An Improved Stereo Matching Algorithm Based on AnyNet

In many applications of depth estimation, accurate disparity map needs to be generated quickly. In order to obtain accurate disparity map, the current mainstream algorithm mostly adopts deep complex network architecture, which requires a large amount of computation and is difficult to be applied in real-time scenes. However, some real-time networks have low disparity accuracy, which also limits their application scenarios. Based on the above shortcomings, this paper improves AnyNet stereo matching algorithm and proposes a stereo matching algorithm with high real-time performance and high accuracy. First, a multiscale feature extraction module is designed to capture and fuse contextual feature information, and then an attention module is constructed to reduce the mismatch problem of ill-posed regions (repetitive, no/weak texture regions). The algorithm proposed in this paper can predict the disparity map in multiple stages during reasoning, weigh the amount of calculation and accuracy according to actual needs, and select the corresponding stage adaptively. Evaluated on the KITTI 2015 dataset, compared with the reference algorithm AnyNet, the final predicted disparity map error rate is reduced by 2.43%, and the running speed is only 1.24% slower.