融合深度学习和结构特征的视觉定位方法  

Visual Localization Method of Structural Features Fused with Deep Learning

作　　者：仉新 毛宇新 李锁 ZHANG Xin;MAO Yuxin;LI Suo(School of Mechanical Engineering,Shenyang Ligong University,Shenyang 110159,China;Shenyang Institute of Computing Technology,Chinese Academy of Sciences,Shenyang 110168,China;School of Software,Northeastern University,Shenyang 110169,China)

机构地区：[1]沈阳理工大学机械工程学院,辽宁沈阳110159 [2]中国科学院沈阳计算技术研究所,辽宁沈阳110168 [3]东北大学软件学院,辽宁沈阳110169

出　　处：《长江信息通信》2023年第5期84-87,共4页Changjiang Information & Communications

基　　金：辽宁省教育厅面上青年人才项目(LJKZ0258);2022年辽宁省科技厅博士科研启动基金计划项目(2022-BS-187)。

摘　　要：为实现复杂场景下移动机器人视觉定位和建图,解决传统方法误差累积及鲁棒性差的问题,提出了融合深度学习和结构特征的视觉定位方法。首先,引入注意力机制改进深度学习网络,提取场景语义信息,剔除动态特征对定位及建图的干扰;其次,结合场景中的特征点、特征线和特征面信息,综合考虑空间的几何结构信息,实现复杂场景下的结构化特征融合;然后,计算多特征融合的视觉词袋模型,提升非线性优化和闭环检测的准确性;最后,提出静态关键帧策略,通过特征匹配实现位姿跟踪,建立无重影的八叉树场景地图,实现机器人自主定位和导航。数据集和真实场景实验表明:改进方法提升了复杂场景下,移动机器人视觉定位的准确性和鲁棒性。In order to realize mobile robot visual localization and mapping in complex scenes and solve the problems of cumulative error and poor robustness of traditional methods,a visual localization method combining deep learning and structural features is proposed.Firstly,the attention mechanism is introduced to improve the deep learning network,extract scene semantic information,and eliminate the interference of dynamic features on localization and mapping.Secondly,the information of point feature,line feature and surface feature in the scene are combined,and the geometric structure information of the space is comprehensively considered to realize the structural feature fusion in complex scenes.Then,the multi-feature fusion visual bag of words model is calculated to improve the accuracy of nonlinear optimization and loop closure detection.Finally,a static keyframe strategy is proposed to realize pose tracking through feature matching,establish an octree map without ghosting,and realize autonomous localization and navigation of the robot.TUM dataset and real scene experiments show that the improved method improves the accuracy and robustness of mobile robot visual localization in complex scenes.

关 键 词：视觉定位 深度学习 目标检测 注意力机制 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]