| note
| - Le sujet de la thèse s'inscrit dans le cadre du projet \"Véhicule intelligent et son intégration dans la ville du futur\" mené au laboratoire Systèmes et Transports de l'Université de Technologie de Belfort-Montbéliard. L'objectif de ce projet est d'assurer la navigation autonome d'un véhicule dans un environnement urbain. Cette thèse s'intéresse plus particulièrement au problème de la perceptioon de l'environnement du véhicule en combinant plusieurs capteurs. Le but est de détecter et suivre des objets dynamiques et de les situer par rapport au véhicule instrumenté. La contribution de la thèse commence par la proposition d'une nouvelle technique de représentation des objets. Cette technique est basée sur l'utilisation des boîtes englobantes orientée (OBB) et exploite deux paradigmes qui sont l'incertitude Inter-Rays (IR) et l'hypothèse de la taille fixe des objets (FS). Pour augmenter la qualité de l'estimation d'état des objets et du suivi, l'algorithme de fusion de deux télémètres laser est présenté. Enfin, deux méthodes d'association de données sont décrites. La première, appelée NNF, est une adaptation de la technique du plus proche voisin à la nouvelle technique de représentation. La deuxième méthode permet de résoudre le problème de clustering des données télémétrique par une fusion d'un télémètre laser et d'un capteur stéréoscopique. Les algorithmes proposés sont testés et évalués à l'aide d'un simulateur développé dans le cadre de la thèse et sur un prototype de véhicule électrique.
- The subject of the thesis lies within the scope of the project \"Intelligent Vehicle and its integration in the city of the future\", led by the Systems and Transportation laboratory of the University of Technology of Belfort-Montbéliard. The objective of this project is to ensure a vehicle autonomous navigation in an urban environment. The autonomous navigation of a vehicle consists of a data processing sequence made up by three stages. The first one relates to the perception of the vehicle environment. In this stage, a dynamic map of entities present around the vehicle is constructed. The second stage consists of localizing globally the vehicle in its environment. The third stage is a path planning of the vehicle displacements, while avoiding obstacles and collisions with dynamic objects. This thesis concerns in particular the perception of the dynamic objects in the vehicle environment by combining several sensors. This is a core part of the autonomous navigation system. The goal is to detect and track the dynamic objects (cars and/or pedestrians) and locate them relatively to the instrumented vehicle. The estimate of objects states is an entry data of path planning and collision avoidance algorithms. Multi-sensor based dynamic objects perception issue can be divided into three main parts: objects representation, data association, and state estimation and tracking. The contribution of the thesis starts by proposing a new Oriented Bounding Box (OBB) object representation model, including Inter-Rays uncertainty (IR) and Fixed Size assumption (FS). To increase the quality of object state estimation and tracking, a two laser scanners based fusion algorithm is presented. Finally, two methods for data association are described. The first one, called NNFS, is an adaptation of the Nearest-Neighbourhood (NN) principle to the OBB object representation using the FS assumption. The second method solves the problem of laser scanner data point clustering by fusing laser scanner (LS) and stereo-vision (SV) sensory data. It is shown in this thesis that the proposed representation model is adequate to be used in the task of object tracking. The integration of the IR and FS algorithms upgrades the quality of objects position and size estimation. The position and size estimation plays an important role in the data association part. The two laser scanners architecture allows an improvement of angular resolution, which is exploited in the fusion algorithm. The proposed fusion technique increases the tracking system reliability by better object velocity, size and orientation estimation. The NNFS data association algorithm gives reliable results. Indeed, it is able to separate coalescing objects (even if the objects touch each other). The second data association algorithm uses two different sensors. Based on this data association algorithm, the fusion method allows clustering correctly the ambiguous LS data points' configurations for which existing clustering approaches fail. To test and evaluate the proposed algorithm, a 3D dynamic based simulator is developed. In the simulator, several sensors are implemented: laser scanner (2D, and 3D), mono- and stereo-vision, odometer and GPS. The proposed simulation platform gives the possibility to create different scenarios by adding static and dynamic entities and defining paths to be followed. The platform based testing allows to find errors in early stages of implementation and to experiment different scenarios (including difficult or very expensive ones e.g. when vehicles touch each other). It also saves time by reducing real world tests. The proposed algorithms are tested and evaluated using the aforementioned simulator and an experimental vehicle (automated electrical vehicle), equipped with several sensors such as: 2D and 3D laser scanners, stereo-vision and GPS- RTK.
|
![[RDF Data]](/fct/images/sw-rdf-blue.png)
