D display systems without wearing 3D glasses. However, due to the lack of 3D contents, 2D-to-3D conversion technique becomes more and more important in the near future. For 2D-to-3D video conversion, the most common way is to generate correspondent depth video from the original 2D video first then synthesizing the desired video by utilizing generated depth videos. The main key point for 2D-to-3D conversion is how to generate the corresponding depth video precisely and efficiently. This thesis proposes an efficient depth maps interpolation method from existing pairs of key frames and depth maps. The proposed method contains forward/backward block-based motion estimation/compensation, block refinement, object refinement, and frame selection mechanism. Experimental results show that the proposed method can successfully and effectively generate depth maps, which will be a great help for 3DTV content generation.

博士<br>國立臺灣大學<br>資訊網路與多媒體研究所<br>99<br>The surveillance industry grows vigorously in recent years, and the monitoring equipments have been mostly digitalized. However, the recorded video files are usually very long, including contents for activities all day long or even several months. For retrieving and browsing the desired video file from the database efficiently, the user needs some visual indices. In surveillance applications, the keyframe approach is more popular in browsing and retrieval on video. However, traditional keyframe extracting methods select video frames as keyframes directly from the input video, and the information in these selected video frames is scattered and does not easily let the users perceive how the events happened in the original video. In addition, they will need to generate a lot of keyframes to express events recorded in the video, which is not very efficient. In this thesis, we propose a new type of keyframe (called an “augmented keyframe”) that is a more meaningful and compact keyframe, augmented with representative objects, important contents (human faces, license plates, etc.), motion information (including trajectories and simple movement situations), and some marks of the moving objects extracted from a surveillance video clip captured by a static camera. This new technique consists of two major phases: content extraction and content synthesis. The innovation is that we propose two different content synthesis approaches (based on the 2D and the 3D information, respectively) to generate the augmented keyframe. In addition, we show through the comparison and the user study in our experiments that the augmented keyframe can generate more comprehensible and meaningful keyframes and collect better user feedback than traditional keyframe approaches.

We present an approach that automatically constructs physically plausible in-between frames, given keyframes of arbitrary implicit surface geometry and feature points registered between adjacent keyframes. This extends to usable keyframe control of computer animated fluid-like materials. Most current implicit surface morphs do not allow feature point tracking and none guarantee physically plausible in-between frames of arbitrary motion. Standard triangle surface mesh morphing techniques do not guarantee physically plausible in-betweens either, nor can they handle topological changes. Current fluid control approaches do not respect keyframes nor track feature points. Our variational approach finds a volume mapping between keyframes which minimizes a physics-based objective function using Gauss-Newton modified to handle linear constraints. We then create as-rigid-as-possible trajectories of the volume respecting this map, which we use to create physically plausible in-between frames.<br>Science, Faculty of<br>Computer Science, Department of<br>Graduate

碩士<br>義守大學<br>資訊工程學系碩士班<br>99<br>As the motion capture technique has been developed in recent years, the movie industry and game industry has widely used this technique to obtain realistic three-dimensional animation. Toward high-level semantic analysis and understanding, it is an important step to provide a compact feature of motion capture data for information retrieval and motion navigation. The subject of this thesis is to explore how to obtain a small number of keyframes from any human motion capture data. Although some strategies were introduced in the past literature such as curve simplification, curve saliency and frame decimation, most of them adopted the data distortion to measure its effectiveness, but did not take into account the human visual perception. In this study, the angle based curve simplification methods, and the other two strategies were proposed to the absolute coordinates of the position based curve simplification method based on the amount of projection based curve detection methods in order to get the keyframes can represent its content, and In order to verify consistent with human visual experience, the three methods to calculate the precision and the recall and compare their advantages and disadvantages.

