Research On End-to-end Binocular Parallax Estimation In Autopilot Scene

Binocular disparity estimation(also known as stereo matching)is a technique to estimate the depth information in a certain visual angle,which first estimates the parallax between the left and right views captured by two cameras,and then uses the camera builtin parameters to convert the parallax map into a depth map.In recent years,autopilot technology has attracted much attention,and obtaining depth information in autopilot scene is very important to perceive the surrounding danger.with the continuous development of deep learning,the research of binocular parallax estimation has made a breakthrough.Therefore,the research of binocular parallax estimation in autopilot scene has high research value.Although the research of binocular disparity estimation based on deep learning improves the problems caused by manual intervention to some extent,these methods in some sick areas,such as large area without texture,glare and occlusion,it will still lead to the inaccuracy of feature matching and disparity estimation due to the lack or error of key feature information.In view of the above problems,the research content of this thesis mainly includes the following aspects:(1)Select PSMNet for in-depth research,analyze the network structure of the model,in order to further reduce the mismatching rate of the model in the ill-conditioned region,in the feature extraction stage,replace the ordinary convolution kernel with a multibranch block based on structural re-parameterization to improve the feature extraction ability;introduce a normalization-based attention mechanism into the residual network to form a channel residual attention module to further restrain the insignificant channels.In order to enlarge the receptive field without increasing the number of parameters,hole convolution is added to the pyramid pooling module.(2)In order to solve the problem of large amount of computation and slow reasoning speed caused by 3D convolution,a binocular parallax estimation model based on adaptive cost aggregation is proposed,which uses adaptive aggregation to replace cascaded 3D convolution cost aggregation.The original top-down/bottom-up hourglass 3D convolution network is replaced by an adaptive cost aggregation module to speed up model reasoning.(3)In order to verify the effectiveness of the improved algorithm,the open datasets Scene Flow and KITTI 2015 are used for experimental verification,and the improved model is analyzed subjectively and objectively.The experimental results show that the reasoning time of the improved model is greatly reduced,the mismatching points in illconditioned areas such as occlusion and weak texture are reduced,and better performance is achieved in both datasets,which are better than the improved model.The improved model proposed in this thesis is compared with other stereo matching network models,such as AANet,PSMNet,etc.The experimental results show that the improved network model also has certain advantages.