Extracting complete 3-dimensional boundary representation from multiple range images
Zukang Xu · Summit (Simon Fraser University) · 1992
An important problem in computer vision is that of generating a 3-dimensional description of objects from sensed data.In robotics applications such as grasping and planning collision-free motion, symbolic 3-dimensional descriptions of objects must be available.Such 3-dimensional descriptions require explicit geometric and topological relations between faces edges and vertices, and are very similar to boundary representation (b-rep) used in solid modeling.In this thesis, we present an approach to extract a complete b-rep description of a polyhedral object from range images taken from multiple view-points.Our system, starting from basic face models of visible surfaces of objects in each local view, extracts features, matches these features, generates rigid-body transformations that relate the local views, and finally merges these local views into a complete 3dimensional b-rep description of the object.In our matching algorithm, a triple branch structure is defined as the main feature to determine correspondence.A coarse-to-fine searching strategy and a prioritizing procedure are designed to reduce the search time even further.The b-rep model is incrementally updated by merging each of the local views in a global description.A convenient and effective termination criterion is designed to monitor the integration process.We have implemented our system in C, running on a SUNISPARC.A synthetic range image generator has also been implemented.The system has been tested on several synthetic range images.Our system in conjunction with the synthetic range image generator also has applications in the area of robotics, CAD and geometric modeling.My sincere thanks to my senior