Dissertação de Mestrado Integrado em Engenharia Electrotécnica e de Computadores apresentada à Faculdade de Ciências e Tecnologia da Universidade de Coimbra.<br>As atividades humanas são caracterizadas como sendo complexas, não só pelos seus padrões espaço-temporais, mas também devido ao facto de que cada ser humano é fisicamente diferente. A mesma atividade é efetuada diferentemente por várias pessoa e até o mesmo ser humano pode executar a mesma atividade com movimentos distintos. Neste sentido, existe a necessidade de incorporar métodos de aprendizagem máquina capazes de lidar com esta variabilidade. No entanto, mesmo com o aumento de atenção que esta área de investigação está a ter, a maior parte destas técnicas de aprendizagem, embora precisas/exatas, são bastante lentas para treinar e classificar. Esta característica torna tais ferramentas inviáveis para aplicações em tempo-real, onde um robô tem de identificar correta e rapidamente uma dada atividade, por forma a responder adequadamente.Neste contexto, é proposta uma metodologia de reconhecimento de atividades humanas que é treinada rapidamente e requer poucos exemplos de treino, consistindo em duas componentes base: uma componente de extração de características (isto é, extração de informação relevante que descreve uma certa actividade) e uma componente de aprendizagem, baseada no classificador de florestas aleatórias (random forests). Inicialmente, é explorado o conceito baseado na segmentação de cada atividade numa sequência de janelas de tamanho fixo (isto é, a classificação não é feita quadro a quadro), extraindo apenas valores máximos e mínimos de cada característica obtida baseada no esqueleto humano. Após a sua validação experimental, foi apresentada e testada uma segunda abordagem, considerando a divisão de cada atividade em janelas de tamanho variável, baseadas em poses chave. Cada janela de ação é delimitada por duas poses chave consecutivas que são identificadas automaticamente, sendo extraídas características estáticas (isto é, geométricas) e dinâmicas (isto é, temporais). Primeiro, estas abordagens foram testadas usando o Cornell Activity Dataset e o classificador de florestas aleatórias disponibilizado pelo programa Weka, obtendo-se resultados médios globais relevantes. Em seguida, foi desenvolvido de raíz um classificador de florestas aleatórias, usando o algoritmo de evolução diferencial como parte do seu núcleo. O classificador de florestas aleatórias desenvolvido foi testado e os seus resultados comparados com os previamente obtidos com o programa Weka, observando-se uma ligeira melhoria em termos dos indicadores de desempenho considerados. Foi construído um dataset, usando um sensor RGB-D, com base no qual foram testados ambos os classificadores de florestas aleatórias. Toda a estrutura foi implementada numa plataforma robótica real (em C++), após a sua validação em ambiente Matlab. Após algumas observações finais, destacam-se novas direcções de pesquisa, por forma a melhorar as características da metodologia propostade reconhecimento proposta e colmatar as suas limitações.<br>Human activities are characterized as being complex, not only for their temporal and spacial patterns, but also because of the fact humans are physically different. When the same activity is performed by different persons, or even by the same one, they can execute the same activity with distinct motion properties. In this sense, there is the need to incorporate machine learning approaches able to integrate all this variability. However, even with the increasingly attention this field of research is getting, current learning approaches, although accurate, are very slow to train and classify. These limitations turn infeasible their use in real-time applications, where a robot must rapidly and correctly identify a performed activity, in order to respond accordingly. A human activity recognition framework, featuring fast training and requiring few training examples, is therefore proposed, consisting of two main components: a features extraction approach component (i.e. extraction of relevant information describing a certain activity) and a machine learning component, based on the random forest classifier. An initial concept of segmenting each activity into a sequence of fixed-size actions is employed (i.e. no frame-by-frame classification), extracting just maximum and minimum values of each extracted skeleton-based feature. After some experimental validation, a second approach was developed and tested, considering the division of each activity into variable-size windows, based on key poses. Each action window is delimited by two consecutive and automatically identified key poses, where static (i.e. geometrical) and maximum and minimum dynamic (i.e. temporal) features are extracted. First, these approaches were tested using the Cornell Activity Dataset and the random forest classifier provided by Weka Software,obtaining relevant overall average results. Then a custom random forest classifier was developed from scratch, using a differential evolution algorithm, as part of its core. The developed random forest was tested and its results compared to the ones obtained with Weka's random forest,suggesting a slight increase in terms of the considered performance indicators. A custom dataset was built, using data from a RGB-D sensor, and both random forest classifiers were tested. The proposed framework, after being validated using Matlab, was implemented in a real robotic platform (in C++). After some concluding remarks, new open research directions are highlighted, in order to improve the framework's characteristics and to bridge its drawbacks.

碩士<br>國立臺灣師範大學<br>資訊工程研究所<br>99<br>Generally speaking, a programmer usually needs to use or maintain the code which is written by someone else. Before he/she is capable of modifying the code, he/she must understand the program. Several entities can help understanding a program. For example, a design document, if available, can greatly help the understanding of a program. However, a design document might be out of date or inconsistent to the code. Tracing the source code is often used as a last resort. It is common to trace a program by debugger. The programmers set break points inside the program and then execute the program in debug mode to observe the change of variable values in order to understand how does the program work. Most debuggers nowadays are textual and numerical based. It is inadequate when a program comprises of complicated data structures. In these years, several software visualization researches have been proposed to visualize software by the graphics, which are called visualize metaphors (VM). However, these research tools are typically limited to specific application domains, without generality. Tools that provide animation capability include ANIMAL[1], PlanAni[2], JEDAS[3], Jeliot3[4], POLKA[5], TANGO[6], and XTANGO[7]. These prototype software tools typically limit their visualization power to fixed problem domains such as education. Some requires writing extra code for the animation. In this thesis, we proposed an animation subsystem in xDIVA[8]. It can display the program dynamic behavior between the break points by animation so that a program can be understood in a more efficient manner. In our method, users do not need to write the extra code for animation. After a layout automatically arranges the VMs in a new break point, our approach computes the differences in VM’s positions and properties so that an interpolation between the two key frames at two break points can be constructed. An animation is then made from such an interpolation.

