Binocular stereo matching algorithm based on multi-feature aggregation

To address the issue of high mismatching rates in binocular stereo matching of objects with weak texture and repeated texture, a multi-feature aggregation-based binocular disparity estimation algorithm is proposed, taking into account the significant influence of cost aggregation on disparity prediction in stereo matching. The feature extraction network combines the left and right image features to create a cost volume, extracts effective multi-scale features using spatial pyramid pooling, and employs dilated convolution to expand the receptive field. In cost aggregation network part introduces 3d attention mechanism module, the module will be the channel distribution characteristics of different weights, after repeated convolution of the advanced features of optimization, improvement of 3d context information generated module, provide identification characteristics, and design characteristics of polymerization more modules, the fusion of high-and low-level features and context information, ultimately generated by means of regression parallax figure. The end-point-error is 1.062 on the SceneFlow dataset, and the three-pixel mismatching rate is 1.79% and 2.21% on the KITTI2012 and KITTI2015 datasets, respectively. The results show that the proposed algorithm effectively reduces the disparity error of prediction.