Abstract:Optical flow estimation is a core task in the field of computer vision. In recent years, deep learning based optical flow estimation methods have made significant progress. However, optical flow algorithms are easily affected by image occlusion, overlapping objects and similar objects, leading to a decrease in the accuracy of motion estimation. To overcome this problem, an optical flow estimation model combining convolutional and cross transformer network was proposed. Firstly, feature extraction and enhancement were performed on two consecutive frames of images using convolution and cross attention mechanisms. Secondly, combining the contextual information provided by the context feature network encoding, the feature aggregation module was used to generate aggregated features, and the ConvGRU module was used to iteratively update the optical flow. The optical flow feature estimation maps were generated through the upsampling module to complete the optical flow estimation task. Finally, a comprehensive experimental comparative analysis was conducted on the MPI Sintel and KITTI 2015 datasets. The experimental results show that the proposed method could effectively improve the accuracy of optical flow estimation and had good generalization ability.