· · · Institutional Repository

Automated landmark detection may prove important for the examination and automatic analysis of real-time three-dimensional (3D) echocardiograms. By detecting 3D anatomical landmark points, the standard anatomical views can be extracted automatically in 3D ultrasound images of left ventricle, for better standardization and objective diagnosis. Furthermore, the landmarks can serve as an initialization for other analysis methods, such as segmentation. In this thesis we describe an algorithm that iteratively applies landmark detection in perpendicular planes of the 3D dataset. The landmark detection exploits a large database of expert annotated images, using an extensive set of Haar wavelet-like features for classification, resulting in fast detection times suitable for real-time applications. The detection is performed using two cascades of Adaboost classifiers, that work in different 2D planes, in a coarse to fine scheme. The method is evaluated by measuring the total detection error for the landmark points between the detected positions and the manual ones.

Traffic congestion has been one of the major problems in our everyday life. It has caused billions and billions of dollars including loss of productivity in our everyday life. Many institutions, governmental organizations, and academics have tried to come up with many initiatives and solutions to minimize congestion. Our research on dynamic routing is meant to help minimizing this problem from individual perspective since we believe that traffic congestion is also directly related and can be influenced by individual’s decision making behavior. We focused on finding a solution that can be integrated to an existing system by using a routing algorithm that can forecast traffic congestion and finding the fastest route to get to the end point. The forecast will be based on the historical traffic information. Deriving on the notion that most traffic congestions happen almost at the same time during the day and the same hours daily, we have combined the historical real traffic data to be used in an algorithm to help an effective Dynamic Intelligent Algorithm for Navigation (DIANa) to avoid congestion or other road blockage.

In spite of numerous road management schemes and developed infrastructure the society nowadays still faces the problem of highly congested roads due to the increasing traffic demand. The focus of this thesis is to develop a complex and integrated system that addresses the challenges of dynamic traffic assignment in modern times. To reach this goal we had to study the theory behind such systems, come with new ideas and implement them into a real world human-centered traveler information system. Taking into account state of the art systems in this field we built a design for a Personal Advanced Traveler Assistant (PTA). The main purpose of PTA is to give routing advices depending on the users preferences and the available capacity in the network. The system incorporates a dynamic traffic algorithm that employs a prediction model. The prediction model should use current traffic updates and the routes of the guided cars in order to achieve an accurate prediction of the future traveling times. Next step was to implement a prototype. This is available through multiple interfaces like smart phones or desktop applications. The prediction model that the algorithm uses is based on historical data. Given the time constraints the prototype uses now only historical data but it can be easily extended to include live updates. This allowed us to build a prediction model that considers typical traffic dynamics in order to estimate the traveling time. The performance measure of the traffic assignment algorithm is the shortest traveling time. The algorithm that we implemented is a time dimensional-extended version of the Dijkstra algorithm. The algorithm is tested on real data collected from the highways in The Netherlands. We evaluate the performance of the algorithm by comparing it to two versions of the static Dijkstra algorithm. To conclude, we successfully implemented a working prototype that uses various technologies such as Java, the Open Street Map API for rendering the map or J2ME for the mobile phone client. The prototype that we have built represents a working proof of concept for a dynamic routing assistant. One advantage of the structure that we chose to implement is that each component can be further on extended independently. In this way we showed that such a system is feasible but we also left the possibility for different parts of the system to be extended into more advanced applications.

Shopping is a daily common activity for all individuals. In that context, there are needs related to security, efficiency and satisfaction. Store owners, customers and producer brands are three parties sharing these needs and there are intelligent systems that offer solutions. Intelligent systems could use the video information from the store and they can offer basic information about what the people in the store are doing by interpreting body language. In this research to address those questions we aim to design a system which can automatically detect the basic actions in a store that are performed by people. We define the basic actions that are most commonly observed in a store and by considering sequences of these actions, higher level information about customers’ behavior can be extracted. We set up a shopping environment for experiments and made recordings in which people are doing shopping and performing the defined basic actions. We analyze the obtained data to examine the common properties and patterns of shopping related actions. Next step is to extract the discriminative features from those recordings that can reveal the actions they belong to. For that part we use two tools, ETH Human Pose Detection framework and Kinect sensor developed by Microsoft. ETH Tool detects and returns the limbs of a person in the scene and we use this information to extract the angles between the limbs automatically. Kinect is capable of returning the depth information, people’s silhouettes and if configured properly also the body skeleton coordinates. Furthermore the information obtained from silhouettes and body skeleton coordinates are used to extract different types of features. Next we evaluate the two tools and the sets of features with different classifiers by employing the developed automatic action detection software module. To conclude we examine the shopping store data, evaluate ETH and Kinect tools with different sets of features and yield to conclusions about those actions and the problem itself. The action detection performance is not very high yet that leads us to a lot of interpretations and deeper knowledge about those actions and possible solutions for addressing the challenges of the analyzed problem.

