基于动态视差分布的立体匹配模型  

Stereo matching model based on dynamic disparity distribution

作　　者：商捷 王逸涵 罗建桥[1] 熊鹰[1] 李柏林[1] SHANG Jie;WANG Yi-han;LUO Jian-qiao;XIONG Ying;LI Bai-lin(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学机械工程学院,四川成都610031

出　　处：《计算机工程与设计》2025年第1期1-8,共8页Computer Engineering and Design

基　　金：四川省科技计划基金项目(重点研发项目)(2021YFN0020)。

摘　　要：为解决目前立体匹配模型中随机不确定性的量化未充分利用训练数据而导致视差估计性能提升受限的问题,在ACV-Net模型基础上提出一种随机不确定性融合标签分布学习的立体匹配模型,记为ACV-Net-AL。利用ACV-Net模型提取特征,形成随机不确定性空间和视差空间;对随机不确定性空间采用高斯分布的概率密度函数进行量化随机不确定性;融合随机不确定性和标签视差值形成正态的动态视差分布,监督视差值的预测。在KITTI数据集上的实验结果表明,相较于ACV-Net,所提模型的视差值异常点的比例下降了1.99%,在表征预测置信度的同时显著提高视差估计性能。To solve the problem that the quantization of aleatoric uncertainty in the current stereo matching model does not make full use of the training data,which leads to the limited performance improvement of disparity estimation,a stereo matching model based on the ACV-Net model with aleatoric uncertainty fusion label distribution learning was proposed,which was denoted as ACV-Net-AL.The ACV-Net model was used to extract features to form aleatoric uncertainty spaces and disparity spaces.The aleatoric uncertainty space was quantified by the probability density function of Gaussian distribution to obtain a measure of aleatoric uncertainty.The normal dynamic disparity distribution was formed by combining aleatoric uncertainty and label disparity value,and the prediction of disparity value was supervised.Experimental results on KITTI dataset show that compared with ACV-Net,the proportion of disparity outliers in the proposed model is reduced by 1.99%,and the performance of disparity estimation is significantly improved while the prediction confidence is characterized.

关 键 词：立体匹配 视差估计 随机不确定性 标签分布 动态视差分布 高斯分布 置信度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]