Study On The Algorithm Of Stereo Image Pair Generation

Binocular stereo imaging is a main branch of computer intelligent vision, which means the observer can perceive three dimension scenes through matching and understanding two captured stereo images pair of the same scene. This technique has an abroad application in the field of virtual reality, multimedia education, digital entertainment, appearance design of industrial products, sculpture and architecture etc. In this dissertation, the theories and techniques related to binocular stereo imaging are firstly studied, and a binocular stereo imaging model based on monitor is proposed. Then the properties of this model as well as the impact of binocular disparity, assemble angle and Panum's fusional area limit to the stereo perception are thoroughly discussed. Then according to the different character of a single planar image and video frames,we work over their stereo conversion algorithm respectively.According to the principle of psychological and physiological depth cues, we convert the single planar image into stereo. Five distributions such as uniform distribution, piecewise uniform distribution, normal distribution, triangular distribution and fitting gray value distribution are constructed firstly, then we discuss the stereo effect when the horizontal movement of sub-images subject to each of them and two quantitative criterions, cross-entropy and root-mean-square error, are used to evaluate the stereo effect; furthermore the impact of number of sub-blocks on stereo effect has been also discussed. The experimental results show that: the stereo effect is preferable when random variables subject to normal distribution and the number of sub-blocks has little influence to the outcome.Known from the stereo conversion theory based on spatio-temporal interpolation, when tracking a moving object each eye sees different portion of the visual field at each instant, but may see the same visual field at different time, therefore aiming at monocular video, we present a method to generate its stereo version by selecting stereo pair among the frames. Primarily the displacement information of the feature points in adjacent frames is acquired via feature track algorithm; then the disparity between frames is estimated, and according this the suitable right-eye view is selected among the sequences. Experimental results shows that, two frames with the binocular disparity between -5 and -9 pixels observers can get suitable stereoscopic perception, and good performance is acquired when the camera or the object in the scene has an approximately horizontal movement.