Abstract: Aiming at the problem of mismatched disparity estimates around object edges in binocular depth estimation algorithms for agricultural machinery's autonomous driving systems, a binocular depth estimation network based on edge detection and multi‑scale cost volume was proposed. In the feature extraction stage, the edge branch network and disparity estimation branch network were designed. The edge branch network extracted image features and embedded learned edge geometric features into the disparity estimation branch network to enhance edge‑aware disparity estimation. During the cost volume construction stage, a multi‑scale cost volume was designed. By imposing mutual constraints between different cost volumes, the approach improved the correlation between matching pixels and candidate pixels. Additionally, the merging of multiple cost volumes captured richer global context information, thereby enhancing regularization performance. The proposed method was evaluated on standard stereo datasets, including Scene Flow, KITTI 2015 and Middlebury v.3. The experimental results showed that the disparity accuracy of the proposed network in Scene Flow, KITTI 2015 and Middlebury v.3 datasets is improved by 45.8%, 14.8% and 41.2%, respectively, compared to the EdgeStereo algorithm. These results highlight the network's effectiveness and provide a valuable reference for advancing autonomous driving technologiesyin agricultural machinery.

Key words: agricultural automatic driving, stereo matching, edge detection, multi?scale cost volume, disparity refinement

摘要： 针对农机自动驾驶环境感知在行驶边缘病态区域存在误匹配的问题，提出一种基于边缘检测和多尺度代价体的立体匹配网络。首先，在特征提取阶段设计边缘分支网络和视差分支网络，利用边缘分支网络有效提取细小物体的几何轮廓信息，并将轮廓作为结构信息嵌入到视差分支网络中；其次，在构建匹配代价阶段设计一种多尺度代价体，不同代价体之间相互约束能够提高匹配像素与候选像素的相关性，同时融合多个代价体能够捕获更多的全局上下文信息进行正则化；最后，在Scene Flow、KITTI 2015以及Middlebury v.3立体数据集进行试验。结果表明，与EdgeStereo算法相比，提出的网络模型在Scene Flow、KITTI 2015以及Middlebury v.3数据集的视差精度分别提高45.8%、14.8%和41.2%，为农业自动驾驶环境感知提供技术参考。

关键词: 农业自动驾驶, 立体匹配, 边缘检测, 多尺度代价体, 视差优化