Scene understanding is the process, often real time, of perceiving, analysing and elaborating an interpretation of a 3D dynamic scene observed through a network of sensors. This process consists mainly in matching signal information coming from sensors observing the scene with models which humans are using to understand the scene. Based on that, scene understanding is both adding and extracting semantic from the sensor data characterizing a scene. This scene can contain a number of physical objects of various types (e.g. people, vehicle) interacting with each others or with their environment (e.g. equipment) more or less structured. The scene can last few instants (e.g. the fall of a person) or few months (e.g. the depression of a person), can be limited to a laboratory slide observed through a microscope or go beyond the size of a city. Sensors include usually cameras (e.g. omni directional, infrared), but also may include microphones and other sensors (e.g. optical cells, contact sensors, physiological sensors, radars, smoke detectors). Scene understanding is influenced by cognitive vision and it requires at least the melding of three areas: computer vision, cognition and software engineering. Scene understanding can achieve four levels of generic computer vision functionality of detection, localisation, recognition and understanding. But scene understanding systems go beyond the detection of visual features such as corners, edges and moving regions to extract information related to the physical world which is meaningful for human operators. Its requirement is also to achieve more robust, resilient, adaptable computer vision functionalities by endowing them with a cognitive faculty: the ability to learn, adapt, weigh alternative solutions, and develop new strategies for analysis and interpretation. The key characteristic of a scene understanding system is its capacity to exhibit robust performance even in circumstances that were not foreseen when it was designed. Furthermore, a scene understanding system should be able to anticipate events and adapt its operation accordingly. Ideally, a scene understanding system should be able to adapt to novel variations of the current environment to generalize to new context and application domains and interpret the intent of underlying behaviours to predict future configurations of the environment, and to communicate an understanding of the scene to other systems, including humans. Related but different domains are robotic, where systems can interfere and modify their environment, and multi-media document analysis (e.g. video retrieval), where limited contextual information is available.

Cette thèse comprend deux axes de recherche. D´abord, un nouveau cadre méthodologique pour évaluer la conformité des données RSO (Radar à Synthèse d’Ouverture) multivariées à haute résolution spatiale est proposé en termes de statistique asymptotique par rapport au modèle produit. Plus précisément, la symétrie sphérique est étudiée en appliquant un test d'hypothèses sur la structure de la matrice de quadri-covariance. Deux jeux de données, simulées et réelles, sont prises en considération pour étudier la performance du test obtenu par l’analyse qualitative et quantitative des résultats. La conclusion la plus importante, en ce qui concerne la méthodologie employée dans l'analyse des données RSO multivariées, est que, selon les différents cas d’usages, une partie considérable de données hétérogènes peut ne pas s’ajuster asymptotiquement au modèle produit. Par conséquent, les algorithmes de classification et/ou détection conventionnels développés sur la base de celui-ci deviennent sub-optimaux. Cette observation met en évidence la nécessité de développer de modèles plus sophistiqués comme l'Analyse en Composantes Indépendantes, ce qui conduit à la deuxième partie de cette thèse qui consiste en l’étude du biais d’estimation des paramètres TSVM (Target Scattering Vector Model) lorsque l’ACP est utilisée. Enfin, les performances de l'algorithme sont également évaluées sous l'hypothèse du bruit gaussien corrélé spatialement. L’évaluation théorique de l'ACI comme un outil de type ICTD (In Coherent Target Decomposition) polarimétrique permet une analyse plus efficace de l’apport d’information fourni. A ce but, deux espaces de représentation sont utilisé, notamment H /alpha et TSVM / This thesis comprises two research axes. First, a new methodological framework to assess the conformity of multivariate high-resolution Synthetic Aperture Radar (SAR) data with respect to the Spherically Invariant Random Vector model in terms of asymptotic statistics is proposed. More precisely, spherical symmetry is investigated by applying statistical hypotheses testing on the structure of the quadricovariance matrix. Both simulated and real data are taken into consideration to investigate the performance of the derived test by a detailed qualitative and quantitative analysis. The most important conclusion drawn, regarding the methodology employed in analysing SAR data, is that, depending on the scenario under study, a considerable portion of high heterogeneous data may not fit the aforementioned model. Therefore, traditional detection and classification algorithms developed based on the latter become sub-optimal when applied in such kind of regions. This assertion highlights for the need of the development of model independent algorithms, like the Independent Component Analysis, what leads to the second part of the thesis. A Monte Carlo approach is performed in order to investigate the bias in estimating the Touzi's Target Scattering Vector Model (TSVM) parameters when ICA is employed using a sliding window approach under different scenarios. Finally, the performance of the algorithm is also evaluated under Gaussian clutter assumption and when spatial correlation is introduced in the model. These theoretical assessment of ICA based ICTD enables a more efficient analysis of the potential new information provided by the ICA based ICTD. Both Touzi TSVM as well as Cloude and Pottier H/alpha feature space are then taken into consideration for that purpose. The combined use of ICA and Touzi TSVM is straightforward, indicating new, but not groundbreaking information, when compared to the Eigenvector approach. Nevertheless, the analysis of the combined use of ICA and Cloude and Pottier H/alpha feature space revealed a potential aspect of the Independent Component Analysis based ICTD, which can not be matched by the Eigenvector approach. ICA does not introduce any unfeasible region in the H/alpha plane, increasing the range of possible natural phenomenons depicted in the aforementioned feature space.