A binocular disparity estimation method based on three-dimensional convolution

Abstract

The invention discloses a binocular parallax estimation method based on three-dimensional convolution. Includes the following steps: calibrating the binocular camera by using the binocular fixing method to obtain a binocular correction map; The binocular image to be estimated is corrected to obtain the corrected binocular image; The corrected binocular image is sent into the preset two-dimensionalconvolution neural network to obtain the feature map after feature transformation. The feature map of the current frame and the feature map of the previous multi-frame images are stitched and sent tothe three-dimensional convolution neural network to obtain the feature map of the multi-frame images. The characteristic images of multi-frame images are transposed and convolution, and then transformed back to pixel domain to obtain disparity estimation map. Compared with the prior binocular parallax estimation method based on the convolution neural network, the invention estimates the binocularparallax map of the current frame by extracting the information on the time dimension through the three-dimensional convolution and combining the binocular information of the current frame and the previous multiple frames. Compared with the original method, this method improves the accuracy and the coherence between the two frames.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and specifically relates to a method for processing input binocular camera information and generating a corresponding disparity map. Background technique [0002] Obtaining accurate binocular disparity maps is a prerequisite for depth estimation. Depth estimation is an important research topic in the field of binocular stereo vision. It is used in many fields such as robot navigation, precision industrial measurement, object recognition, virtual reality, scene reconstruction, and surveying. There are applications. Observing an object with the left and right cameras, acquiring the image under the binocular perspective, and obtaining the disparity map according to the pixel matching relationship between the images. The offset between pixels is calculated by the principle of triangulation to obtain the three-dimensional information of the object. After obtaining the depth of field informat...

AI Technical Summary

Problems solved by technology

However, the existing convolutional neural network binocular disparity estimation method needs to be further improved in terms of accuracy and coherence between front and back frames.

Images