Disparity map algorithm using hierarchical of bitwise pixel differences and segment-tree from stereo image

• Main Author: Zainal Azali, Muhammad Nazmi
• Format: Thesis
• Language: English; English
• Published: 2024
• Online Access: http://eprints.utem.edu.my/id/eprint/28375/; https://plh.utem.edu.my/cgi-bin/koha/opac-detail.pl?biblionumber=124264

In computer vision technology, stereo matching algorithm plays an important role in generating disparity map or depth map through a correspondence process from stereo images. The algorithm development can be categorized into local, global, and semi-global methods. Global method produces high computational complexity and slow implementation, deferring its suitability for real-time application. Local methods excel in solving matching problems through local-based analysis with fast execution and low computational demands. Combining attributes from both, the semi-global method introduces more complex structure and high computational complexity. This thesis presents a local-based stereo matching algorithm to increase the accuracy on complex regions. These regions are low texture, repetitive patterns, illumination differences, discontinuity, and occlusion. The proposed algorithm has four stages that start with a novel bitwise pixel-based differences at matching cost computation. This stage utilizes XOR gate to produce the initial disparity map. The next stage involves the utilization of Segment Tree (ST) to eliminate the noise at aggregation step. Then, an optimization stage employs Winner-Take-All (WTA) strategy. The final step of the proposed algorithm framework is refinement stage. At this stage, Bilateral filter (BF) and Weighted Median (WM) filter are utilized. These filters not only increase the accuracy but are also capable of preserving the object’s edges. Then, hierarchical Gaussian pyramid is applied at each stage to further enhance the final disparity map. The performance evaluation of the proposed algorithm is conducted using two standard online benchmarking databases, which are the Middlebury Stereo for quantitative metrics and Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) for qualitative assessments. The adaptability of the algorithm is demonstrated through a 3D surface reconstruction using a final disparity map. In conclusion, the proposed algorithm displays significant efficiency, yielding an average non-occlusion error of 5.61% and an all-error rate of 8.92%. Hence, the proposed algorithm is competitive with other existing methods, especially when incorporating the pyramid method over non-pyramid approaches.