CASIA OpenIR  > 毕业生  > 博士学位论文
基于移动视觉平台的行人检测与跟踪技术研究
Alternative TitleResearch on the Technologies of Pedestrian Detection and Tracking Based on the Mobile Vision Platform
王敏
Subtype工学博士
Thesis Advisor乔红
2012-06-01
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword智能交通 流形学习 行人检测 跟踪 Intelligent Transportation Manifold Learning Pedestrian Detection Tracking
Abstract本文主要针对行人保护系统中基于移动视觉平台的行人检测和跟踪这些关键技术进行研究。行人保护系统力图在汽车等移动平台上建立一个自主、智能的行人检测、跟踪、辅助驾驶系统,具有提高驾驶安全性、保障行人生命安全的重要应用意义。另一方面行人保护系统的研究技术涉及到传感器技术、自动化与控制、人工智能、模式识别、信息融合等多个领域,是多学科交叉的一个研究热点,是相关领域良好的验证平台,具有很高的科学理论意义。 和一般的目标检测与跟踪问题相比,基于移动视觉平台的行人检测和跟踪因其动态应用场合的特殊性有其自身的特点,主要体现在:1)算法应用环境的动态性;2)应用范围的广泛性;3)计算速度的实时性;和4)算法结构的精炼性。 针对这些问题和挑战,本文对移动视觉平台的行人检测和跟踪算法进行了深入的研究分析。 首先,本文从获取稳定本质特征的角度出发,对于行人检测问题,认为相关性(Context)特征是一类比较好的描述行人这种复杂非刚性物体的方式,进行改进并自己提出一种新特征;对于行人跟踪问题,认为运动连续性是跟踪问题的特点,基于流形学习的方法提出了运动连续性保持的本质特征。 其次,本文从建立特征模型的角度出发,对于行人检测问题,建立了最优层次结构分类器的模型,减少了分类计算的复杂度;对于行人跟踪问题,基于流形学习建立了行人的本质流形这样的生成模型,在本质低维空间实现了行人跟踪。 最后,本文从算法框架的角度出发,使用基于近邻搜索、贝叶斯理论等跟踪框架对算法有效性进行实验验证;同时提出了前景增强的预处理机制、双向映射机制等来增强算法的鲁棒性,使其适合多种应用环境。 本文的主要研究工作和贡献主要包括: 1) 针对行人检测算法中目标、环境带来的稳定有效特征提取的困难,本文从相关性(Context)的角度出发,分析了行人复杂姿态的特点及相关的特征提取算法,改进了一种基于形状相关性(Shape-Context)的特征提取算法,使得算法复杂度更小而且适合Adaboost组合分类器的最优特征选择机制。将其和最优层次结构分类器组合,能够准确的进行行人检测与姿态鉴别。 2) 针对行人检测算法中的Context特征提取算法计算复杂度高的问题,本文提出了一种新的Context特征提取算法――统计梯度相关性特征提取算法。该算法基于积分图像运算,能够满足实时性要求,同时能够对复杂多变的行人观测提取稳定有效的Context特征,并不受光照、多种姿态、复杂背景等影响,能够满足行人检测的鲁棒性要求。 3) 针对行人跟踪问题中一般具体特征(颜色、纹理、梯度等)容易受目标时空变化的影响,本文基于流形学习理论提取行人目标的本质、低维的特征表示,构建了行人的本质变量连续性保持的低维流形子空间,基于此低维流形空间实现了鲁棒的动态行人跟踪。并进一步提出了一种自适应最优特征选择机制作为候选样本的预处理,增强样本前景区域的显著性,从而提高了算法的准确性。 4) 针对基于流形学习的行人跟踪中的模板漂移、低维表示有误差等问题,本文从提出了一种双向映射机制将流形空间与图像观测空间紧密联系。该机制包括:一种新的从...
Other AbstractThis thesis focuses on the researches of pedestrian detection and tracking on the mobile platform, which are the key technologies in the Pedestrian Protection Systems (PPSs). PPSs strive to build up an independent and intelligent driver-assistant system for pedestrian detection and tracking. These technologies can enhance the safety of driving, and have applicative signification for protecting the pedestrian’s life.On the other hand, the technologies applied in PPSs refer to many research areas, such as sensors, automatic control, artificial intelligence, pattern recognition, information in- tegration and so on. They are research interests of cross-disciplines, and are also the evaluation platform of the theory from these areas, with high academic signification. Compared with general object detection and tracking, the pedestrian detection and tracking on mobile platform has its own characteristics due to its application, which includes 1) the dynamic of application environment, 2) the wide range of application, 3) real-time processing requirement and 4) the succinct procedure of the algorithm program. Focusing on these problems and challenges, this thesis makes deep research on the pedestrian detection and tracking algorithm for mobile vision platform. First, from the view of extracting intrinsic feature, this thesis introduces context feature to be a good description to represent the pedestrian appearance which is a nonrigid target. A feature extraction method is improved, and a new feature extraction method is also proposed for pedestrian detection. Moreover, we consider the movement continuity is a key characteristic of pedestrian tracking. An intrinsic feature with movement continuity preserving is extracted for pedestrian tracking based on manifold learning. Then, from the view of feature model, we introduce the hierarchical classifier model with multi-templates of target postures for pedestrian detection, which can re- duce the computational complexity. We also learn a generative model of intrinsic pedestrian manifold based on manifold learning. The pedestrian tracking can be real- ized in the intrinsic but low-dimensional space. At last, from the view of detection and tracking framework, we use neighbor search, Bayesian theory and so on to evaluate the effectiveness of these algorithms. Meanwhile, we proposed the preprocessing mechanism for foreground segmentation, bi-direction mapping and so on to improve the robustness of the algorithms, in w...
shelfnumXWLW1746
Other Identifier200918014628054
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6472
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
王敏. 基于移动视觉平台的行人检测与跟踪技术研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2012.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20091801462805(8536KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王敏]'s Articles
Baidu academic
Similar articles in Baidu academic
[王敏]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王敏]'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.