基于改进SIFT的室内机器人惯性视觉定位系统  

An Indoor Robot Inertial Vision Positioning System Based on Improved SIFT

作　　者：李港 赖钦伟 蒋林 雷斌[1,3] 梁铧杰 LI Gang;LAI Qin-wei;JIANG Lin;LEI Bin;LIANG Hua-jie(Institute of Robotics and Intelligent Systems,Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,China;Zhuhai A Micro Semiconductor Co.,Ltd.,Zhuhai 519000,China;Key Laboratory of Metallurgical Equipment and Control Technology,Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,China)

机构地区：[1]武汉科技大学机器人与智能系统研究院,武汉430081 [2]珠海市一微半导体有限公司 [3]武汉科技大学冶金装备及其控制教育部重点实验室,武汉430081

出　　处：《组合机床与自动化加工技术》2023年第2期155-159,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家重点研发计划项目(2019YFB1310000)。

摘　　要：对移动机器人室内单目视觉定位存在的定位精度较低,光照鲁棒性差的问题,提出一种基于改进SIFT的实时惯性视觉定位系统。首先压缩SIFT尺度空间,降低描述子维度;再利用结合位置信息的描述子曼哈顿距离进行FLANN匹配,PROSAC迭代优化。在跟踪阶段结合IMU获取的初值约束匹配,剔除误匹配。在回环检测阶段利用惯导位姿给定位姿初值范围,局部搜索SIFT特征点匹配实现重定位。在室内天花板数据集上验证改进SIFT算法,匹配准确率平均提升12.4%。在室内多个场景验证定位精度,平均定位误差为0.076 m。实验结果表明,所设计的算法能够在算力受限平台实现较精准的实时室内定位。A real-time inertial visual positioning system based on improved SIFT is proposed for the problems with low positioning accuracy and poor illumination robustness in indoor mobile robots.Compressing SIFT scale space and reduce the descriptor dimensions.FLANN matching using the descriptor Manhattan distance combining position information was optimized by PROSAC iteratively.In the tracking phase, mismatches were excluded by combining the initial value constraint matching of IMU acquisition.In the loop detection stage, the inertial navigation pose is used to give the positioning pose initial value range, and we can locally search for SIFT feature point matching to achieve relocation.The improved SIFT algorithm was validated on the indoor ceiling dataset, with an average matching accuracy improvement of 12.4%.The positioning accuracy was verified in multiple scenarios indoors, with an average positioning error of 0.076 m.The experimental results show that the designed algorithm can achieve more accurate real-time indoor positioning in the computing power limited platform.

关 键 词：SIFT 单目视觉 惯性视觉定位 回环检测 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]