先进驾驶辅助系统中基于单目视觉的场景深度估计方法  

Scene Depth Estimation Based on Monocular Vision in Advanced Driving Assistance System

作　　者：丁萌[1,2] 姜欣言 Ding Meng;Jiang Xinyan(College of Civil Aviation,Nanjing University of Aeronautics and Astronautics,Nanjing,Jiangsu 211106,China;Key Laboratory of Aircraft Health Monitoring and Intelligent Maintenance,Civil Aviation Administration of China,Nanjing,Jiangsu 211106,China)

机构地区：[1]南京航空航天大学民航学院,江苏南京211106 [2]中国民用航空局飞机健康监测与智能维护重点实验室,江苏南京211106

出　　处：《光学学报》2020年第17期131-139,共9页Acta Optica Sinica

基　　金：国家自然科学基金(61673211);国家自然科学基金联合基金(U1633105);中央高校基本科研业务费专项(NS2020049)。

摘　　要：针对先进驾驶辅助系统对车辆前视景深信息的需求,在无监督学习框架下提出了一种基于单目视觉的场景深度估计方法。为了降低不同尺寸的前视目标对景深估计结果的影响,采用金字塔结构对输入图像进行预处理;在训练过程中,将深度估计问题转化为图像重建问题,利用双目图像设计了新的损失函数代替真实深度标签,解决了真实场景景深数据难以获取的问题;将中间多尺度的视差图与原输入图像的尺寸统一,改善了深度图中的空洞现象,提升了景深估计精度。在KITTI和Make3D数据集上的定量与定性对比结果表明,本方法可以获得准确度较高的绝对景深数据,且具有良好的泛化能力。在真实道路场景下的实验结果表明,本方法可以利用单张车载前视图像得到对应的像素级景深信息。Aiming at that the requirements advanced driving assistance system for vehicle forward looking depth of field information,this paper proposes a scene depth estimation method based on monocular vision under the framework of unsupervised learning.In order to reduce the influence of forward looking targets with diverse sizes on the depth estimation results,the proposed method uses a pyramid structure to preprocess the input image.In the training process,the depth estimation problem is transformed into an image reconstruction problem,and a new loss function is designed using binocular images instead of the true depth label,which solves the problem that the depth data of the real scene is difficult to obtain.The size of disparity map and original input image is unified,which improves the hole phenomenon in depth map and improves the accuracy of scene depth estimation.The quantitative and qualitative comparison results on the KITTI and Make3 D datasets show that the proposed method can obtain high accuracy absolute depth of field data and has good generalization ability,Experimental results in real road scenes show that the proposed method can obtain pixel level depth of field information from a single vehicle forward looking image.

关 键 词：深度估计 卷积神经网络 无监督学习 多尺度统一 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]