摘要
提高地图构建的一致性和解决地图重影等问题是即时定位和地图构建(SLAM)领域的基本任务,在地点识别领域已得到广泛关注。对于地点识别中识别和描述的关键环节,常见的策略包括从点云框架中提取关键特征并生成描述符,计算其相似度以获得地点相似度。然而,许多传统方法只能提取单一特征,导致环境描述能力不足。文中提出一种基于复数的鲁棒位置重识别框架,这一框架利用复数的可扩展性,融合高度和强度两种信息进行描述符构建,旨在增强环境描述能力和效率,提供更稳健的环境描述并提高整体准确性。与传统的过滤策略不同,文中将原有的两步过滤策略扩展为三步:首先,设计了旋转不变性快速检索策略;其次,针对复杂特征的分布差异,采用分布过滤策略;最后,利用多角度复合向量进行精确评分。在公共数据集和真实自动驾驶车辆上进行的实验,验证了所提出的方法与传统地点识别方案相比具有优越性和灵活性。提出的位置重识别方法实现了超过99%的准确率,同时将时间效率提高了30%以上。
Improving the consistency of map construction and eliminating map overlapping are fundamental tasks within the domain of simultaneous localization and mapping(SLAM).They have gained significant attention in the field of place identification.For the critical steps of identification and description in the place identification,the common strategies involve extracting key features from a point cloud frame,describing them as descriptors,and calculating their similarity to determine place similarity.However,many traditional approaches are capable of extracting single features only,leading to inadequate environment description capabilities.In this paper,a robust place re⁃identification framework based on complex numbers is proposed.In the framework,the scalability of complex numbers is utilized to fuse two types of information of height and intensity,so as to construct descriptors.This aims to enhance environmental description capabilities and efficiency,providing a more robust environment description and improving overall accuracy.Departing from the traditional filtering strategies,the original two⁃step filtering strategy is expanded into the one with three steps.A fast retrieval strategy for rotational invariance is designed first,and then,for the distribution differences of complex features,a distributional filtering strategy is implemented.Lastly,multi⁃angle complex vectors are utilized for accurate scoring.Experiments on public datasets and real autonomous vehicles show that the proposed method is superior and flexible in comparison with the traditional place re⁃identification scheme.The proposed place re⁃identification method achieves an accuracy of more than 99%.Meanwhile,its re⁃identification efficiency has been improved by more than 30%.
作者
王智佼
王治才
晏鹏博
王武祺
谢耀华
WANG Zhijiao;WANG Zhicai;YAN Pengbo;WANG Wuqi;XIE Yaohua(Gansu Changda Highway Co.,Ltd.,Lanzhou 730000,China;School of Civil Engineering,Southwest Jiaotong University,Chengdu 611756,China;School of Information Engineering,Chang’an University,Xi’an 710018,China;China Merchants Chongqing Communications Technology Research and Design Institute Co.,Ltd.,Chongqing 400067,China)
出处
《现代电子技术》
北大核心
2025年第11期180-186,共7页
Modern Electronics Technique
基金
甘肃省科技计划项目(22YF11GA299)。
关键词
回环检测
即时定位
地图构建
位置重识别
复数描述符
快速检索
loop closure detection
simultaneous localization
mapping
place re⁃identification
complex number descriptor
fast retrieval
