This project deals with the problem of automatically generating a qualitatively geometric scene understanding using planar surfaces in multiple views. Geometric scene understanding are useful for, among other things scene analysis's and autonomous robots. Much research have been done on the multiple view qualitative geometric scene understanding methods and the single image qualitative methods. This thesis focus on the generating a qualitative geometric scene understanding using multiple frames but without doing reconstruction.
The system consists of four module which are evaluated quantitatively at the end of each module. "Pre-processing", which detect blur and noise in
each frames and discard bad frames. "Feature extraction", which extract feature point correspondences between two frames using SIFT and KLT. The extracted feature points are improved with respect to the epipolar geometry to archive correct and accurate correspondences. "Layer extraction", which extract the planar surfaces in two views using an extended RANSAC. The initial layer segmentation are improved using graph cut. "Geometric scene understanding", which estimates layer connections, relative depths, and orientations. These features are then combined using simple to generate a qualitative geometric scene understanding. The entire system is evaluated qualitatively on two videos. It is concluded that accurate layer segmentation is vital for the system performance. Provided with accurate layer segmentation is the system able to generate a good and precise geometric scene understanding.
This project makes the following contribution, an extension to RANSAC and a simple and intuitive reasoning algorithm for generating a qualitative geometric scene understanding using homographies and layer segmentations.