Research Of Stereo Matching For Autonomous Driving Scenes

Binocular disparity estimation(or Stereo Matching)is the task to obtain dense depth information for that view.Binocular disparity estimation techniques are crucial for cars to understand their position and sense surrounding hazards in autonomous driving scenarios.Although with the development of deep learning,network learning-based binocular disparity estimation compensates for the shortcomings of traditional methods in feature extraction and cost aggregation.However,these methods are still limited by the lack or error of discriminative features in large texture-free,occluded,and glare areas,and suffer from the problem of less accurate matching and estimation.To solve the above problems,this paper finds that depth interval classification method based on a single view is batter in texture,occlusion,and glare areas.Accordingly,this paper proposes a stereo matching network based on the depth interval classification method to compensate for the degraded accuracy of the stereo matching algorithm in areas with missing matching features.Firstly,this paper proposes a disparity cost-volume resolution recovery module using planar normal vector information to obtain a high-precision disparity map.Secondly,in order to make the boundary of the region with missing matching features correctly predicted,this paper designs a localization module to obtain soft labels for the localization of feature missing areas.Thirdly,this paper also proposes a post-processing refinement module which selectively adjusts the cost distribution to preserve the cost peak of fine structure while fusing the classification results of depth intervals.Notably,experiments demonstrate that the algorithm in this paper is still effective in the limit case of single camera defacement failure.Multiple sets of comparison experiments on standard datasets such as KITTI and Scene Flow illustrate that the method proposed in this paper effectively combines the advantages of depth interval classification methods,solves the mis-matching problem of binocular depth estimation algorithms in texture-free,occluded,and glare-free areas,and brings significant accuracy improvement on standard datasets.