The accuracy of the position data is examinated by collection position data from a Shipmate RS58000 GPS reciever connected to a PC.

The background for USA's Selective Availablity politic and developments in American politics on this subject is examined.

Furthermore Differential GPS and the possibility of combining GPS and the Russian GLONASS system is treated.

Report in Adobe PDF (813k)

Sourcecode (1M)

On basis of the construction of a software-based digital mixer device with a 3-band equalizer, it is demonstrated that the new development leads to a series of advantages.

Based on the experience gathered from the construction, the requirements for software development methods used for development of this kind of software is considered.

It is concluded that there is a need to focus on developing new methods for software development.

Report in Adobe PDF (538k)

Sourcecode (144k)

Report in Adobe PDF (1.3M)

Report in Adobe PDF (2.1M)

Sourcecode (155k)

Report in Adobe PDF (1.4M)

Sourcecode (40k)

For training ALSAC utilizes a clustering mechanism, incorporating the Dynamic Time Warping (DTW) algorithm. The recognition is implemented using the OnePass algorithm. Modularity is obtained by use of the Strategy Design Pattern. This is done to ensure the possibility to change algorithms (strategy) of the various parts independently of the rest of the system. The ALSAC-system is a C++ library which ITE can use as a building block in their products. It is concluded that a system performing the intended task of classifying the contained sequence of visemes and time stamps in speech signals has been implemented correctly. However the test shows that ALSAC does not produce an accuracy score as high as required. This can partly be explained by the fact, that the training data used were not optimal. It is expected that with better training data and adjusted system parameters the performance of ALSAC can be increased. ALSAC is believed to be able to accelerate the currently manual animation process at ITE, even though the system's accuracy score was not found to be optimal in the system's present state.

Report in Adobe PDF (2.1M)

Sourcecode (50k)

Article in Adobe PDF (52K)

Slides in Magicpoint (215K)

Semcon Poster in .ps.gz (61K)

The system can in general be divided into 3 core parts; segmentation, stabilization, and tracking.

The stabilization, which estimates the orientation of the camera relative to the surroundings, utilizes window based correlation to estimate the disparity between images and a 2D affine method to convert the disparity to angles in world-space.

The segmentation of objects is solved by using a bipartite approach. Initially pixels belonging to the table are found using global thresholding in the hue channel of the HSV color model, and subsequently pixels belonging to objects are found using the convex hull.

The tracking part, is based on feature extraction of among others invariant moments and histograms. Mahalanobis distance is used for the feature matching, and feature prediction is based on simple linear extrapolation. Furthermore, a general window based object recognition method for handling occlusions is proposed which works on histograms in the HSV color space.

The system is designed using Matlab, and a prototype is implemented in C++ based on the Intel Corporation's Open Computer Vision Library (OpenCV). The prototype contains the segmentation and tracking component of the system. Occlusions are, however, not handled currently.

The test of the prototype showed that the system is capable of tracking separate objects and is robust with respect to scaling and rotations. Unique objects are tracked more robustly than objects with similar color. The demand regarding a frame rate of 10 frames per second has been tested and confirmed, as long as there are 6 or fewer objects in the camera's field of view.

It is concluded that the system with the chosen delimitations works as specified, but a number of future improvements is suggested. Specifically stabilization and the method for handling occlusions must be added to the current prototype in order to achieve better results.

Lowres Adobe PDF (3M)

Slides (Danish) (5.2M)

The developed system consists of 3 parts: Generation of initial occlusion mask using background subtraction, refinement of occlusion mask using visual hull with post processing, and visualization.

The initial occlusion mask is created using background subtraction after an image warping step which transforms images into a reference viewpoint.

Refinement is done in two steps: First edges are refined using a modified visual hull algorithm, and secondly interior holes are classified as real holes or is caused by camouflage effects using an algorithm developed by the project group.

Visualization produces the final augmented image using the calculated occlusion mask. The method presented is optimized for the capabilities of modern 3D hardware.

A prototype has been implemented in C++ using OpenGL and The Visualization Toolkit (VTK) on a Linux platform.

Test of the prototype show that the background subtraction process is able to generate a good initial occlusion mask even in cases with complex silhouettes. Test on synthetic data show that the two-step refinement process is able to eliminate errors in the initial occlusion mask.

It is concluded that the augmented images generated by the developed system contain less visual artifacts than using the raw depth information of the reconstruction system, even without the refinement step. Adding the refinement step will most likely decrease the number of artifacts caused by camouflage effects in the interior of the silhouettes.

Lowres gzipped Postscript (3.7M)

Slides (Danish) (1.3M)

MPEG Videos
Botanical Sequence
LQ (2.2M)  HQ (11M)

Image Warping
LQ (1.3M)  HQ (8M)

Background Subtraction
LQ (1.2M)  HQ (8M)

Augmented Sequences
LQ (3.8M)  HQ (23M)