Traditional content creation for computer games is a costly process. In particular, current techniques for authoring destructible behaviour are labour intensive and often limited to a single object basis. We aim to create an intuitive approach which allows designers to visually define destructible behaviour for objects in a reusable manner, which can then be applied in real-time. First we present a short introduction into the way that destruction has been done in games for many years. To better understand the physical processes that are being replicated, we present some information on how destruction works in the real world, and the high level approaches that have developed to simulate these processes. Using criteria gathered from industry professionals, we survey previous research work and determine their usability in a game development context. The approach which suits these criteria best is then selected as the basis for the approach presented in this work. By examining commercial solutions the shortcomings of existing technologies are determined to establish a solution direction. To separate destructible behaviour from particular objects, we introduce the concept of destructible materials: where the material of an object usually defines the way an object looks, a destructible material will determine how it breaks. Destructible materials provide a reusable definition and intuitive way of designing and tweaking destructible behaviour of objects, which can then be applied in real-time. Using a prototype implementation we show the viability of the presented approach and how it extends previous research with reusability, making it more designer friendly and allowing the same destructible behaviour to be easily applied to different objects. While the prototype can only apply this destructible behaviour in real-time for simple cases, it still takes us a step in the right direction.

There exist several formalisms for representation and reasoning in dynamic systems, for example, Dynamic Influence Diagrams (DID), Influence Diagrams (ID), Dynamic Bayesian Networks (DBN), Bayesian Networks (BN), Hidden Markov Models (HMM), Markov Decision Processes (MDP), and Partially Observable Markov Decision Processes (POMDP). All these formalisms belong to graphical models based on probability theory. It has been shown that all probability models can be seen as variants of one generalization model. The purpose of this thesis is to review these models, to try to propose a unifying representation of these models at some generalization level (assuming DID level), and to test them in practice.

The demands for the quality and scale of videogames have become so high in the last few years, that game studios are having increasing difficulties meeting them, unless the development process is changed. The goals of this project are to create a platform aimed at supporting a new iterative game development paradigm, improving collaboration within studios, automating trivial tasks, supporting project management and improving the overall workflow. The solution we came up with is a framework that has a modular design and was built with usability in mind. Different components are created to facilitate the different goals for this platform. Together, these components form a solid framework to further build the Cannibal Game Development Platform on.

Since the stone age the human race seeks for strategies to extend its viewing range. With the rise of technology in the twentieth century, cameras are found to be a very useful tool to survey a large area with limited resources. With an increasing numbers of cameras, it becomes more difficult to watch every monitor and prevent incidents in the surveillance area. For the last decades, research seeks for possibilities to automatize the process of video surveillance. For this thesis, we approach the surveillance task from the human perspective: we try to emulate what human operators do when they watch the monitors. To perform this task, state-of-the-art techniques from Computer Vision and Artificial Intelligence are applied. An object tracking technique called P-N Learning is used that enables the tracker to learn from its mistakes. The Java Agent Development Framework (JADE) is used to enable communication between agents in the FIPA Agent Communication Language standard. A surveillance system model is designed that detects suspicious behavior in a non-public area. Its task is to alert the operators about suspicious events to give them the chance to investigate and take action. Two prototype applications are implemented and experiments are conducted to show the performance. We showed the proof-of-concept of a system which is able to emulate operators and can potentially outperform a human being. Once the system knows what is considered suspicious behavior it can be automatically detected.

In this work we propose a template-based tracker for simultaneously tracking multiple targets such as facial features. Our work falls within the particle fil- tering framework and more specifically it extends the Particle Filtering with Factorized Likelihoods (PFFL) tracking scheme. In this work we introduce the idea of using a distribution per constellation of facial features that. These distri- butions are updated in a two-stage observation process in which appearance and shape likelihoods are evaluated separately. Information from all distributions is used to estimate the head pose, which is needed to apply the shape prior onto the observations. In this work we propose the use of a hierarchical shape prior that is modeled to have at its root the shape described by the centroids of the feature constellations and its leaf nodes the shapes for the feature constellations.

