· · · Institutional Repository

SKION is a foundation that facilitates research of children with cancer. The research consists of investigating slides and other material gotten from the treating hospitals at SKION. The results of these investigations are saved into a database, while other information gathered by SKION through forms is stored in another database. Four times a year, Dutch oncologists from all around the country meet in real life to review several slides. This is a very inefficient way of working, since all participants have to come to a central location. As a solution, Martijn Tijsma introduced the possibilities of a Digital Review. This requires an application that stores forms and combines data from the two databases and merges them on screen for all participants of the review to see. The goal of this project was to create a proof that such an application can be build (Proof of concept).

In dit onderzoeksverslag zullen we laten zien welke keuzes we in het ontwikkelproces hebben moeten maken. Een aantal keuzes konden we vooraf maken en hebben we bepaald in het vooronderzoek. Tijdens het project echter zijn we op een aantal problemen gestuit waardoor we de plannen toch moesten aanpassen. We laten daarom zien welke keuzes we konden behouden en welke we moesten aanpassen.

This document gives an overview of the progress and achievements of R. Spliet and S. Smit during their internship at SKION, including all delivered documentation.

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.

The goal of this research is to improve a system capable to detect, track a single person and recognize poses real time for controlling a spatial game. After performing background subtraction, the human blob is segmented in order to track the torso and hands. Angles and distances between hands and torso center are used to compute the features. Finally, a 10-Nearest-Neighbor classifier recognizes 9 predefined Poses which are used by the player to control the game. This work contributes with two improvements. The first one is a more robust improved hand detection combining the current skin color detection with human blob information. The second improvement is a classifier that recognizes Non-Poses in addition to the 9 predefined Poses that are used in the game.

In this thesis, we describe a system that performs online transformation estimation between pre-calibrated stereo cameras. This allows the stereo cameras to be moved around and automatically re-calibrated without the use of a calibration object. This also allows the set-up to deal with ad-hoc camera relocation and to recover from accidental nudges that invalidate the extrinsic (external to the stereo camera) calibration. The obtained transformations can for example be used in virtual view rendering for 3D Video. The relative positions and orientations of the stereo cameras are obtained using sparse point correspondences found in different views of the scene. For each stereo camera, 3D coordinates of salient scene points are triangulated and their image feature descriptors are used to locate the same points in the views of other stereo cameras. The salient point descriptors SIFT, ASIFT, SURF and FAST are compared for this purpose, together with two different error measures to see which selects the best transformation. Two strategies are proposed and evaluated to deal with three or more stereo cameras. Given enough salient image points, the proposed solution accurately finds the transformation between stereo camera pairs with a reprojection error less than 1 pixel. Additionally we use the obtained transformations to perform virtual view rendering and evaluate the perceptual quality.

With the current situation of bad economy and substantial increases in energy costs, people are becoming more and more eager to do something about their energy bill. Although people are willing to spend money to save money, few are willing to spend a lot of money on a complete insulation of their homes. There is a need for products that will help them point out the most cost effective remedies. One thing that is difficult to pinpoint is draft. Our company wants to sell a product that allows people to track where a draft is coming from. We imagine an electronic handheld device that can measure the strength and direction of draft, but we are open for suggestions. Therefore, we are looking for a company that is willing to take on a short-term project to invent and build a prototype raft-detector.

With the release of RGBD-cameras (cameras that provide both RGB as well as depth information) researchers have started evaluating how these devices can be used in fields of localization, mapping and ubiquitous computing. Intel Seattle Research proposed an indoor mapping algorithm making use of such a camera. The algorithm itself is vulnerable to violations of the static environment assumption and image based localization failures that can be caused by, for example, featureless environments. The goal of this master thesis is to augment the indoor mapping algorithm with additional Inertial Measurement Unit (IMU) data to enhance the robustness to these vulnerabilities. To this end the characteristics and limitations of the Microsoft Kinect are investigated and an enhanced mapping algorithm is proposed. IMU orientation estimates are fused with pose estimates based on image data, which give an initial guess to the Iterative Closest Point (ICP) algorithm that is used to align point cloud data to create a final map. In case visual localization fails, the algorithm of Intel uses a constant velocity assumption as fallback mechanism while the IMU data provide more accurate orientation estimations than the constant velocity assumption can provide. The IMU-enhanced algorithm shows similar mapping quality in ideal mapping conditions compared to the plain mapping algorithm. While a series of corner case tests show that the IMU-enhanced algorithm was unable to improve the results compared with the plain mapping algorithm, it potentially generates improvements in mapping quality when dealing with non-static environments.

In this master's thesis a framework for the rendering of scenes for 3DTV applications is presented. The goal of our work is to investigate ways to generate virtual views of a scene captured using a relatively small number of real cameras. The approach adopted uses per-pixel operations and therefore is suitable for fast parallel implementation. A depth map is implicitly generated for every of the virtual viewpoints. The approach is refined by the use of camera proximity weights and an outlier removal method. The rendering system can handle occlusions and can cope with a broad range of different sceneries. Our framework allows for flexible and scalable generation of a virtual view, in terms of the number of capturing cameras.