A NEW POST-PROCESSING METHOD FOR STEREO MATCHING ALGORITHM
Stereo matching is continuing to be a critical and challenging problem due to continuous progression in computer vision including the widely applied in various Artificial Intelligence (AI) applications to echo the human visual system. New developed algorithm approaches with more quality of information extraction and high frame rate of data processing are unique targets to control huge volume data and provide convenient solutions for the massive evolution in this field. Fundamentally, the process of stereo matching involved several stages to implement the disparity map which provides the depth information required in 3D reconstruction. Numerous studies and sophisticated algorithms have been developed in the stereo vision area with different concepts and properties to achieve disparity map implementation. However, their accuracy still low and algorithm structures are complicated which far from current requirements. As a popular breakthrough, post-processing algorithms have proved to achieve outstanding performances in stereo matching in terms of low error rate, less complex algorithm structures, and highly computational in their processing speed. In this paper, we present a novel post-processing framework known as Multistage Hybrid Median Filter (MHMF) with a new segment-based algorithm to surge up the accuracy and maintain low computational complexity. The developed framework consists of two main stages: in Stage 1, the Basic Block Matching (BBM) and Dynamic Programming (DP) are applied to obtain the initial disparity map. While, Stage 2 concerns on segment-based, hybrid median filtering, and merging process for the result of Stage 1 as the main contribution of MHMF. To prove the reliability with current available state-of-the-art algorithms, the experimental results and quantitative measurements on standard indoor and outdoor datasets have been compared. The results demonstrate that the proposed method outperforms many existing methods and works as a powerful approach especially in terms of low accuracy, computational cost, and handling of horizontal stripes.
Stereo vision, stereo matching, Artificial Intelligence (AI),3D vision, post-processing algorithms, disparity depth map.