Abstract:
Reliable extraction of man-made structures from 3D terrestrial laser scanner point clouds and color images is crucial to the problem of scene modeling and robot navigation in urban enviroments.
However, a great deal of variability exhibited in the large datasets makes it a challenging problem. 3D point clouds density varies drastically over the scanning volume, and object views in 2D color images are unconstrained. Furthermore, a large amount of clutter which does not follow any simple predictive noise model is always present in 3D and 2D data. Due to the large amounts data, some of the structures of interest may contain only a small percentage of the overall data set.
Our approach starts from extracting multiple coherent man-made structures from 3D point clouds and regions from the corresponding color images, and labels them based on prior models of the objects expected.
A local structure based minimum volume ellipsoid (MVE) estimator is used to compensate for the non-uniform and directional nature of point clouds data sampled in outdoor environments. For color images, a causal multiscale random field (MSRF) is used as a prior model on the class labels to capture spatial dependencies of the labels as well as of the image data.
Ultimately, the research aims to extract man-made structures robustly by utilizing combined features in 3D point clouds and 2D color images.
About the speaker:
Hang Zhou has a BE(5 years) in Electronics Engineering from Tsinghua University, China. After spending 7 years commercializing image compression technology in the video surveillance industry, she came to Monash University and started working on a MEngSc by Research degree at Intelligent Robotics Research Centre (IRRC) and Institute for Vision System Engineering (IVSE) from March 2005. Her supervisors are A/Prof David Suter and Prof Ray Jarvis. Her research interest includes: 2D & 3D image segmentation, application of image analysis and machine vision.
.
.
