第16卷第1期 智能系统学报 Vol.16 No.1 2021年1月 CAAI Transactions on Intelligent Systems Jan.2021 D0L:10.11992tis.202009009 基于迁移学习的移动机器人单帧图像坡度检测算法 辛菁，杜柯楠，王媛媛，刘丁 (西安理工大学自动化与信息工程学院，陕西西安710048) 摘要：针对未知环境下移动机器人平稳上坡控制对坡度感知精度的要求，本文提出了一种基于迁移学习的移 动机器人单帧图像坡度检测算法。利用室内图像标准数据集训练深度卷积神经场一全连接超像素池化网 络(deep convolutional neural field-fully connected superpixel pooling network,DCNF-FCSP)并获得室内单帧图像深度 估计网络模型：将DCNF-FCSP模型中前5个图像特征提取层的网络参数迁移至室外图像深度估计网络中：固 定室外图像深度估计网络中图像特征提取部分的网络参数，利用室外图像数据集对剩余5层的网络参数进行 训练，从而得到室外单帧图像深度估计网络：将其应用在移动机器人坡度检测中，根据单帧斜坡图像估计出斜 坡角度。标准数据集和实际场景的深度估计和坡度检测实验表明：本文所提出的基于迁移学习的移动机器人 单帧图像坡度检测算法能够仅根据移动机器人车载相机采集的单帧斜坡RGB图像就可估计出精确的斜坡角 度，满足移动机器人在未知环境中对坡度感知精度的要求。 关键词：未知环境：移动机器人：坡度检测：室外：单帧图像：深度估计；迁移学习：深度卷积网络 中图分类号：TP391文献标志码：A文章编号：1673-4785(2021)01-0081-11 中文引用格式：辛菁，杜柯楠，王媛媛，等.基于迁移学习的移动机器人单帧图像坡度检测算法J川.智能系统学报，2021， 16(1):81-91. 英文引用格式：XIN Jing,DU Kenan,,Wang Yuanyuan,ctal.Single frame image slope detection algorithm for mobile robots based on transfer learning[Jl.CAAI transactions on intelligent systems,2021,16(1):81-91. Single frame image slope detection algorithm for mobile robots based on transfer learning XIN Jing,DU Kenan,Wang Yuanyuan,LIU Ding (School of Automation and Information Engineering,Xi'an University of Technology,Xi'an 710048,China) Abstract:To meet the requirement of slope perception accuracy for stable uphill control of mobile robots in an un- known environment,a single frame image slope detection algorithm for mobile robots is proposed in this paper based on transfer learning.First,the deep convolutional neural field-fully connected superpixel-pooling network(DCNF-FCSP)is trained using a standard indoor image dataset,and the depth estimation network model of indoor single frame images is obtained.Second,the network parameters of the first five image feature extraction layers in the DCNF-FCSP model are transferred to the outdoor image depth estimation network.Then,the network parameters of the image feature extrac- tion part in the outdoor image depth estimation network are fixed,and the network parameters of the remaining five lay- ers are trained using the outdoor image dataset;thus the outdoor single frame image depth estimation network is ob- tained.Finally,it is applied to the slope detection of a mobile robot,and the slope angle is estimated according to the single frame slope image.The depth estimation and slope detection experiments on a standard dataset and in actual scenes show that the proposed algorithm can estimate the accurate slope angle according to only a single frame slope RGB image captured by the vehicle-mounted camera of a mobile robot.The proposed algorithm meets the requirements of the slope perception accuracy of a mobile robot in an unknown environment. Keywords:unknown environment;mobile robot;slope detection;outdoor;single image;depth estimation;transfer learning;deep convolutional network 收稿日期：2020-09-01. 未知环境地形的自主感知是移动机器人自主 基金项目：国家自然科学基金项目(61873200,61833013， 导航的基础和根本保证，也是移动机器人进行其 U20A20225). 通信作者：刘丁.E-mail:liud@xaut.edu.cn 他行为决策的前提。