View-Registration for 3-D Model Generation from Sensor Data


Martial Hebert

Robotics Institute

Carnegie Mellon University





Generation of accurate models from sensor data is critical for constructing virtual environments. In this talk, I will explore several aspects of the problem, emphasizing automatic construction of models, i.e., with little or no human intervention. I will illustrate the discussion with examples ranging from small-scale objects, such as 3-D models of desktop objects, to larger models such as room-size maps, to models of extended areas, such as terrain maps. Applications include the creation of virtual environment for building mapping, mine inspection, and terrain map simulation.


All the examples share in common the same underlying technology for automatic surface registration. I will discuss current status of our approach to this problem. In particular, I will discuss the issues of globally optimal and automatic registration and fusion of large number of views, the manipulation of very large data sets, and the issues associated with manipulating non-smooth surfaces. All of these issues are of paramount importance from a practical standpoint but also lead to interesting scientific problems.



Brief Biography:


Prof. Hebert is professor at the Robotics Institute, Carnegie Mellon University. His interests include the development of techniques for representing, comparing, and matching 3-D data from sensors, with application to object recognition and model building. His group has developed techniques for fully automatic modeling of 3-D scenes from many views and for recognizing objects in complex scenes. His interests include also the development of perception techniques for autonomous mobility for mobile robots navigation and techniques for object recognition in images.