A Stereo Matching and Depth Map Acquisition Algorithm Based on Deep Learning and Improved Winner Takes All-Dynamic Programming

Abstract

The stereo matching technology has been widely used nowadays for robot guidance, depth of field rendering, and video processing applications. The binocular stereo matching technology collects two images of the same scene from different views, computes the disparity, and then determines the 3-D depth of the object using the triangulation theory. In this paper, we propose a new deep-learning-based stereo matching algorithm called convolutional neural network WTADP (CNN-WTADP). The commonly used WTA method ignores the disparity constraint of the neighboring pixels and is prone to be affected by noise, so the obtained original disparity map may contain a lot of abnormal values and mismatched points. To solve this problem, we construct a CNN for matching cost computation and use a novel global method of dynamic programming based on WTA for disparity selection. Steps for the CNN-based computation of the matching cost are presented, and the post-processing algorithm is used to obtain the final determination of the disparity map. The Middlebury stereo dataset and stereo evaluation are used to compare the proposed algorithm with other methods. The experiments and its results demonstrate the ability of the proposed algorithm to match the images more effectively.