Tree Extraction from Urban Image Sequences

• Applicant: Professor Dr.-Ing. Helmut Mayer
• Subject Area: Geophysics
• Term: from 2007 to 2011
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 62030927

The goal of the project is the automatic extraction of individual trees from urban, high resolution terrestrial image sequences and laser-scanner data, The trees will be reconstructed with their particular characteristics, i.e., shape and texture, adding a natural touch to three dimensional (3D) city models and thus improving the realism of 3D geoinformation systems (GIS). Strong object models about vegetation encoded by means of L-systems are employed together with Markov Chain Monte Carlo - MCMC resulting into a statistical generative approach. With it, plausible trees can be reconstructed from only a couple of images in spite of weak contrast, clutter by background objects, and occlusions. The second period aims at more complex, e.g., context-sensitive, L-systems based on a more detailed geometrical modeling with generalized cylinders for the reconstruction of the branching structure of trees. The estimation of empirical distributions for the parameters from a larger number of examples is expected to lead to better results and a more efficient modeling. By integrating reversible jumps - RJ (Green 1995) in conjunction with model selection into MCMC, search is optimized by an informed selection of competing hypotheses while at the same time avoiding overfitting. For coniferous and foliaged deciduous trees dense surface models are to be generated by matching, leading together with silhouettes generated by segmentation of the images to the 3D hulls of the trees. From them, possibly based on 3D medial axes, the parameters of the L-system will be determined again via RJMCMC. Finally, we want to test if our model and search procedure might also be applicable for terrestrial laser-scanner data, when a suitable 3D likelihood function is used.

• DFG Programme: Research Grants