基于自注意力机制和多尺度代价聚合的双目深度估计方法

打开文本图片集

中图分类号：TB9 文献标志码：A文章编号：1674-5124（2025）08-0122-09

Abstract： To address the challnges of depth estimation for slender and low-texture objects in outdoor scenes for unmanned systems， a stereo depth estimation method based on self-atention mechanism and multi-scale cost aggregation is proposed. Firstly， using deformable convolutions and atrous spatial pyramid convolutions enhances the feature extraction capability of the feature extraction module; secondly， adopting multi-scale matching cost calculation balances the global continuity and detailed information in disparity estimation; and then，the matching cost aggregation module incorporates a self-attntion mechanism to address the uneven distribution of cost volume values; subsequently， the final estimated disparity is obtained through disparity regression; Finally，ablation and comparison experiments are used to validate the performance of the depth estimation method.The experimental results indicate that， while achieving the basic real-time requirements of unmanned systems，this method reduces the D1 metric to 1.28% and the EPE metric to 0.614 pixels， effectively enhancing the accuracy of disparity estimation. Furthermore， qualitative evaluations demonstrate that this method achieves good results in depth estimation for slender and low-texture objects. Keywords： depth estimation; convolutional network; cost computation; self-attention

0 引言

近年来，得益于计算机和通讯领域的高速发展以及人工智能在各领域的成功应用，以无人车为代表的自主无人系统得到广泛发展[1-2]。（剩余12281字）