Object Tracking has been a very active area in the field of C omputer Vision. Over the years, a variety of approaches have been put forth to solve this problem and though many of them have demonstrate considerable success none of them have been completely successful. With more methods being written each day, the evaluation of such systems becomes a very important task. If an evaluation system exists that is able to point out specific flaws in the stage of development, it can lead to a very robust and improved algorithm. This work attempts to create such an evaluation framework. Given an algorithm that detects people and simultaneously tracks them, we evaluate its output by considering the complexity of the input scene. Some videos used for the evaluation are recorded using the Kinect sensor and a benchmark dataset from the PETS workshop is also used. To analyze the performance of the tracking system,the reasons due to which the algorithm might fail are investigated and quantified over the entire video sequence. A set of features called Scene C omplexity Measures are obtained for each input frame. The variability in the algorithm performance is modeled by these complexity measures using various regression models. From the regression statistics, we show that we can compare the performance of two different algorithms and also quantify the relative influence of the scene complexity measures on a given algorithm.

Robust automatic detection of surface and air objects in a maritime environment is a problem that is of growing importance to the Royal Netherlands Navy (RNLN). Due to a shift in the field of operation from the open oceans towards the littoral waters, the RNLN is forced to operate in complex environments with cluttered backgrounds and the presence of many small vessels and a wide range of other objects. Traditional radar systems are not optimal in these circumstances due to their minimum detection range, lack of sensitivity to small, non-metallic, objects and poor classification power. Complementation by Electro-Optical (EO) camera systems is therefore desired, which resulted in the start of the development of a detection algorithm based on polynomial background estimation. Automated object detection in the maritime environment is a complex problem however, due to various complicating factors. These factors include the highly dynamic background, camera motion, the variety in possible objects and their appearance, and the diversity in meteorological as well as environmental circumstances. Although the developed detection algorithm is quite well capable of detecting the objects, it also produces an extensive amount of false detections. This study investigates whether these false detections can be eliminated, while maintaining the true detections, by means of classification of the detections as either target or background. To this end, the initial detection algorithm is optimised to detect as much objects as possible in a carefully constructed dataset of eight hundred Visible Light (VL) images. The resulting detections from the optimised algorithm are used accordingly to train and test various basic classifiers, using a set of features found in the literature. The best performing classifier is selected and the performance of this classifier, and the two-stage detection algorithm as a whole, is subsequently further analysed by means of various tests involving the features used, the evaluation procedure and the fusion of detection results. Results show that especially the features as well as the clustering procedure for detected pixels are important parameters with respect to a good performance of the algorithm. This works shows that the linear discriminant classifier is best to use with the problem among the classifiers considered. Moreover, it is demonstrated that including features of histogram equalized boxes in combination with features of the entire image increased the performance the most, that determining the features on a slightly larger area than the originally detected area is beneficial and that fusion of detections after classification can be used to optimise the detector output. Although the developed classification approach is capable of eliminating many false detections and to retain a majority of the true detections, further research is required. Suggested are separate classifiers for the sea- and sky part, inclusion of the time dimension, optimisation of the operating point of the classifier and preprocessing steps.

The human eye is a complex organ of the human body which is able to cope with a lot of different situations. One of these situations is adapting to different light intensities by changing the dynamic range of the eye, so-called temporal adaptation. This report gives insight into what temporal adaptation means for luminance visibility while driving on an urban street during nighttime. The first step needed was collecting data of luminance values falling onto the retina while driving in the dark. As driving with a head-mounted eye-tracker was impractical and insufficiently accurate, the experimental set-up was changed into a lab set-up. An HD camera was used to record a video during driving in the dark, while a luminance camera was used to measure the corresponding luminance values. The resulting video was shown to 20 participants in the lab, and they were asked to watch the movie as if they were driving themselves. Their gaze positions on the screen were measured using a table mounted eye tracker. The gaze positions were converted into saliency maps, which were then combined with the luminance values of the video frames to generate weighted luminance maps. From the weighted luminance maps luminance profiles were deduced, representing the luminance value falling on the center of the retina as a function of time. The luminance values measured in these profiles varied between about 1 and 18,000 cd/m2. About 54% of the fluctuations in luminance within a one-second time frame had a difference below 1000 cd/m2, while about 6% of the fluctuations had a difference larger than 10,000 cd/m2. The effect of temporal adaptation on these luminance fluctuations was estimated by superimposing a temporal adaptation model existing in literature on the luminance profiles. Light adaptation (i.e. from a dark to a bright environment) is so fast that we considered it instantaneous for our application. Dark adaptation (i.e. from a bright to a dark environment) is somewhat slower and so was expected to delay the perception of the lower luminance values. However, also for dark adaptation the initial drop in perceived luminance is so fast that temporal adaptation only had a small effect on perceived luminance for the highest light intensity fluctuations. It should be noted, however, that literature on temporal adaptation only reports adaptation to homogeneous light intensity fields, while in our particular case of driving on an urban street during nighttime the visual field was characterized by a rapidly changing spatially complex light intensity field. Models to describe temporal adaptation to spatially complex light fields do not exist yet, and so, may affect the conclusions of our research.

