Binocular deep learning method based on adaptive single-peak stereo matching cost filtering

Abstract

The invention discloses a binocular deep learning method based on adaptive single-peak stereo matching cost filtering. The method is characterized in that single-peak distribution supervision with real parallax as the center is directly applied to matching cost of network prediction, self-adaptive matching cost filtering is achieved. The method comprises the following steps that (1) a data set isconstructed, the data set comprises a left image and a right image, and the left image and the right image serve as a three-dimensional image pair; 2) taking the PSMNet as a stereo matching model basic network, inputting the stereo image pair into the PSMNet stereo matching model basic network, and outputting three matching cost bodies (Cost Volume) aggregated by the stacked hourglass 3D convolutional neural network by the PSMNet stereo matching model basic network; (3) for each matching cost body (Cost Volume), calculating the matching cost of each matching cost body (Cost Volume); the methodcomprises the following steps of: estimating a self-confidence degree graph by using a self-confidence degree evaluation network (Confidence Estimation Network) respectively, and adjusting a real matching cost volume (Group Truth Cost Volume) by using the self-confidence degree graph; the method comprises the following steps of: generating unimodal distribution of a pixel level to serve as a network training mark; the device has the advantages that the defects in the prior art can be overcome, and the structural design is reasonable and novel.

Description

technical field [0001] The invention relates to a binocular deep learning method based on adaptive unimodal stereo matching cost filtering, and belongs to the technical field of binocular stereo matching visual image processing. Background technique [0002] Binocular stereo vision obtains rich three-dimensional data, especially depth information, by imitating the principle of human vision. After years of development, binocular stereo vision has played a huge role in industrial measurement, 3D reconstruction, unmanned driving and other fields. Binocular stereo vision is based on the principle of parallax and uses imaging equipment to obtain two images of the measured object from different positions, and obtains the three-dimensional geometric information of the object by calculating the position deviation between the corresponding points of the image. The binocular stereo matching process generally includes four steps: matching cost calculation, matching cost aggregation, d...

AI Technical Summary

Problems solved by technology

For example, DispNetC proposed to use the correlation layer as an approach to the cost function, and then use the parallax regression loss to constrain the network to learn image feature extraction. Because the correlation layer calculates the matching cost process and loses too much information, the accuracy of the binocular matching result is low; while GCNet is It further releases the flexibility of network learning image features and cost functions, and proposes to connect the left and right image features in the channel dimension, and use a series of three-dimensional convolutional layers to learn matching cost calculations. However, the end-to-end network design uses parallax regression (by soft argmin Function regression) loss supervised network learning, lack of clear constraints on the matching cost calculation process, resulting in image feature extraction and cost calculation functions cannot be effectively learned

Images