Research On Real-time Binocular Stereo Matching Algorithm Based On Attention And Multi-scale Information

Binocular stereo matching technology is a research hotspot in the fields of computer vision,computer graphics and machine learning.It has a wide range of applications in three-dimensional reconstruction,robot navigation,autonomous driving,unmanned driving,VR/AR and other fields.Estimating disparity from a pair of rectified binocular stereo images is an ill-posed problem.In recent years,due to the great success of deep learning-based stereo matching technology in the field of 2D and 3D vision,it has received extensive attention,but it is still affected by slender structures,textureless areas,and complex scene areas,causing the problem of low matching accuracy.In some industrial applications,not only the accuracy of stereo matching,but also the speed of disparity estimation.This paper studies the real-time binocular stereo matching technology based on deep learning methods.The main content of the paper can be summarized as follows:(1)A real-time stereo matching algorithm based on attention mechanism is proposed.First introduced the basic principle of attention mechanism,and analyzed the use process of common attention mechanism.Then the attention module designed in this paper and the disparity optimization module based on low-level feature guidance are proposed.The ablation experiment and the benchmark comprehensive experiment of the overall network structure show that the attention module and the disparity optimization module based on low-level feature guidance can significantly improve the matching accuracy in the slender structures area and the textureless areas,making the model in the Scene Flow data set the EPE(End-Point-Error)is reduced to1.25 pixels,and the speed can reach 71 FPS on the one GTX1080-GPU.(2)A real-time stereo matching algorithm based on multi-scale information is proposed.First,the multi-scale model is introduced,and the advantages of the multi-scale model and the feasibility of the method are analyzed.Then,the feature construction of multi-scale method designed in this paper and the multi-scale disparity optimization module based on RGB image guidance are proposed.The ablation experiment and the benchmark experiment of the overall network structure show that the multi-scale disparity optimization module based on multi-scale information and RGB image guidance proposed in this paper can significantly improve the accuracy of stereo matching in complex scenes,and make the model EPE(End-Point-Error)on the data set it is reduced to 1.02 pixels,and the speed can reach 64 FPS on the one GTX1080-GPU,maintaining a good balance of accuracy and speed.