Stereo Vision and 3D Reconstruction on a Distributed Memory System

An important research topic in image processing is stereo vision. The objective is to compute a 3-dimensional representation of some scenery from two 2-dimensional digital images. Constructing a 3-dimensional representation involves finding pairs of pixels from the two images which correspond to the same point in space. Several stereo matching algorithms are developed to find matching pairs. Hierarchical matching can be applied to improve the effectiveness and reliability of the matching algorithms. Hierarchy is established by multi-resolution images, also known as pyramids. Pixel based matching algorithms require lots of processing power and memory space. However, stereo vision applications are often subject to strict real-time requirements. In order to meet timing constraints, a multiprocessor implementation of the hierarchical matching algorithm is realized. The parallel algorithm is implemented on a distributed memory system on top of a communication library. The resulting system is integrated in a prototype apparatus for surface measurements during the construction of tunnels.