Extracting 3 D Free - Form Surface Boundaries of Man - Made Objects from Multiple Calibrated Images : a Robust , Accurate and High Resolving Power Edgel Matching and Chaining Approach

This paper describes a purely geometrical approach for matching and chaining edgels, i.e. contours points with their tangent direction, to reconstruct in a general way 3D rectilinear or curved surface boundaries from multiple calibrated images. Results show that the 3D reconstructed edgels are highly reliable and accurate, and correctly reproduce the 3D limits of very small image structures thus demonstrating the high resolving power of this technique and enabling a finer description and understanding of urban scenes. 3D Edgels are a key feature for reconstructing man-made objects. Practical examples of edgels reconstruction are given on various type of 3D objects (roundabouts, zebra crossings, cars, classical and odd shape buildings, dormer windows) in surveys at a ground pixel size of 25 cm. Practical examples of edgels integration within a surface reconstruction or 3D pattern recognition and object extraction processes are also provided. Indeed 3D edgels can be used as characteristic and constraining lines for the surface topography reconstruction and used to build the limits of cartographic objects thus ensuring a natural geometrical and topological coherence between Digital Surface Models and Digital Object Models.