Tree Extraction from Urban Image Sequences

• Applicant: Professor Dr.-Ing. Helmut Mayer
• Subject Area: Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: from 2005 to 2009
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 5456400

The goal of the project is the automatic reconstruction of individual trees with their particular characteristics, i.e., shape and texture, from urban, if possible high resolution terrestrial image sequences, to add a natural touch to three dimensional (3D) city models. The model will be described by a simplified L-system from computer graphics and will be estimated by means of reversible jump Markov chain Monte Carlo (RJMCMC). The jumps will be used to control the branching of the Lsystem into branches, twigs, and leaves or needles starting from the trank. The chosen modeling allows not only to view the tree from every direction, but also to change the lighting, to animate the tree, e.g., by wind, or change the season. For a more flexible acquisition, it will be possible to extract coniferous and foliaged deciduous as well as unfoliaged deciduous trees, the latter also because they provide detail about the branching. We will mainly concentrate an the branching structure, because the foliage can be seen in first approximation as a distribution and computer graphics problem. Trees will be detected based an their vertical trunks or texture of the leaves or twigs, with hypotheses for trunks for the latter generated from local symmetries detected by Hough transformation. For unfoliaged trees, the 3D structure of the branches needed for the L-system will be obtained by matching, if possible based an dense image sequences. The matching is difficult as the sequence of matching branches can be totally different for each image because of the complex 3D structure. We will tackle this problem by knowledge about the branching structure modeled by the L-system and RJMCMC. For coniferous and foliaged deciduous trees dense surface models will be generated by matching, leading together with silhouettes generated by segmentation in the images to the 3D hulls of the trees. From them, possibly based an 3D medial axes, the parameters of the L-systems will be determined via RJMCMC.

• DFG Programme: Research Grants
• International Connection: China, China (Hong Kong)
• Participating Persons: Zhilin Li; Zuxun Zhang; Qiming Zhou