汽车工程 ›› 2023, Vol. 45 ›› Issue (12): 2291-2298.doi: 10.19562/j.chinasae.qcgc.2023.12.011

所属专题: 智能网联汽车技术专题-感知&HMI&测评2023年

• • 上一篇    下一篇

基于局部平面引导层的无监督单目红外图像深度估计

石琴1,2,陈雅芳1,2,程腾1,2(),张强3,王文冲3,石本义4   

  1. 1.合肥工业大学汽车与交通工程学院,合肥  230000
    2.安徽省智慧交通车路协同工程研究中心,合肥  250000
    3.奇瑞汽车股份有限公司,芜湖  241000
    4.合肥英睿系统技术有限公司,合肥  230000
  • 收稿日期:2023-06-02 修回日期:2023-06-22 出版日期:2023-12-25 发布日期:2023-12-21
  • 通讯作者: 程腾 E-mail:cht616@hfut.edu.cn
  • 基金资助:
    安徽省自然科学基金(2208085MF171);中央高校基本科研业务费专项资金(JZ2023YQTD0073);国家自然科学基金(82171012);安徽省新能源汽车暨智能网联汽车创新工程项目(JZ2021AFKJ0002)

Unsupervised Monocular Infrared Image Depth Estimation Based on Local Plane Guidance Layer

Qin Shi1,2,Yafang Chen1,2,Teng Cheng1,2(),Qiang Zhang3,Wenchong Wang3,Benyi Shi4   

  1. 1.School of Automobile and Transportation Engineering,Hefei University of Technology,Hefei  230000
    2.Anhui Intelligent Transportation Vehicle-Road Collaborative Engineering Research Center,Hefei  250000
    3.Chery Automobile Co. ,Ltd. ,Wuhu  241000
    4.Hefei Yingrui System Technology Co. ,Ltd. ,Hefei  230000
  • Received:2023-06-02 Revised:2023-06-22 Online:2023-12-25 Published:2023-12-21
  • Contact: Teng Cheng E-mail:cht616@hfut.edu.cn

摘要:

当前无监督单目红外图像深度估计方法难以处理低纹理和低对比度区域,导致估计效果差,因此本文提出了一种基于局部平面引导层的无监督单目红外图像深度估计算法。该算法由连续视频帧输入、多尺度特征提取、ASPP和局部平面引导层、计算损失、联合训练、输出图像模块组成。首先,通过采用多个小分辨率灰度块和多尺度特征融合,解决了红外图像边缘模糊和遮挡物体等问题。其次,通过利用局部平面引导层在深度图像上引入一个平面约束,减小了深度图像中的噪声和不连续性,以及传统算法对低纹理区域处理不明显等问题。实验结果表明,所提出的深度估计算法有效提高了单目深度估计的精度并减小误差,在Iray数据集上的Abs RelSq RelRMSRMS(log)分别为0.262、3.621、9.473、0.332,在阈值指标小于1.25、1.252、1.253时,准确率达到了60.5%、85.2%、94.5%。

关键词: 无监督, 红外图像, 单目深度估计, 局部平面引导层, 损失函数

Abstract:

At present, the unsupervised monocular infrared image depth estimation method is difficult to deal with low texture and low contrast areas, resulting in poor estimation effect, so an unsupervised monocular infrared image depth estimation algorithm based on local plane guide layer is proposed in this paper. The algorithm consists of continuous video frame input, multi-scale feature extraction, ASPP and local planar guidance layer, computational loss, joint training, and output image module. Firstly, by using multiple small-resolution grayscale blocks and multi-scale feature fusion, the problems of blurring edges and occluding objects in infrared images are solved. Secondly, by using the local plane guidance layer to introduce a plane constraint on the depth image, the noise and discontinuity in the depth image are reduced, and the problem of lack of clear processing of low texture areas of the traditional algorithm is solved. The experimental results show that the proposed depth estimation algorithm effectively improves the accuracy of monocular depth estimation and reduces the error, and the Abs RelSq RelRMSRMS(log) on the Iray dataset is 0.262, 3.621, 9.473 and 0.332, respectively, and the accuracy reaches 60.5%, 85.2% and 94.5% when the threshold indicators are less than 1.25, 1.252 and 1.253.

Key words: unsupervised, infrared image, monocular depth estimation, local planar guidance layer, loss function