摘要
本文针对动态场景,在ORB-SLAM的基础上,提出一种新的语义SLAM地图构建方法,提高智能移动机器人对环境感知和场景认知的能力。以RGB-D为输入信息,对RGB信息和深度信息分别作ORB特征匹配和尺度判断,利用RANSAC算法进行位姿估计判断关键帧。通过基于金字塔池化改进的MASK-RCNN神经网络对关键帧进行语义分割。在分割好的关键帧上,通过查找表法结合语义信息剔除动态目标。处理好的关键帧用于构建语义地图,同时进行局部集束调整,最后再回环检测。原语义分割网络精确率为81.2%,改进的网络精确率达到90.5%。
Aiming at dynamic scenes,based on ORB-SLAM,a newsemantic SLAM map construction method is proposed to improve the ability of intelligent mobile robots to sense the environment and recognize scenes.With RGB-D as input information,ORB feature matching and scale judgment are performed on RGB information and depth information,respectively,and key frames are determined by pose estimation using RANSAC algorithm.Then MASK-RCNN neural network based on improved pyramid pooling is used to perform semantic segmentation on key frames.On the segmented key frames,the dynamic target is eliminated by the lookup table method combined with the semantic information.The processed key frames are used to construct a semantic map and perform local clustering adjustments at the same time,and then perform loop detection.The accuracy of the original semantic segmentation network is 81.2%,and the accuracy of the improved network reaches 90.5%.
作者
丁佳惠
刘翔
奚峥皓
DING Jiahui;LIU Xiang;XI Zhenghao(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201610,China)
出处
《智能计算机与应用》
2020年第7期17-22,共6页
Intelligent Computer and Applications
基金
上海市科委地方能力建设项目(15590501300)
作者简介
丁佳惠(1995-),女,硕士研究生,主要研究方向:计算机视觉;通讯作者:刘翔(1972-),男,博士,副教授,主要研究方向:人工智能Email:xiangliu@outlook.com
