基于多维注意力的立体匹配网络  

Stereo matching network based on multi-dimensional attention

作　　者：孙国栋[1] 张航 李超 杨雄 SUN Guodong;ZHANG Hang;LI Chao;YANG Xiong(School of Mechanical Engineering,Hubei University of Technology,Wuhan 430068,China)

机构地区：[1]湖北工业大学机械工程学院,湖北武汉430068

出　　处：《传感器与微系统》2023年第6期133-136,145,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(51775177)。

摘　　要：针对基于深度学习的立体匹配算法在挑战区域(如细节区域、弱纹理区域)存在一些误匹配的问题,提出一种基于多维注意力的立体匹配方法。首先,设计空间金字塔注意(SPA)模块,通过将空间金字塔结构与注意力机制相结合,获取更有效的全局上下文信息,来提高匹配精度;然后,构建注意力堆叠沙漏聚合(ASA)模块,在堆叠沙漏结构中引入注意力机制(AM),对匹配代价体进行重新校准,以进行更精确的视差计算;同时,采用可微分的Patch Match(DPM)方法,通过减少候选视差数量,构建轻量级匹配代价体,在保证模型匹配精度的同时,降低了计算资源的消耗。在Scene Flow、KITTI2015和KITTI2012数据集上的实验结果表明,与基准方法相比,所提算法在减少运行时间的同时提高了匹配精度。Aiming at the problem that stereo matching algorithm based on deep learning has some mismatches in challenge regions,such as detail regions and weak texture regions,a stereo matching method based on multi-dimensional attention is proposed.Firstly,the spatial pyramid attention(SPA)module is designed to improve the matching precision by combining the spatial pyramid structure with the attention mechanism to obtain more effective global context information.Then,the attention stacked hourglass aggregation(ASA)module is constructed,and the attention mechanism(AM)is introduced into the stacked hourglass structure to recalibrate the match cost volume for more accurate disparity calculation.At the same time,the differentiable Patch Match(DPM)method is used to reduce the number of candidate disparity and construct a lightweight match cost volume,which not only ensures the model matching precision,but also reduces the consumption of computing resources.Experimental results on Scene Flow,KITTI2015 and KITTI2012 datasets show that compared with the benchmark method,the proposed algorithm not only reduces the running time,but also improves the matching precision.

关 键 词：深度学习 立体匹配 注意力机制 匹配代价体 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]