Concurrent Stereo Reconstruction

This thesis presents a probabilistic approach to multi view stereo reconstruction from calibrated images. Together with a voxel based world model a Markov Random Field is used to describe the problem of stereo reconstruction as a structural recognition task. A specific scene representation has been developed for stereo reconstruction with the ability to keep ambiguities within the reconstruction result for further evaluation. Based on several implementation specific methods being used in combination with Gibbs sampling an efficiently computable and flexible framework for stereo reconstruction is presented to solve the Bayesian estimation task. Different a priori and observation models are investigated to adapt the general model to specific reconstruction tasks and unsupervised learning methods have been used to reduce the number of model parameters. Evaluation on sufficient test data proves the viability of the proposed method