基于深度知觉和表面完全化的立体视觉研究

Stereo vision is a primary means for recovering depth information from two images taken from different viewpoints. A large number of methods have been developed to solve the so-called stereo matching problem, a key step in stereo vision. According to the nature of the measurement primitives, most of them fall into one of two categories: feature-based or region-based. The feature-based methods extract image features, such as edges or line segments and match them across stereo images. Owing to fewer measurement primitives to deal with, the methods are usually more efficient than the region-based ones. However, a surface interpolation step has to be included to obtain a dense disparity map. Therefore, a common problem in feature-based methods is that if the features are sparse, then the surface interpolation step will be difficult. The region-based methods use the intensity patterns as the measurement primitives. A correlation or sum of squared differences is computed to match the intensity patterns across stereo images. Since the intensity values are used, the region-based methods can obtain a dense disparity map. However, most of them give poor results in occlusion regions due to the discontinuities in depth. Therefore, how to segment the image into appropriate regions is a key problem in the region-based cases. In order to overcome the difficulties in obtaining a dense disparity map from sparse data, such as a result from stereopsis, Some researchers attempted to use a regularization theory to solve the problem. A common drawback of the methods is that if the images are sparsely textured in the regions behind occluders, then the methods tend to give false results in such cases. The results are not consistent with what humans perceive. Why do the methods have the drawback? The reason is that the information available for reconstruction purpose has not been used globally in those cases. As a result, locally optimal solutions arise. .In this research project, we have proposed a new multi-layered model for stereo vision. The model is based on the psychophysical observations. We have given a relation of interaction between depth and surface perception and have showed that the problems of both depth determination and image segmentation can be solved simultaneously by using the proposed multi-layer representation model and the interaction mechanism of depth and surface perception.

本项目将提取视差信息和生成景物描述作为一个整体来进行研究。通过引入深度知觉和表面完全化之间的相互作用机制来融合灰度和双眼视差信息以得到与观测数据不矛盾、并能正确反映景物深度分布的解答，进而生成景物基于深度分布的分层表象模型。所提出的理论将为用工学方法实现与人类相近的双眼立体视觉机能提供理论和技术上的框架。