【作者】 王智灵；

【导师】 陈宗海；

【作者基本信息】 中国科学技术大学 ， 控制科学与工程， 2009， 博士

【摘要】 视觉信息是典型的非接触式传感信息。基于视觉信息的目标跟踪可以在不干涉观测目标,不对观测目标造成影响的前提下实现期望的功能。基于视觉信息的复杂环境中特定/非特定运动目标的可靠检测与鲁棒跟踪是当前备受关注的前沿方向。在实际应用中,观测数据集常含有大量离群数据,这就要求跟踪方法必须具有一定的鲁棒性。鲁棒性是跟踪算法实用化的前提。随着计算机硬件技术的发展,跟踪算法的实时性问题越来越多的依赖硬件方案解决;跟踪任务的核心问题,就是如何增强对各种应用环境和各种干扰因素的鲁棒性。研究如何提高复杂自然环境下运动目标检测与跟踪过程中的鲁棒性,不仅有助于实现和推广视觉信息的智能化自动处理,提升下一代移动机器人的智能行为,还将进一步推进人们对于生物视觉认知特性本身的认识。本文从鲁棒性需求、鲁棒性描述、鲁棒性处理策略和鲁棒的跟踪算法设计与评估的角度对基于视觉信息的运动目标跟踪任务进行研究。分别从鲁棒统计学和生物认知特性出发,针对目标跟踪任务不同环节——背景建模和背景抑制、运动估计和运动跟踪、目标跟踪——中的不同鲁棒性需求,对传统的鲁棒估计子进行改进,提出新的跟踪策略和实现算法。实验表明,这些方法与现有的算法相比,在效率、精度、鲁棒性等方面都具有一定优势。文章的主要工作和贡献如下:(1)首先阐述鲁棒估计的基本含义和基本理论,介绍鲁棒统计学中关于估计子鲁棒性的几个关键概念——离群数据、溃点和影响函数,着重讨论本文使用的几种鲁棒估计子——中位数估计子、RASNAC估计子、M估计子——的基本原理和各自的优缺点;然后分析非参数化估计方法与鲁棒估计之间的本质联系,介绍几种典型的非参数估计方法,着重分析基于核密度估计的Mean-Shift的性能和收敛性,并给出了实验验证。(2)以背景建模任务为示例,分析不同应用条件下的背景建模任务的鲁棒性需求,考察应用过程中干扰的引入原因,分析各种干扰现象相应的样本结构形态;为实现对多结构混杂数据的适应能力,本文对传统的RANSAC方法进行了改进,并首次将其推广到动态背景像素建模任务当中。通过MAD尺度估计子自适应估计阈值参数,经由随机采样、模型估计、模型数据扩充、多级建模的方式,使算法在保持高鲁棒性的同时仍能拥有较好的描述精度。在此基础上,采用三元组模型描述方式,以改进的多级RANSAC方法建立动态像素背景模型,以单端截尾均值估计子建立静态像素模型,并在一个统一的框架中快速更新。该方法有效的解决了RANSAC方法的固有缺陷,能容忍高比例的离群数据;该方法具有较好的描述能力,在保持算法准确性的前提下有较快的更新速度。通过对溃点的理论计算和统计分析验证了方法的鲁棒性,并与GMM估计、Median估计进行了对比。(3)研究基于运动信息的目标跟踪方法,应用光流方法和Kalman方法进行跟踪实验并分析其不足之处。针对Kalman估计子不能克服外点的影响,难以解决的运动分析过程中的非线性、粗差的问题,结合Turkey型M-估计子对UKF的迭代过程进行改造。M-估计子是一类重要的鲁棒估计子,可以通过选择不同类型的损失函数,实现不同的鲁棒效果。提出了一种M-UKF估计方法,该方法将M-估计的加权原理应用到UKF的迭代递推的过程中,可以有效地减弱或消除观测值中粗差的影响;在解决了运动模型非线性估计的同时,能较好的克服离群数据的干扰,大大提高了估计的鲁棒性。结合包含运动信息在内的多层次信息,给出了一个人的精细轮廓的跟踪方法和比较实验。(4)从生物视觉认知的角度对运动目标跟踪过程进行全面考察,提出了一种基于蛙眼认知特性的“模糊化区域理解”跟踪策略。首先详细分析了蛙眼视觉系统的行为特性、生理特性和神经特性,然后通过借鉴蛙眼认知的外部特性和神经生理特性,提出了一种与蛙眼跟踪方式类似的鲁棒跟踪策略;并以“模糊化区域理解”的方式,给出了这种跟踪策略的一种实现模式。将Mean-Shift滤波应用到“模糊化区域理解”的过程中,分析尺度参数解耦对理解效果的影响结果,并设计了一种自适应的局部尺度参数选择策略。最后,利用该方法解决存在背景场景突变、目标外观变化、环境光照变化、目标干扰运动等现象的复杂环境下的目标跟踪问题,提高跟踪方法对这些干扰因素的抵抗能力,并将该方法与传统的理解方法进行了对比。更多还原

