Binocular vision stereo matching algorithm based on improved PSMNet

Abstract

The invention discloses a binocular vision stereo matching algorithm based on an improved PSMNet. The method comprises the following steps: performing feature extraction on the left image and the right image by adopting a dimensionality reduction starting module; obtaining feature maps respectively, inputting the obtained feature map into an SPP module; enabling the SPP module to compress the feature map and then carry out up-sampling; synthesizing the feature maps of different levels into a final SPP feature map; combining each parallax value in the left and right images, wherein the featuremap corresponding to each parallax value and the SPP feature map form a four-dimensional matching cost volume; enabling the three-dimensional convolution module to aggregate environment information, obtaining a final disparity map through up-sampling and disparity regression, calculating the possibility of each disparity according to a prediction cost Cd obtained through operation of a normalization exponential function, using constant mapping to perform optimization, and wherein a prediction disparity value is obtained through summation of each disparity value and the corresponding possibility.

Description

technical field [0001] The invention relates to the technical field of visual stereo algorithms, in particular to a binocular visual stereo matching algorithm based on improved PSMNet. Background technique [0002] After years of development, binocular stereo vision has played an important role in 3D reconstruction, industrial measurement, unmanned driving and other fields. Stereo matching is the core content of binocular vision research, and it is also a difficult research point of binocular vision. So far, traditional binocular vision stereo matching algorithms are mainly divided into the following three categories: global matching, local matching and semi-global matching. Global matching generally includes matching cost calculation, disparity calculation, and disparity optimization. Its core is to construct a global energy function and minimize the global energy function to obtain an optimal disparity map. The global matching algorithm gets better results, but generally ...

AI Technical Summary

Problems solved by technology

Although compared with the traditional binocular vision stereo matching algorithm, the CNN-based stereo matching method has a certain improvement in speed and accuracy, but in ill-posed areas (such as occlusion areas, discontinuous disparity areas, weak texture areas, reflective surfaces, etc.) is still not ideal

Images