A local multi-block with various constraints approach for stereo vision in StereoPi

Abstract

The conventional local block matching algorithms are fast but have an accuracy issue, making them unsuitable to be implemented on real-time applications. This paper studies a local block matching algorithm with different cost computations, combining a cost aggregation with various blocks' shape and sizes, using the computer and StereoPi. Compared to other local block matching algorithms, the Ranking-Gradient-Similarity Multi-Block (RGSMB) algorithm can extract information with a higher cost uniqueness within the predefined local window. Using KITTI 2015 as input data, RGSMB has better accuracy than other local block matching approaches. It is at least 5% better than Census Transform (CT), 23% better than Ranking Transform (RT), and 62% better than Sum of Absolute Differences (SAD) in terms of accuracy. Furthermore, it can achieve lower computational costs on the StereoPi, a low-cost computer vision embedded system. The RGSMB can achieve 2.5 fps with the right image resolution and parameter settings.