Skip to Main Content
Stereo vision is the process of recovering 3D spatial information from a pair of 2D images. It is very difficult problem due to the fact that stereo matching problem tends to produce as large number of plausible solutions. Thus we need to restrict the solution space in some manner. Trellis-based stereo matching algorithm places hard constraints on the solution by considering the geometry of stereo imaging and by making assumptions about the real depth shape. In that algorithm, permitting the use of the highly parallel Viterbi algorithm (a special case of dynamic programming), real-time stereo vision system can be realized. However, by using a single scan-line for matching, there are too many stripe noises in the result. In this thesis, a new method which can reduce the stripe noise is proposed. Using the disparity information of the upper scan-line, the proposed algorithm achieves more exact disparity map. The proposed algorithm uses Average of Squared Difference (ASD) instead of Euclidean distance in calculating matching cost in order to alleviate the horizontal stripe noise. The experimental results show more precise disparity map compared with the existing algorithm, especially, at low-to-moderate SNR range. Random dot stereogram (RDS) is used to evaluate the quantitative experiment.