SIA OpenIR  > 机器人学研究室
基于深度学习和序列匹配的长航程场景识别算法研究
Alternative TitleLong-term Place Recognition based on Deep Learning and Sequence Matching
殷鹏
Department机器人学研究室
Thesis Advisor徐卫良 ; 何玉庆
Keyword同步定位建模 场景识别 无监督学习 序列匹配
Pages96页
Degree Discipline机械电子工程
Degree Name博士
2018-11-27
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract同步定位建模指移动机器人在未知环境中,根据运动状态信息和环境观测信息进行同步的地图建模和自身定位,并被普遍应用于无人驾驶、最后一公里物流、仓库分拣等机器人系统中。同步定位建模的工作开始于1986年,至今已经有32年的历史。其中,场景识别作为移动机器人定位导航的核心模块之一,指移动机器人根据当前观测信息检测出历史相似轨迹的能力,对于全局地图的建模优化和自身的精确定位起着决定性的作用。但在实际的长航程导航任务中,场景识别的实时性和稳定性往往受到特征提取和场景匹配两个方面的挑战:1) 特征提取:指从局部中提取出具有唯一性的特征向量。传统的特征提取方法在光照强烈变化(室内、室外、地下、废墟场景变化)下无法提取稳定的特征;同时由于季节、天气变化等环境因素,难以提取出与环境条件信息无关的特征信息,导致定位建模方法只能适用于局部短时的导航任务。2) 场景匹配:指定移动机器人根据当前的特征信息检测轨迹中闭环的能力。传统的定位建模方法基于单帧图片进行闭环检测,但环境的条件改变、机器人快速运动与环境中的动态对象会对单帧特征的提取引入噪声,降低场景匹配精度。2012年,Milford等人提出了基于序列帧的场景匹配方法,降低对单帧特征的置信度依赖,提高场景匹配的精度和鲁棒性。但基于序列匹配的方法计算复杂度远远高于单帧匹配方法,如何提高场景匹配的效率是一个亟待解决的问题。基于以上导航定位、特征提取与场景匹配的分析,本文的工作旨在提出一种基于序列匹配的长航程的鲁棒且高效的定位建模方法。本文具体贡献如下: 针对当前基于视觉传感器场景特征提取能力不足的问题,提出了一种条件无监督学习的特征提取方法,本方法从数据的分布中自动学习环境中潜在的特征描述。在特征提取过程中,考虑了环境条件因素,从而提高了特征提取的稳定性和鲁棒性,为后续的场景匹配提供了稳定可靠的特征表达;? 激光传感器对光照变化、气候变化具有鲁棒性,针对激光点云的有效特征提取能力不足的问题,本文提出了一种无监督学习的视角无关特征提取方法,在不依赖精确人工标签的前提下,从局部投影地图中提取视角无关的激光特征信息;为了提高原始序列匹配算法的匹配效率,本文提出了一种基于多分辨率采样的快速匹配方法,平衡在长航程序列匹配中的精度和效率。围绕移动机器人长航程导航与定位,本文的工作依次从视觉场景特征提权、激光点云特征提取和场景快速匹配三方面的内容进行了深入研究并进行实验验证。于此同时,我们开发了一套数据采集平台,方便进行长航程的数据采集和场景匹配性能测试。实验结果验证了本论文提出的方法的有效性和实用性,为后续移动机器人长航程定位建模提供了一定的理论基础和技术支撑。
Other AbstractSimultaneous localization and mapping (SLAM) is the ability of a robot system to relocate itself within an unknown environment and build the map at same time, based on the sensor measurements and motion updating. The original SLAM work began from 1986, 32 years ago till now. Currently SLAM has became the core module in Autonomous Driving, last-mile delivery and Warehouse sorting robotics. Place recognition is the key module in the SLAM framework, which is used to find loop closure within the history trajectory and plays an important role in the map optimization and localization. But in real long-term navigation task, place recognition methods still face the challenges from the following three aspects: 1) Feature Extraction: is to extract unique feature descriptions for local scenes. Traditional traction extraction in SLAM methods may get affected by the illumination changes (from day to night, and abrupt light condition changing), thus can not provide stable features for long-term navigation task. On their other hand, appearance changed caused by seasons and weathers will also introduce measurement uncertainty in feature extraction; 2) Scene Matching: is to retrieval previous visited scenes based on the feature descriptions, and constructed a loop on the global map. Traditional SLAM methods rely on single-frame for LCD detecting, but this approach is unstable and will get external noise under variant environment conditions, when the robot system is moving fast or the environment has dynamic objects. In 2012, Milford et.al proposed a sequence matching base place recognition based, which is robust to environment conditions and dynamic objects. But since the computation complexity of sequence based matching method is higher than the single-frame based method. According to the previous analysis on place recognition, in this paper we aim to provide an efficiency and accurate place recognition method based on sequence matching for long-term navigation task. The main contributions of this paper are: ? Aim to learn robust feature in for local place description, we proposed an unsupervised feature learning method, which can learn the hidden features of scenes without human labeling. And the extracted feature is more robust to the environment condition changing and dynamic objects; For the sake of practical loop closure detection, we design an multi-resolution sampling based sequence matching method, which can improve the efficiency and accuracy of sequence matching greatly in the long-term navigation task; ? We design an multi-resolution sampling based sequential matching method in largescale place matching. Finally, we build a platform and relative software to gather the raw data in the cloud online for raining the robust feature and further algorithm evaluation requirements.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/23647
Collection机器人学研究室
Recommended Citation
GB/T 7714
殷鹏. 基于深度学习和序列匹配的长航程场景识别算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习和序列匹配的长航程场景识别算(8692KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[殷鹏]'s Articles
Baidu academic
Similar articles in Baidu academic
[殷鹏]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[殷鹏]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.