摘要
通过深度学习模型对室内楼道环境的视觉信息进行处理,帮助移动机器人在室内楼道环境下自主行走。为达到这个目的,将楼道环境对象分为路、门、窗户、消防栓、门把手和背景六类,通过图像的语义分割实现对象识别。在对楼道环境的六类对象进行分割的实验中发现,由于门把手比起其他对象小很多,影响了对它的识别效果;将六分类模型改为"5+2"分类模型,解决了这个问题。分类模型的基础是全卷积神经(FCN)网络,可以初步实现图像的分割。为了提高FCN网络的分割效果,从三个方面进行了实验研究:a)取出FCN网络的多个中间特征层,进行多层特征融合;b)考虑到移动机器人行走过程中视觉信息的时间序列特点,将递归神经网络(RNN)的结构纳入到FCN网络中,构成时间递归的t-LSTM网络;c)考虑到二维图像相邻像素之间的依赖关系,构成空间递归的s-LSTM网络。这些措施都有效地提高了图像的分割效果,实验结果表明,多层融合加s-LSTM的结构从分割效果和计算时间方面达到综合指标最佳。
This paper processed visual information from corridor scene inside buildings through deep learning models to help mobile robot walk autonomously in this environment. To this end,objects in corridor environments were classified into 6 classes： road,door,window,hydrant,door handle and background,and these objects in images of corridor scene were recognized through semantic segmentation. The recognition for door handle was not satisfied in experiments because of its relatively small size to other objects; instead of 6-classification,it used ＂5 ＋ 2 ＂-classification to solve this problem. The basis of this model was a fully convolution neural networks,which could segment images of corridor scene primitively. In order to improve the performance of the FCN network,this paper conducted experiments in three aspects： a） Combining features from multi-intermediate-layers of FCN network instead of only using features of last layer to form multi-layer-fusion FCN network. b） Introducing recurrent neural network（ RNN） into FCN network to form the temporal recursive t-LSTM network with the consideration of the time series feature of images from mobile robots＇ video cameras. c） Introducing recurrent neural network into FCN network to form the spatial recursive s-LSTM network with the consideration of the dependence among adjacent pixels of the two-dimensional images. The experimental results show that the combining of multi-layer-fusion and s-LSTM achieves good performance in segmentation and computational time consuming.
作者
徐风尧
王恒升
Xu Fengyao;Wang Hengsheng(College of Mechanical ＆ Electrical Engineering,Central South University,~7~angsha 410083,China;State Key Laboratory for High Per formance Complex Manufacturing,Changsha 410083,China)
出处
《计算机应用研究》
CSCD
北大核心
2018年第6期1863-1866,1886,共5页
Application Research of Computers
基金
国家"973"计划资助项目(2013CB035504)
中南大学中央高校基本科研业务费专项资金资助项目(2017zzts639)
关键词
图像语义分割
全卷积神经网络
递归神经网络
多层特征融合
移动机器人导航
image semantic segmentation
fully eonvolutional neural （ FCN ） network
recurrent neural network （RNN）
multi-layer feature fusion
mobile robot navigation
