Research And Implementation On Deep Learning Based Stereo Depth Measurement Method

With the rapid development of artificial intelligence technology,many applications have put forward higher and higher requirements for understanding the 3D environment.Augmented reality technology,autonomous driving technology,3D reconstruction and positioning technology,unmanned aerial vehicle technology,simultaneous localization and mapping technology and other application scenarios have high requirements for depth measurement technology.As a non-contact measurement solution,binocular depth measurement technology has been widely studied and applied due to its low cost,easy deployment and good reliability.In this thesis,deep learning technology is applied to the field of binocular vision depth measurement,and the practicality of the algorithm is gradually improved by continuously combining the geometric characteristics of the binocular vision depth measurement task itself.The main contents of this thesis are as follows:1.This thesis studies a method of fully convolutional network disparity estimation method combined with edge information.Based on the analysis of the input and output forms of the binocular disparity estimation problem,this thesis models the disparity estimation problem as image generation problem.On this basis,this thesis designs a fully convolutional network architecture to estimate the disparity map,with the characteristics of the disparity map,and combination of edge and visual information,and constructs an edge-aware disparity smoothing loss function to improve the performance of the algorithm.2.This thesis studies a pseudo-3D convolutional network disparity estimation method that combines multi-layer geometric information.In this thesis,the geometrical characteristics of the binocular vision problem are introduced into the neural network design.By combining the multi-layer information of input image and construct 4D cost volume through shfted concatenation,the neural network can explicitly use the geometric characteristics of disparity.After obtaining the 4D cost volume,the 4D cost volume is further processed by pseudo-3D convolution and transposed 3D convolution,and finally the disparity map is obtained by soft argmin function.The disparity map obtained by this method has good accuracy and generalization ability.3.This thesis studies an unsupervised adaptive disparity estimation method based on left and right image reconstruction.Through the in-depth analysis of the relationship between the left and right images and the disparity map in the binocular vision problem,this thesis uses the left and right input images and disparity maps to reconstruct the left and right images,and calculates the image reconstruction loss as the back propagation signal for neural network training.Finally,the disparity estimation method using unsupervised learning mechanism is realized.The method has relatively good accuracy,and has online adaptive update capability,which is practical in applications.