基于CTH的不适定区域精确无监督立体匹配算法  

Accurate Unsupervised Stereo Matching Algorithm for Ill-Posed Regions Based on CTH

作　　者：冯强龙 王晓峰 王海宇 陆正霖 丁坤岭 舒航 FENG Qianglong;WANG Xiaofeng;WANG Haiyu;LU Zhenglin;DING Kunling;SHU Hang(School of Mathematical and Physical Sciences,Chongqing University of Science and Technology,Chongqing 401331,China;School of Electronic and Electrical Engineering,Chongqing University of Science and Technology,Chongqing 401331,China)

机构地区：[1]重庆科技大学数理科学学院,重庆401331 [2]重庆科技大学电子与电气工程学院,重庆401331

出　　处：《重庆科技大学学报（自然科学版）》2024年第5期85-91,共7页Journal of Chongqing University of Science and Technology(Natural Sciences Edition)

基　　金：重庆市自然科学基金面上项目“复杂场景下融合全局图神经网络和三维内蕴几何约束的精准立体视觉研究”(CSTB2022NSCQ-MSX0398);重庆市研究生科研创新项目“复杂环境自动驾驶下融合双目立体视觉和激光雷达多模态的深度学习快速精确深度感知算法研究”(CYS240804);重庆科技大学硕士研究生创新计划项目“复杂场景下基于BEV视角的融合双目立体视觉和激光雷达多模态精确深度感知算法研究”(YKJCX2321103)。

摘　　要：无监督立体匹配算法的准确性直接影响深度估计的质量。目前的无监督立体匹配算法普遍缺乏全局特征提取能力,在边缘和遮挡等不适定区域的视差估计精度较低。为此,提出了一种基于CNN和Transformer的不适定区域全局与局部特征提取的精确无监督立体匹配算法。首先,提出了一种结合多尺度和跳跃连接特性的沙漏网络,设计了适用于立体匹配的CNN与Transformer并行特征提取(CTH)算法;其次,利用特征耦合单元(FCU)将局部特征与全局表示相融合,提升了模型整合全局信息的能力;最后,将原始视差监督损失函数引入基准模型的损失函数中,通过一致性损失引导网络学习更全面的特征。实验结果表明,与其他经典算法相比,该算法在不适定区域的精度和鲁棒性均有所提升。The accuracy of unsupervised stereo matching algorithms directly affects the quality of depth estimation.Current unsupervised stereo matching algorithms generally lack the ability to extract global features,resulting in low disparity estimation accuracy in ill-posed regions such as edges and occlusions.To address this issue,an accurate unsupervised stereo matching algorithm based on CNN and Transformer for global and local feature extraction in ill-posed regions is proposed.First,a sandglass network combining multi-scale and skipping connections is introduced,and a CNN-Transformer Hybrid(CTH)algorithm is designed for stereo matching to enable parallel feature extraction.Second,a Feature Coupling Unit(FCU)is used to fuse local features with global representations,enhancing the model′s ability to integrate global information.Finally,the original disparity supervision loss function is incorporated into the baseline model′s loss function,guiding the network to learn more comprehensive features through consistency loss.Experimental results demonstrate that compared to other classical algorithms,the proposed method improves accuracy and robustness in ill-posed regions.

关 键 词：无监督立体匹配 CNN TRANSFORMER CTH 全局和局部特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]