In the field of computer animation, character animation using keyframes remains a popular technique. In this, the animation sequence is represented compactly by just the more important character poses, termed as keyframes and the rest of the frames, known as in-betweens are generated when needed, say during playing the animation, using the keyframes. The animator specifies (creates) the keyframes while the in-between frames are computed using a suitable interpolation scheme. Interpolation parameters are usually under the control of the animator. Most animation software today will include support for keyframe animation. However, specifying the parameters so that the generated animation sequence fulfils the animator expectations and other motion requirements, say like preserving area or volume, or satisfying a physical constraint, can be quite difficult. The number of degrees of freedom is very high for skeleton-based animation and much higher for mesh-based animation. Physically-based animation techniques have been proposed for character poses to satisfy physics constraints. But animators find it difficult and non-intuitive to specify physics parameters, like body mass, forces. etc. and seem to very much dislike the fact that they lose control over the final animation. Trial and error using the keyframe technique is presently the most popularly adopted solution by animators. Our main thesis is that for a character action, given just the keyframe representation or the entire animation sequence, we can recover a characterizing motion representation in lower dimension space using manifold learning. This characterizing motion recovers distinguishing information hidden in the huge amount of correlated and coherent character animation data in high dimensional space. Then we can use it to enhance keyframe animation techniques, which can considerably reduce animator effort required for specifying the keyframes for desired quality of animation. Further, these new techniques are equally applicable to 3D skeleton and mesh animation. In our first major contribution, we present a formulation to adopt the technique of locally linear embedding (LLE) to project the given character animation data into a much lower dimension embedding space. We show that animations depicting distinguishable activities, say walking, running, jumping, etc. take distinct characteristic shapes in lower dimensional embedding space. Based on this embedding, we present a new framework consisting of a reconstruction matrix combined with motion represented in the low dimension embedding space. This framework enables us to generate complete animation sequences in the original high dimensional space while maintaining desirable physical properties in a deformable shape. The latter is done by introducing the concept of a property map in the embedding space of values for the different physical properties of the mesh, for example area, volume, etc. A probability distribution function in embedding space then helps us rapidly choose the required number of in-between poses with desired physical properties. The reconstruction of the animation sequence is achieved by using the whole set of keyframes during interpolation for generating each in-between. This framework has many other applications and we demonstrate this by introducing two other new techniques which further enhance key frame animation. In another contribution of this research, we present a non-physically based method for incorporating perceivable variations in repetitive motion of an autonomous virtual character while retaining its principal characteristics. Usually, this is rather difficult to achieve using standard keyframe animation techniques, since even small changes in keyframes can result in less predictable changes in the final interpolated animation. The basis for our method is provided by asymmetric bilinear factorization of a given animation derived using the above framework. Keeping the action unchanged (that is the characterizing motion extracted as the embedding in the low dimensional space), we define a method to incorporate controlled perturbations into the reconstruction matrix so as to yield variations of the same motion. Further, to join the varied motion segments into a longer animation sequence, we present an embedding space method, which again makes use of the property map to maintain the desired physical properties. We also present an effective method for optimized keyframe selection from complete animations or motion capture sequences. Given the fact that most animators are very comfortable with the keyframe animation technique, this will enable animators to work easily with previously created animations or motion capture data. Our solution uses animation saliency and combines it with the embedding. For this, we use the representation of the animation in the form of matrix multiplication of reconstruction matrix with combination weights and then cast the keyframe selection problem into a constrained matrix factorization problem. The error metric that is minimized however uses animation saliency computed in the original high dimensional space. This way, the time consuming optimal search required by the matrix factorization problem in high dimensional space is simplified to a much more efficient method in low dimensional space. These enhancements to the character keyframe animation technique are possible because of capability of the manifold learning technique like LLE to effectively capture in low dimensional space the characterizing motion information that is present in a given character animation. Together these form the most significant contributions of this thesis.

