Stereopsis is one of the most important research topics in computer vision. Unlike the mainstream stereo matching methods in computer vision field, which employ purely mathematical and engineering approaches, we seek a physiology-motivated approach. The main work in this thesis is to explore new theories and methods to stereo vision based on the physiological data of cortical disparity detection mechanisms, and to apply the new methods to stereo matching on our robot platform. Here are the primary contributions of the work: First, a computational model inspired by vergence eye movement mechanism is proposed to deal with large disparity, which poses a problem for standard disparity energy model. The model employs a hierarchical disparity information loop to simulate vergence eye movement and a Bayesian inference to obtain an optimal solution. The proposed model yields better estimation accuracy and reliability over the existent models based on the disparity energy model. Second, a novel image similarity measurement is introduced by spatially pooling the population response curves from different spatial frequency channels in a coarse-to-fine manner. Matching cost function can be constructed using this measurement and a global cooperative process is engaged to refine the locally estimated disparity map. Third, a computational model for binocular disparity gradient estimation is proposed and numerical simulation results of our model are in agreement with the physiological data of neuronal sensitivity to surface slant and orientation. Fourth, the proposed disparity computation methods in this thesis are applied to the stereo matching in robot navigation. An extended disparity model is proposed to cope with small vertical disparities resulting from severe image distortions.
修改评论