A stereo vision algorithm and its DSP implementation for real-time geometrical reconstruction of indoor architectural scenes

The work described in this thesis is the development of a technique to extract geometrical models of indoor architectural environment by using image sequences made by a video camera undergoing a planar motion in a hallway along its direction. The algorithm is primarily correlation-based. Some conventional image processing and analysis methods, such as the Sobel Operator and the Hough Transform, are used as preliminary processing steps to remove redundant information contained in the images. The major contribution of the work is that during the construction of the scene models, geometric constraints are imposed by the orthogonality of typical architectural environment; that is, the scenes are composed of geometrical objects, such as lines and planes, that are either in parallel or perpendicular to one another.;The main algorithm is first prototyped in the Matlab environment and is implemented on a DSP coprocessor board inserted in a PCI slot in a PC host computer. The second contribution of our work is that the technical details concerning the implementation of the algorithm on the embedded DSP system, the interaction between the DSP system and a host PC, and the host's user interface are described.;The technique developed in this thesis can be applied to various areas such as robot navigation and aid for the visually impaired.