Research On Binocular Stereo Matching Based On Deep Learning

Binocular stereo vision understands the depth information of the target object by simulating the binocular of human beings,and has a wide range of applications in smart phones,autonomous driving,intelligent robots and other fields.Nowadays,deep learning technology shows strong image understanding capabilities,which can be used to extract robust depth representations directly from stereo image pairs,and its performance far exceeds traditional algorithms based on manual settings.However,in practical applications,for complex texture regions,the stereo matching technology based on deep learning may still have errors in estimation,resulting in incomplete local details and unclear object edges.In view of the above problems,this paper mainly studies the stereo matching algorithm based on deep learning,and the main work includes the following three aspects:First,to solve the problem of lack of feature diversity expression in existing stereo matching algorithms,a dual attention residual stereo matching algorithm is proposed.The dual attention residual module is constructed based on the original residual network,and the channel and spatial attention feature maps are respectively inferred along two different paths.Channel attention uses avg-pooling and max-pooling to integrate interdependent features in all channels of the object.Spatial attention also uses two pooling methods to emphasize location information and provide rich and reliable complementary features.The improved three-dimensional residual cost aggregation module further learns feature correlation from the cost volume through three sets of three-dimensional residual blocks and down-sampling operations,and performs jump connections to capture a wide range of contextual information.Then,in view of the lack of effective coordination in the two modules of feature extraction and cost aggregation in existing stereo matching algorithms,a multi-dimensional attention feature aggregation stereo matching algorithm is proposed.Design a twodimensional attention residual module,by introducing non-dimensionality reduction adaptive two-dimensional channel attention into the original residual network,local crosschannel interaction and extraction of significant information,providing rich and effective features for matching cost calculation.A three-dimensional attention hourglass aggregation module is proposed,which uses a stacked hourglass structure as the backbone to construct a dual pooling three-dimensional attention unit,captures multi-scale geometric context information,further expands the multi-dimensional attention mechanism,adaptively aggregates and recalibrates the cost volumes from different network depths.Finally,considering that the cost volume is the hub of the two modules of feature extraction and cost aggregation,the existing cost volume structure does not make full use of multi-scale information and global context information,and a pyramid cost volume aggregation stereo matching algorithm is proposed.The residual module of improved convolution is proposed,and the depth information is extracted by the deep convolution operation with lower computational complexity,and the potential internal information is fully revealed by using redundant features.Construct four-dimensional spatial pyramid cost volumes with four different scales based on series cost volumes,and use multi-scale cost volume three-dimensional aggregation modules to fuse pyramid cost volumes,reduce the number of three-dimensional hourglass modules,infer spatial information from multiple levels,and capture more robustly structural clues to preserve the complete disparity range.