碩士<br>國立成功大學<br>資訊工程學系碩博士班<br>97<br>A novel robust blind watermarking algorithm for three-dimensional (3D) keyframe animation is proposed in this thesis for copyright protection. In this algorithm, the geodesic distance of each vertex pair on the 3D mesh model extracted from a 3D keyframe animation to be watermarked is calculated, and then this model is segmented into meaningful pose-invariant components based on the calculation result. Each meaningful component is further divided into non-overlapping cells according to the distribution of geodesic distances from its vertices to the cutting boundary. Moreover, the defined watermark is embedded into these cells by adjusting the mean value of geodesic distances according to surface roughness estimation. Finally, we insert this watermarked model to the original animation content to obtain a watermarked 3D keyframe animation. Our experimental results demonstrate that the proposed algorithm has the ability to resist the common attacks applied to the 3D animation content, these attacks include timeline attacks, pose changing, similarity transformations, vertex reordering, local deformation, cropping, simplification, additive random noise, and smoothing.

碩士<br>國立臺灣大學<br>資訊工程學研究所<br>106<br>The algorithm of multi-robot cooperative localization has become more and more powerful. However, it is needed to consider the limited computing resources and network resources in implementation. Therefore, in order to reduce the use of redundant information, many studies make use keyframe detection to measure the importance of information before localization process. In robot soccer scenario, we adopt multi-robot cooperative localization and tracking, or MR-SLAT, which can provide good localization result under stable communication situation. But, in the measurement update step, the number of measurements sharing among robots can cause significant overhead to the system. Existing keyframe detection solutions against image content can be mainly classified into two categories, color-based and feature-based. Color-based approaches execute efficiently, but they lose the spacing information of interested objects. Feature-based approaches provide more robust performance but spend more computational cost for extracting features. This thesis base on the MR-SLAT framework and propose a keyframe selection mechanism as a preprocessing for MR-SLAT. We design an utility function based on Shannon Entropy to provide a mechanism for calculating the utility of measurements. The proposed utility function, without the use of feature points, measure low-level data including moving objects and stationary objects which are extracted from an image. The experimental results show that, our method can detect the change of scene. In localization experiment, the trajectory of localization can be significantly improved for up to 3x from odometry by using 25% of image data. when the trajectory error ratio between using all frames and odometry is close to 1, the best performance of the keyframe-based localization system is about 85% of by using all frames.

碩士<br>國立中央大學<br>資訊工程研究所<br>94<br>In this thesis, an intelligent video-based surveillance system for the detection of dangerous situations related to the presence of abandoned objects in unattended or guarded environment is presented. In the proposed system, image processing content-based retrieval capabilities have been added to make the operator inspection task easier. When a potential abandoned object event is detected, the operator is notified and the system provides a keyframe for interpreting the incident. In addition to tracking a moving person and deciding if he drops an object automatically, the system also detects the presence of abandoned object in a monitored area simultaneously. Depending on the detection of either an abandoned object or someone abandoning an object, different alarms will be signaled. Then, the system finds the relationship between the causes of the alarms. If it is verified that someone has lost the abandoned object, in other words, the owner and abandoned objects are detected, the system will alert the operator and provide him the keyframe, i.e., semantic information of which relating to both of the dangerous object and the person who left it in the surveillance environment. Experimental results are illustrated in terms of abandoned object detection, owner detection, keyframe detection, the probability of correct detection of abandoned objects as well as examples of abandoned objects event sequences. The results demonstrate that the proposed system is feasible and effective in abandoned object detection.

Στην παρούσα εργασία υλοποιούμε ένα σύστημα δημιουργίας περιλήψεων από ακολουθίες βίντεο. Υλοποιούνται όλα τα βήματα που θα πρέπει να ακολουθηθούν (εξαγωγή χαρακτηριστικών-ανίχνευση πλάνων-εξαγωγή χαρακτηριστικών καρέ) έτσι ώστε να εξαχθεί ένα σύνολο καρέ (χαρακτηριστικά καρέ) τα οποία να συνοψίζουν νοηματικά το περιεχόμενο μιας ακολουθίας βίντεο. Η επεξεργασία του βίντεο γίνεται απευθείας στο συμπιεσμένο πεδίο και συγκεκριμένα σε συμπιεσμένα αρχεία MPEG-1-2, έτσι ώστε τα αποτελέσματα να εξάγονται σε σχετικά μικρό χρόνο και με σχετικά χαμηλές απαιτήσεις σε αποθηκευτικό χώρο και επεξεργαστική ισχύ.<br>In this paper a video summarization system is being constructed. We acomplish all the needed steps (feature extraction -shot detection-keyframe extraction) in order to extract a set of frames (keyframes) that capture the semantic content of the video sequence. The processing of the video takes place directly at the compressed domain (at MPEG-1-2 video files). Thus we obtain results at relatively little time and with relatively low storage and computer power demands.