摘要
视觉位置识别是移动机器人维持高精度定位和维护地图一致性的重要手段。然而,受视点和外观变化的双重干扰,位置识别问题仍然极具挑战性。本文提出了一种基于局部与全局描述符紧耦合联合决策的分层式视觉位置识别方法。该方法基于多任务知识蒸馏来学习描述符提取能力。经过良好训练的轻量化模型以紧耦合的形式同时提取图像的全局和局部描述符,并进一步实现局部描述符的二值化表示和词袋空间映射。在所构建的位置识别架构中,提出了分层式识别策略进行由粗到精的位置检索,并基于相位相关法分配全局和局部描述符的联合决策权重。在多项基准数据集上的评估结果证实,所提方法在可接受的匹配效率下实现了匹配性能的显著提升,在多种复杂环境下表现出较强的泛化性和鲁棒性。
Visual place recognition(VPR)is an important means for mobile robots to maintain high-precision localization and map consistency.However,due to the interference of viewpoint and appearance changes,the VPR problem remains extremely difficult.We propose a hierarchical VPR method based on joint decision-making of tightly coupled local and global descriptors.The proposed approach learns the ability to extract descriptors based on knowledge distillation.The well-trained lightweight model extracts global and local descriptors of an image in a tightly coupled form,further converts local descriptors into a binary representation,and maps it to the Bag of Visual Words space.In the constructed VPR architecture,a hierarchical recognition strategy is presented for coarse-to-fine place retrieval and a phase-correlation-based approach is employed to assign the joint decision weights of global and local descriptors.The evaluation results on several benchmark datasets confirm that the proposed approach achieves a significant improvement in performance with acceptable matching efficiency and exhibits strong generalization and robustness in various complex environments.
作者
李康宇
王西峰
朱守泰
LI Kangyu;WANG Xifeng;ZHU Shoutai(China Academy of Machinery Science and Technology Group,Beijing 100037,China;Machinery Technology Development Co.,Ltd.,Beijing 100037,China;Beijing Jiaotong University,Beijing 100091,China)
出处
《信息与控制》
CSCD
北大核心
2024年第3期400-415,共16页
Information and Control
基金
国家重点研发计划项目(2020YFB1313304)
关键词
视觉位置识别
闭环检测
深度学习
视觉SLAM
场景识别
visual place recognition
loop closure detection
deep learning
visual SLAM
scene recognition
作者简介
通信作者:李康宇,liky@mtd.com.cn,(1995-),男,博士生。研究领域为位置识别,深度学习;王西峰(1964-),男,研究员。研究领域为机电一体化;朱守泰(2001-),男,本科生。研究领域为计算机视觉。
