Generation of Stereoscopic Video from Monocular Image Sequences Based on Epipolar Geometry

Stereo Visualization is visually perceiving the properties of a three-dimensional object using two flat images, superimposed using appropriate techniques such that each eye would perceive a different image when viewed through a 3D glass. It started in 1838 when Sir Charles Wheatstone narrated the "Phenomenon of Binocular Vision". Stereo Vision supports the fact that Human Eye perceives two slightly different views of the same object to perceive depth in real life. The biggest disadvantage of using a conventional stereo capture method is the requirement of bulky cameras with double the cost of production. Using Computer Vision techniques with videos captured from a single camera to generate stereo pairs not only reduces the cost of video capture, but also gives the freedom of increasing or decreasing the convergence between the two scenes to decrease or increase depth perception. This thesis introduces an automatic approach to generate stereo pairs from monocular video cues using Epipolar Geometry. The proposed method uses the input video frames to select best match stereo pairs. Since this algorithm does not depend on rendering images using depth maps, it is more robust and generates high quality images. Matched feature coordinates between two frame sequences generate the Fundamental Matrix relation between them. We considered zero convergence angle and unit translation between the frames for parameterization of Fundamental Matrix for pure translation as the basis for predicting the pairs. Camera parameters can be estimated using the projection matrix obtained from the scene geometry. The predicted scenes are tested for their Peak Signal to Noise Ratio and Computation time, apart from their visual plausibility and depth maps.