Method Of Acquiring Perspective Views Based On Multi-viewpoint Image

Due to its flexible designing and low cost, the multi-viewpoint imaging (MVI) technique is widely used in situations where specific field of view or image resolution is needed, such as video conference, panoramic monitoring, robot navigation etc. Multi-viewpoint images are often used to generate perspective views of the scene they represent. However in general, perspective views cannot be computed from a single MVI unless scene structures are known.Most current approaches assume the scene as being at infinity and only use the light-ray directions to compute perspective views. This kind of map is referred to as infinity-map. The infinity-map algorithm however, results in severe geometric distortions when the viewpoints of the system are scattered.With regard to this problem, this thesis provides a new way to extract depth information from structured features in an MVI, and then generate perspective views with the acquired image depth. The result of this approach is better than the infinity-map. This thesis mainly derives the method of acquiring scene depth of structured features from a single MVI, and provides specially designed mathematical approaches to compute depth of common scene curves such as lines and circles. Meanwhile, it analyzes the relationship between the accuracy of depth calculation and the geometry of the imaging system.At last, this thesis verifies the results of two perspective view generating approaches by simulating and real experiments respectively, comparing with the infinity map result and the ground truth. The algorithm of this thesis shows an obvious superiority.