Hybrid image segmentation using lidar 3D planar primitives

Airborne laser systems are nowadays well-known to provide regular and accurate altimetric data. The aim of this paper is to investigate the potential of using extracted features from a Lidar 3D point cloud for data fusion purpose through an hybrid image segmentation algorithm. The general context of this study is the building reconstruction. We first describe an efficient algorithm for extracting 3D planar primitives from a laser survey over urban areas. It is based on a normal driven random sample consensus (ND-RANSAC) which consists of randomly selecting sets of three points within laser points sharing the same orientation of normal vectors. A robust plane is then estimated with laser points that are likely to belong to the real roof facet. The number of random draws is managed automatically with a statistical analysis of the distribution of normal vectors within an approximation of the Gaussian sphere of the scene. These 3D facets are then introduced into an image segmentation algorithm based on a bottom-up region merging scheme. Initial regions are computed by a watershed transform onto the gradient of the aerial image. Regions adjacency are represented with a Region Adjacency Graph (RAG). Series of successive merging generates a hierarchy of RAGs where edges are mutual inclusion relationships. The process terminated when the entire image is represented as a single region. A cut in the hierarchy provides a desired image partition. Results of the facet extractions as well as of the segmentation are shown and discussed.