Kadmin, Ahmad Fauzan (2023) Local-based stereo matching algorithm using multi-cost pyramid fusion, hybrid random aggregation and hierarchical cluster-edge refinement. Doctoral thesis, Universiti Teknikal Malaysia Melaka.
![[img]](http://eprints.utem.edu.my/style/images/fileicons/text.png) |
Text (Full Text)
Local-based stereo matching algorithm using multi-cost pyramid fusion, hybrid random aggregation and hierarchical cluster-edge refinement.pdf - Submitted Version Available under License Creative Commons Attribution. Download (48MB) |
Abstract
The estimation of Stereo Matching Algorithm (SMA) is one of the extensive research topics for obtaining the disparity map from two images. The depth measurement provided by the stereo matching framework is used to rebuild the three-dimensional coordinates of point and obiect detection. Similar to how human eyes and binocular vision perceive depth, the visual disparity information obtained from this pair of images captured by the cameras represents the impression of perceived depth. The stereo vision algorithm computes disparity using local, global, and semiglobal optimisation methods established by the researcher. However, the computation needed for the creation of SMA is more difficult, particularly for images comprising complex scenes. The influencing factors include low-texture regions, repetitive patterns, illumination variation, depth discontinuity, and occlusion. Several issues have been challenges to researchers, especially for local methods, such as producing an accurate correspondence between pixels that lie around the boundaries due to different illumination conditions. Besides that, window-based approaches and pixel-based intensity comparison between central pixels and neighbour pixels may cause problems at incorrect disparities, while similar matching costs at low textures cannot be efficiently solved with increasing window aggregation size or implementation of global optimisation. Therefore, this thesis proposes a local-based SMA that enhances the accuracy of complex regions detection by focusing on these issues. The four stages of the proposed SMA were centred on the matching cost computation. The first stage comprised of Tr uncated Absolute Differences (TAD), Gradient Magnitude CLAHE (GMC), and Modified Census Edge (MCE), which were then combined through Planar Pyramid Fusion (PPF) to obtain the initial cost volume. Then, a new proposed cost aggregation based on the Hybrid Random Aggregation (HA) was implemented that utilized modified Iterative Non-Local Guided Filter (iNLGF), Simple Linear Iterative Clustering (SLIC), Graph Segmentation (GS) and Extended Restart Random Walk (eRWR) for error reduction. Next, a Winner-Take-All (WTA) approach was used to select the location of minimum aggregated value corresponding to the disparity value for each pixel. During the refinement stage, the Left-Right (LR) consistency checking process and the Confidence Disparity Filling (CDF) were conducted. Then, the K-means clustering, and Side Window Filter (SW were used to recover the low texture and to remove the remaining noises. In this thesis, the accuracy of the proposed algorithm was evaluated using two standard online benchmarking database systems. For the quantitative and qualitative assessments, systems from the Middlebury Stereo, the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI), and actual images from UTeMLab-Stereo were applied. Once accurate results were achieved, the proposed SMA's disparity maps were generated to be used for the 3D surface reconstruction. As a result, the proposed SMA was able to deliver accurate validation process with findings from the Middlebury system showing 5.11% nonce error and 9.02% all error and the KITTI system showing 7.90% nonce error and 7.07% all error. Therefore, the proposed framework is proven to be competitive with other established methods and can be used as a complete algorithm.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Uncontrolled Keywords: | Computer vision, Algorithms, Computer vision, Computer programs, Image processing, Digital techniques |
| Subjects: | T Technology > T Technology (General) T Technology > TA Engineering (General). Civil engineering (General) |
| Divisions: | Library > Tesis > FKEKK |
| Depositing User: | Muhamad Hafeez Zainudin |
| Date Deposited: | 25 Feb 2026 08:11 |
| Last Modified: | 25 Feb 2026 08:11 |
| URI: | http://eprints.utem.edu.my/id/eprint/27194 |
| Statistic Details: | View Download Statistic |
Actions (login required)
 |
View Item |