Unsupervised Depth Prediction Based On Convolutional Neural Network And Binocular Parallax

As one of the important directions in the field of computer vision,depth prediction enables the computer to estimate the depth information of the scene through twodimensional images.At the same time,the convolution neural network has a good effect on the depth prediction of the scene by virtue of its powerful image feature extraction and function fitting ability.However,many depth prediction methods based on convolution neural networks need a lot of real depth information as training data,and the acquisition of real depth information is affected by equipment and environment,which requires a lot of manpower and resources.To solve the above problems,this paper presents a unsupervised depth prediction method based on convolution neural network and binocular parallax.The specific work of this article is as follows:Firstly,aiming at the problem that the depth prediction method needs a lot of real depth data,this paper puts forward the unsupervised depth prediction method,uses the convolution neural network to fit a nonlinear function to estimate the scene depth information,and then unifies the parallax between the left and right images to realize the unsupervised prediction scene depth information.Because the disparity between the left and right images is not only non-linear relationship with the depth information,but also is closely related to the position transformation between the left and right cameras,so this method can further estimate the relative motion between the left and right cameras.In this paper,the unsupervised depth prediction method,which takes only the left and right images as input without any real depth information,can be used to estimate the depth information of the scene and the rotating translational relation between the camera and the left side.Secondly,in order to improve the ability of predicting depth information of this neural network,this paper proposes an improved network structure based on the network structure of other depth prediction methods.The network structure is as follows: 1 in order to simulate the human binocular acquisition scene information,the neural network takes about two images as input and 2 to merge the feature map acquired by the left and right image convolution operation,then carries on the multiple convolution operation to extract the image feature.Finally,the continuous deconvolution operation is performed to gradually recover the depth information.This paper uses this network structure to simulate the perceptual mechanism of the depth information of human reasoning scene.Compared with the results of other depth prediction methods,the method of unsupervised depth prediction has better performance in prediction precision with the same evaluation standard,and the depth prediction method proposed in this paper can further estimate the position transformation between left and right camera.