In this thesis, a projector-based Spatial Augmented Reality (SAR) system designed and developed to be applied to support physical and virtual 3D Rapid Prototyping in the field of Industrial Design Engineering is presented. The main contribution is a 3D scanner to get a virtual model of a physical model and tools to support the design of features interactively, on the object’s surface. More specifically, this work contains an approach to set the hardware to support SAR, else known as hardware calibration for SAR. Each hardware entity is calibrated with respect to a common “world”, in order to achieve effective communication. This world is set to coincide with the graphics world, and this allows us to imagine being inside the 3D graphics world while the virtual content is rendered onto the scene’s objects. In order to identify the basic ingredients that enable interaction in our SAR system, we take into account the limitations of Rapid Prototyping process, background knowledge for SAR systems and related work. Therefore, we designed the interaction components of the system according to characteristics of our setting. The SAR system was designed to perform in the peripersonal region. In this region, the user inserts input via a constructed IR tracked pen and a dynamic menu is used as interface to the system. Functionality such as selection, feedback and annotation is enabled for interacting with the SAR system. The system’s application is divided into two parts. The first part includes the use of the RGB-D camera and the IR tracker for the construction of a 3D scanner, in order to produce a virtual model of an object through sampling, segmentation, and registration of sequential point clouds. In the second part, the result of the scanning process, which is a polygonal mesh of the scanned object, is added to the SAR system’s application that enables interaction with virtual models and the ground level of the world. These two parts of the SAR system application aim to support industrial designers in the scanning of a freshly made physical prototype, and enable the design of features on the corresponding virtual model by using the SAR system during the rapid prototyping process. In order to identify the strong points and weaknesses of the current state of the SAR system application, we carried out a user evaluation. 21 students from the Faculty of Industrial Design Engineering evaluated the two parts of the system’s application. The results show that the SAR system is useful and that it has great potential in the field of Industrial Design Engineering. Nevertheless, there is still room for improvement and future work, in order to be fully applied in the field.

Nowadays almost everything is digital. Even testing new products can be done by simulating them digitally, also called virtual prototyping. This report describes research that is done to improve virtual prototyping at Philips Research. Two experiments have been performed with the aim to evaluate how image quality and realism are affected by the integrator, i.e. the method used to calculate the light at each point in the scene, and the time spent on the rendering process. The first experiment was conducted to evaluate whether a different integrator than an implementation of a standard path tracer might result in more realistic renderings. Mitsuba was used as render software and support integrators both based on path tracing and photon mapping. Two images of the same scene were simultaneously shown to 20 participants separately. The participants had to indicate which of the two they considered most realistic. They had to do this for different lighting conditions and with and without a reference to the real scene. The results showed that people preferred a photon based integrator over the path based integrators. The second experiment focused more on how material characteristics influence the rendering time. Renderings of a teapot with different materials and under different lighting conditions were created with Mitsuba and Indigo software. Participants had to compare a rendering to a reference rendering, which was part of a standard quality ruler, based on one type of material and light condition. As a consequence, they had to compare quality across different materials and lighting conditions. The results were widely spread across the 20 participants, mainly because comparing quality of different materials was a difficult task. The task became even more complicated because of the different kinds of noise generated by the different integrators. To overcome the issue found in the second experiment, a follow-up study was conducted. In this follow-up study the participants not only saw a stimulus and a ruler rendering but also the perfect versions of those two. The task was now to identify if the stimulus or the ruler rendering had the smallest difference with their perfect version. The results for the 4 participants that performed this follow-up study were less spread than those of the second experiment. Therefore the interface used in the follow-up study is considered to be useful for future experiments where comparisons across different materials are needed.