The Research On A Two-phase Stereo Matching Algorithm

Stereo matching has an important role in vision based navigation and 3D reconstruction for the robots. Based on the applications of the mobile robots, this thesis addresses four problems for image matching i.e. an image noise filter, feature extraction, the computing optimization of the similarity function, feature and area based matching.(1) To reduce the noise which will reduce the accuracy of the stereo matching, an adaptive median filter based on noise detected is proposed. An impulse noise filter for 1D signal has been introduced and then extended for the image proceeding. In this algorithm, the position and magnitude of the noise are detected to compute the size of the template of the median filter. The simulation results show the proposed method can effectively filter the impulse noise and deduce the negative impact on image smoothing by the median filter.(2) For the feature extraction, an improved Barnard-LoG feature extraction algorithm is proposed. This algorithm combines the advantages of Barnard, which has good tolerance for the intensity of the illumination and contrast, and the advantages of LoG, which has good tolerance for the noise. The comparable results of the Barnard-LoG algorithm show that it has good tolerance of the illumination and contrast. Based on this feature extraction, a stereo matching algorithm is implemented to support its efficacy.(3) To reduce the computing complexity of the cross correlation in the area based stereo matching, an increment based optimized algorithm for SSD and cross correlation algorithm is proposed. In the cross correlation computing, the mean value and standard deviation are applied by the incremental computing separately to achieve the optimization of the whole cross correlation computing. Considered incremental relationship of the SSD when applied on the consecutive pixels in the image, an optimized computing algorithm is proposed.