坡度检测作为移动机器人自DOI: 10.11992/tis.202009009 基于迁移学习的移动机器人单帧图像坡度检测算法 辛菁，杜柯楠，王媛媛，刘丁 （西安理工大学 自动化与信息工程学院，陕西 西安 710048） 摘 要：针对未知环境下移动机器人平稳上坡控制对坡度感知精度的要求，本文提出了一种基于迁移学习的移 动机器人单帧图像坡度检测算法。利用室内图像标准数据集训练深度卷积神经场−全连接超像素池化网 络 (deep convolutional neural field-fully connected superpixel pooling network, DCNF-FCSP) 并获得室内单帧图像深度 估计网络模型；将 DCNF-FCSP 模型中前 5 个图像特征提取层的网络参数迁移至室外图像深度估计网络中；固 定室外图像深度估计网络中图像特征提取部分的网络参数，利用室外图像数据集对剩余 5 层的网络参数进行 训练，从而得到室外单帧图像深度估计网络；将其应用在移动机器人坡度检测中，根据单帧斜坡图像估计出斜 坡角度。标准数据集和实际场景的深度估计和坡度检测实验表明：本文所提出的基于迁移学习的移动机器人 单帧图像坡度检测算法能够仅根据移动机器人车载相机采集的单帧斜坡 RGB 图像就可估计出精确的斜坡角 度，满足移动机器人在未知环境中对坡度感知精度的要求。 关键词：未知环境；移动机器人；坡度检测；室外；单帧图像；深度估计；迁移学习；深度卷积网络 中图分类号：TP391 文献标志码：A 文章编号：1673−4785(2021)01−0081−11 中文引用格式：辛菁, 杜柯楠, 王媛媛, 等. 基于迁移学习的移动机器人单帧图像坡度检测算法 [J]. 智能系统学报, 2021, 16(1): 81–91. 英文引用格式：XIN Jing, DU Kenan, Wang Yuanyuan, et al. Single frame image slope detection algorithm for mobile robots based on transfer learning[J]. CAAI transactions on intelligent systems, 2021, 16(1): 81–91. Single frame image slope detection algorithm for mobile robots based on transfer learning XIN Jing，DU Kenan，Wang Yuanyuan，LIU Ding (School of Automation and Information Engineering, Xi’an University of Technology, Xi’an 710048, China) Abstract: To meet the requirement of slope perception accuracy for stable uphill control of mobile robots in an un￾known environment, a single frame image slope detection algorithm for mobile robots is proposed in this paper based on transfer learning. First, the deep convolutional neural field-fully connected superpixel-pooling network (DCNF-FCSP) is trained using a standard indoor image dataset, and the depth estimation network model of indoor single frame images is obtained. Second, the network parameters of the first five image feature extraction layers in the DCNF-FCSP model are transferred to the outdoor image depth estimation network. Then, the network parameters of the image feature extrac￾tion part in the outdoor image depth estimation network are fixed, and the network parameters of the remaining five lay￾ers are trained using the outdoor image dataset; thus the outdoor single frame image depth estimation network is ob￾tained. Finally, it is applied to the slope detection of a mobile robot, and the slope angle is estimated according to the single frame slope image. The depth estimation and slope detection experiments on a standard dataset and in actual scenes show that the proposed algorithm can estimate the accurate slope angle according to only a single frame slope RGB image captured by the vehicle-mounted camera of a mobile robot. The proposed algorithm meets the requirements of the slope perception accuracy of a mobile robot in an unknown environment. Keywords: unknown environment; mobile robot; slope detection; outdoor; single image; depth estimation; transfer learning; deep convolutional network 未知环境地形的自主感知是移动机器人自主 导航的基础和根本保证，也是移动机器人进行其 他行为决策的前提。坡度检测作为移动机器人自 收稿日期：2020−09−01. 基金项目：国家自然科学基金项 目 (61873200， 61833013， U20A20225). 通信作者：刘丁.E-mail：liud@xaut.edu.cn. 第 16 卷第 1 期 智 能 系 统 学 报 Vol.16 No.1 2021 年 1 月 CAAI Transactions on Intelligent Systems Jan. 2021