Real-time stereo matching via multi-volume attention concatenation volume

Stereo matching accuracy and real-time performance are vital for autonomous driving and robot navigation. We introduce muti-volume attention concatenation volume (MVACV), a novel cost volume creation method, enhancing real-time and high-accuracy stereo matching. MVACV employs lightweight L1 and group-wise correlation volumes to filter concatenation volume, improving feature representation by eliminating redundancy and emphasizing vital details. Additionally, our Feature Aggregation Module (FAM) aggregates parallel multiscale features, boosting disparity estimation in texture-poor areas. Both modules are lightweight, easily integrated into networks like PSMNet and GwcNet, enhancing efficiency and accuracy. This leads to MVACVNet, an end-to-end, real-time stereo matching network. Experiments on Scene Flow, KITTI 2012, KITTI 2015, and Middlebury 2014 datasets prove our method’s real-time capability, high accuracy, and superiority over leading approaches, showcasing strong generalization.