This thesis presents the development of an automatic mesh generator, named WoodFrameMesh, using object oriented C++. The program developed is capable of generating complete finite element models of wooden houses incorporating frames, linear links, springs, nodal loads and restraints at the desired locations. The finite element mesh generated by the program may be triangular or quadrilateral. The triangular mesh can be generated over any arbitrary domain with multiple openings and line constraints. The program implements the advancing front method for triangulation as discussed by Lee and Hobbs. The difference is made by implementing the algorithm using object oriented concepts and the extensive use of the powerful C++ Standard Template Library (STL). Quadrilateral mesh generation is limited to simple quadrilateral domains with no openings or constraint lines. A simple structured technique is implemented to generate the quadrilateral mesh. The amount of time spent in manual generation of the complete finite element model of wooden houses has been considerably reduced by automating the modeling process. Overall, the use of object oriented design has facilitated the code development and has provided a platform for further additions. The program relies on the use of STL as it provides dynamic data structures, algorithms for storage, searching, sorting, etc. Efficiency of the program is improved by the use of the in-built features in STL instead of developing new code. Analysis of the finite element models generated by the automatic mesh generator is performed using SAP 2000 and WoodFrameSolver. WoodFrameSolver is a finite element analysis engine for WoodFrameMesh, which was developed at Virginia Tech by a group of graduate students (including the author) and professors as a separate project. A chapter discussing the WoodFrameSolver architecture, its extensibility features and its verification is also presented in this thesis. The solver performance and accuracy are similar to those of SAP 2000, which was chosen as the benchmark for testing the analysis results. / Master of Science

Nous abordons le suivi d’objets dans des séquences d’images de scènes déformables selon deux axes de recherche. Il s’agit de déterminer les transformations d’un objet, d’une image à l’autre, lorsque celui-ci s’est éventuellement déformé ou déplacé et lorsque le point de vue de la caméra a éventuellement été modifié (déplacement, zoom...). Pour cela, nous nous sommes inspirés de l’algorithme de Jurie et Dhome qui permet de suivre un objet plan indéformable. D’une part, nous en améliorons les performances. D’autre part, nous le généralisons au cas d’objets déformables. Le premier axe de recherche consiste à améliorer les performances de l’algorithme de Jurie et Dhome en termes de précision et robustesse. Le suivi s’appuie sur un ensemble de points d’intérêt, dont dépendent fortement les performances. Ces points d’intérêt sont issus d’une sélection des points obtenus par des détecteurs reconnus, à savoir SIFT, KLT, SUSAN, HARRIS et MORAVEC. Nous avons étudié et mis en oeuvre, sur différentes classes d’images, des heuristiques de sélection fondées sur des approches statistique et algébrique. Nous montrons : – qu’il n’existe pas de détecteur universel, – que l’approche statistique est à privilégier dans tous les cas. Le second axe de recherche est une proposition d’un nouvel algorithme de suivi s’appuyant sur le maillage 2D des images de la séquence. Cet algorithme généralise celui de Jurie et Dhome aux scènes déformables. Il repose sur : – des transformations élémentaires (nodales) du maillage, directes et inverses, que nous avons caractérisées tant d’un point de vue géométrique qu’analytique, – l’utilisation des coordonnées barycentriques généralisées pour approcher la composition de deux transformations d’un maillage. Cet algorithme donne des résultats similaires à celui d’appariement hexagonal de Nakaya et Harashima tout en étant plus rapide. / We deal with object tracking in videos of non-rigid scenes with two main purposes. We aim at determining the transformations of an object, from one frame to the next, when it may be distorted or moved and when the camera focus may change (movement, zoom...). To do this, we were inspired by the Jurie and Dhome algorithm, which enables the tracking of plane rigid objects. On the one hand, we improve its performance. On the other hand, we generalize it to non-rigid objects. The first goal consists in improving the performance of the Jurie and Dhome algorithm, in terms of accuracy and robustness. The tracking is based on a set of interest points, which has a great effect on the algorithm’s performance. These interest points come from a selection among the points extracted with some common detectors: SIFT, KLT, SUSAN, HARRIS, and MORAVEC.With various pictures classes, we have studied and implemented some selection heuristics based on statistical or algebraic approaches. We show that : • there is no universal detector, • the statistical approach is the best in all cases. The second goal is a proposal of a new tracking algorithm based on a 2D mesh of the video frames. This algorithm generalizes the Jurie and Dhome one for non-rigid scenes. It is based on : • elementary (nodal), direct or inverse, mesh transformations that we geometrically and analytically characterize, • generalized barycentric coordinates to approximate the composition of two mesh transformations. This algorithm gives similar results to the hexagonal matching algorithm of Nakaya and Harashima while being faster.