Research On Fast Image Registration And High Accuracy Sub-pixel Stereo Matching

Image-based Modeling and Rendering(IBMR) methods attract more attention inthese years. As IBMR is independent on specific equipment, it is widely used in vari-ous applications, such as robot navigation, object reconstruction, object extraction andtracking. Among which image registration and3D information recovery are specificallyimportant modules. For image registration, local feature based algorithms have beenproved to be more accurate and robust, and become the most popular kind. But thehigh computational cost is a problem, both in detection and description. For3D recovery,benefited from simple modeling and dense information of depth map, stereo matching be-comes one of the most popular topics. Nevertheless, the performance in textureless andoccluded regions is still the critical problem. This work conducts an in-depth research,and proposed a fast and stable image registration method and a high accuracy sub-pixelstereo matching method for above two aspects respectively. The main contributions ofour paper are as follows.1. We proposed a new topology-based descriptor for afne-covariant local featuredetectors, such as MSER, in image registration. Considering that the texture-based de-scriptors usually have high computational cost and are sensitive to the nonlinear illumi-nation transformation, we treat the region-pair as a unit and extract the geometry afneinvariant parameters as the feature descriptor. Due to the independence of image texture,the new descriptor is robust to the illumination change and is much faster than texture-based descriptor. Real-time feature extraction and frame matching on320*240video isachieved.2. We proposed a local-based stereo matching algorithm by Confidence-based Sup-port Window(CSW). In order to make full use of the matching results from windowswith diferent sizes, the proposed method defines CSW to select suitable neighbors, andthen applies local plane fitting to solve the textureless and occlusion problems. The pro-posed method is fast and has good results on smooth surface. Experiments show that thenew method has a better overall performance compared with the state of the art in localmethod.3. We proposed a high accuracy sub-pixel global-based stereo matching algorithm. Firstly a new global energy function is presented based on the local plane smooth itemand ground control points. To optimize the energy function, we introduce an interleavingupdating framework of disparity and confidence maps. In this paper, disparity map isupdated by the adaptive CSW method, while confidence map is updated based on thedisparity volatility in local area. The proposed algorithm is competitive with relatedworks. The top ranks under diferent error threshold in Middlebury platform show thatour algorithm achieves sub-pixel precision. The experiments on diferent datasets provethe efectiveness of our algorithm.4. We proposed the adaptive initialization and multi-step refinement modules instereo matching. Based on confidence maps obtained by multiple cost function, the al-gorithm selected most suitable depth and confidence maps adaptively. Experiments showthat our initialization integrates the advantages of diferent cost functions. In the refine-ment, we proposed a new color-weighted median filter to improve the performance ondiscontinuity regions. By applying multi-step refinement, the discrete noise is suppressedwhile the sub-pixel precision is preserved.