Ph.D. Research

  • Detection and characterisation of vegetation in urban areas from high-resolution aerial imagery
    Abstract:
    Significant progress has been made in recent years concerning the automatic reconstruction of man-made objects or environments from multiple aerial images. Yet, a lot of challenge concerning the modelling of other objects present on the terrain surface, such as trees, shrubs, hedges, or lawns still exists. An accurate reconstruction of such types of vegetation areas is a challenge due to their complex nature and to their intricate distribution between man-made objects in dense urban areas. This thesis presents an image analysis system for vegetation detection and characterisation from high resolution colour infrared aerial imagery for 3D city modelling. The aim of the system's first module is to extract vegetation areas. The approach developed is based on a Support Vector Machines (SVM) classifier and its performances are compared to traditional remote sensing methods for vegetation detection. The system's following modules aim at characterising vegetation areas thus identified, according to their morphology. Separation into high- (tree) and low- (lawn) height vegetation areas is based on texture characteristics computed on the digital surface model (DSM). Individual tree crown delineation is performed by using a region-growing algorithm based on geometrical characteristics of trees. 3D morphological characteristics (height, crown diameter, tree trunk position) are estimated for each tree crown. A supervised classification to characterise each tree by its species is performed on each tree crown. The set of parameters extracted by each of the modules are used to enrich 3D city models by virtual realistic tree models.
  • Keywords:
    • Computer vision: object and scene recognition, object localization
  • Supervisors:
  • Funding: IGN & Terra Numerica Project, Cap Digital Business Cluster

Selected Past Projects

  • Street scene classification from street level panoramic images - iTowns ANR Project
     
      The goal of the iTowns project is to develop a new generation of multimedia web tools for 3D navigation through panoramic images acquired at street level and an image-based search engine based on intelligent queries. I workd on the task of street scene classification based on local descriptors. These works were performed in the framework of the iTowns ANR (French National Research Agency) project. The aim of this research was to evaluate the performances of state of the art local features and kernels approches to classify different types of street furniture (porches, buildings, ...), different types of buildings (commercial, residential, ...) and their styles (Haussmann, modern, ...).
     
     
  • Map revision system based on pattern detection from scanned traditional cartographic data
     
      An image processing application integrated in a map revision system based on images obtained from scanned traditional cartographic data. Its aim is to detect specific patterns and define their outline. These areas will later be replaced by an adequate uniform colour. The application is composed of a pattern recognition module based on the two-dimensional cross-correlation operation. The algorithm was tested in the map production department of IGN, Institut Géographique National and is currently used for map production.
     
 

Present & Past Affiliations

  • Institut National de Recherche en Informatique et Automatique (INRIA)
    Project Team DIGIPLANTE, Ecole Centrale Paris (ECP)
  • Machine Learning and Information Retrieval Team (MALIRE)
    Laboratoire d'Informatique de Paris 6 (LIP6), Université Pierre et Marie Curie (UPMC)
  • Institut Géographique National (IGN)
    Laboratoire Méthodes d'Analyses et de Traitement d'Images pour la Stéréo-restitution (MATIS)