An improved binocular stereo matching algorithm based on AANet

Stereo matching is an important part of establishing stereo vision. Parallax information obtained by stereo matching directly affects the three-dimensional information of an object. End-to-end stereo matching algorithms can directly derive parallax maps from the designed network. However, at the same time, the structure of the network is complex, and a large number of parameters take up much memory. The network increases the device burden, which increases the time required to obtain the parallax map, lowering the efficiency of the network movement. Thus, an improved stereo matching algorithm based on AANet (adaptive aggregation network for efficient stereo matching) is proposed in this paper: AEDNet (adaptive end-to-end depth network for stereo matching). In the feature extraction module, the network simplifies the network structure by limiting the convolution kernel size to obtain the features with low abstraction. In cost aggregation, the intra-scale aggregation module is used to achieve adaptive cost aggregation through deformable convolution, and the inter-scale aggregation module uses the traditional cross-scale aggregation method to compensate for the missing global information to a certain extent. The network is verified the performance on the KITTI dataset. The results show that the algorithm can still complete stereo matching efficiently and accurately and obtain a better disparity map when the network is simplified. These provide preconditions for accurate three-dimensional reconstruction.