深浅层特征结合的自监督立体匹配  

Self-supervised stereo matching combining deep features and shallow features

作　　者：葛兰 贾振堂 Ge Lan;Jia Zhentang(College of Electronical and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海201306

出　　处：《电子测量技术》2023年第12期143-149,共7页Electronic Measurement Technology

基　　金：国家自然科学基金(62105196)项目资助。

摘　　要：针对现有的立体匹配算法物体细节部分估计效果较差、有监督算法依赖大量真实视差图等问题,本文提出了一种深浅层特征结合的自监督立体匹配算法。该算法在特征提取网络中嵌入通道注意力机制来提取图片的浅层和更具表征能力的深层特征。基于深层特征构建代价体积预测初始视差图,并用浅层特征指导初始视差图进行优化。此外在损失函数部分在左右视差一致性损失的基础上本文提出左右特征一致性损失,加强浅层特征信息对视差的约束作用,提高算法的鲁棒性。本文在KITTI 2015数据集上训练评估,并应用到拍摄的实际场景中。实验结果表明,本文提出的方法与其他算法相比能获得更好的效果,特别是在视差突然变化的细节区域。Aiming at the poor estimation of the existing stereo matching algorithms on the details of objects and the fact that supervised algorithms relying on a large number of groundtruth disparity maps,this paper proposes a self-supervised stereo matching algorithm combining deep and shallow features.The algorithm embeds Efficient Channel Attention in the feature extraction network to extract shallow and more expressive deep features of the picture.The cost volume predicting initial disparities are constructed based on the deep features,and the shallow features are used to guide the optimization of the initial disparities.In addition,in the loss function section,on the basis of the left and right disparity consistency loss,this paper proposes the left and right feature consistency loss,which strengthens the constraint effect of shallow feature information on disparity maps and improves the robustness of the algorithm.This article trains and evaluates on the KITTI 2015 dataset and applies it to the actual scenes taken by us.Experimental results show that the proposed method can achieve better results than other algorithms,especially in the details where the disparity changes suddenly.

关 键 词：立体匹配 自监督 通道注意力机制 浅层特征 深层特征 视差图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]