Publication: Ranking - gradient – similarity multi - block matching algorithm for stereo vision
Loading...
Date
2020-10-01
Authors
Kok, Kai Yit
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Various research has been conducted on the development of stereo vision whereby the local approach, the global approach, and the convolutional neural network (CNN) approach have different pros and cons. Nevertheless, the local approach still outperforms the others in terms of simplicity and computation cost, generally. Therefore, a new local block matching algorithm known as Ranking-Gradient-Similarity Multi-Block (RGSMB) has been proposed in this research work. This algorithm developed to resolve the accuracy issue of the local approach. A new cost computation has been introduced in the proposed algorithm, which fully utilizes the information from the limited local window region by combining the cost from three different constraints, including a Ranking constraint, a Gradient constraint, and a Similarity constraint. The performance evaluation of the RGSMB was done using the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) dataset. A total number of 200 image pairs are used in the parameter optimization analysis to find the optimum parameter settings of the RGSMB for stereo vision. Performance comparison with other local approaches carried out has substantiated the superiority of the RGSMB. The proposed algorithm achieved the lowest average errors with moderate computational cost in comparison with the tested algorithms. It is also compared with recent CNN approaches using both KITTI 2012 and 2015 datasets to justify the capability of the RGSMB further. In this comparison, the RGSMB can outperform the CNN approaches with fewer disparity errors. This comparison indicates that the RGSMB exhibits excellent potential for stereo vision. Next, RGSMB is tested experimentally by implementing the algorithm on the StereoPi system. With appropriate strategies of increasing computational efficiency, the RGSMB can achieve near to 2.5 frames per second using a low-cost hardware system. Hence, a new local approach with satisfying accuracy and performing in real-time is developed in this research.