【Abstract】 Contributed to the non-contact characteristic,visual-information-based object detection and tracking tasks can be achieved without interfering desired objects.The detection and tracking of moving objects under dynamic scenes becomes one of the most significant topics in computer vision research.In realistic application,the observed data sets often contain high percentage of outliers,which make it necessary to use robust methods for estimation.For any practical tracking algorithms,robustness of the tracker must be concerned.Recently,with the rapid development of computer hardware,the requirements of real-time computing are more dependent on hardwares’ performace.Then,robust problem,which still remains very difficlut,has been more and more essential and critical in tracking.The research on how to improve tracking robustness in complex natual circumstance,not only give benifit to the achievement and promotion of intelligent automatic visual information processing,but also will further promote the comprehension of biological visual understanding mechanism itself.This dissertation discusses the theories and applications of moving object tracking through robustness requirements,robustness description,robust treatment strategies and robust tracking algorithms.We analysis the different sub-tasks of tracking,such as background modeling and background substraction,move estimation and move tracking,object tracking.And its different robustness requirements are taken into account.Traditional robust estimators are improved.New tracking tactic and algorithms,which based on robust statistics or biological technology,are presented. Experiments show that these algorithms compared with the existing algorithms,have an advantage on the efficiency,accuracy,robustness,and so on.The main tasks and contributions of this thesis are:(1) First,some fundamental concepts about estimation theory are introduced.The statistical definitions of robust estimator(such as outliers,breakdown point and influence function) are given and some robust statistics methods(such as Median, M-estimator,RANSAC) are discussed.Their respective advantages and limits are examined.Nonparametric statistics has a close relationship with robust estimation.We also analysis some nonparametric estimation method,focus on mean-shift’s performance and convergence,and give experimental verification. (2) Taking the task of background modeling as an example,its robustness requirements under different conditions are analyzed.We consider the reasons of noise in different applications;give a structural analysis of its sample set.To dealing with multiple structural contaminated data,the traditional RANSAC is improved by MAD estimator and multi-procession.On this basis,a novel robust background modeling algorithm is presented.The model is established by an improved Multi-RANSAC approach for dynamic background pixels and by one-tail trimmed mean estimation for static pixels.A three-component cell is derived for the model so that it can be updated quickly in a unified framework.The method effectively overcomes the inherent deficiency in RANSAC and can tolerat a high percentage of oulier.It proves right even when there are more than 70 percent outliers and is fit for extraventricular natural scenes.Quantitative evaluation and comparisons with GMM estimator and Median estimator verified that the proposed method has much better performance.(3) M-estimator is one of the most important robust estimators,which can achieve different performance of robustness by choosing different loss function.In the process of estimating and tracking of moving object,the traditional Kalman filter-based method can’t overcome the influence of outliers,and has trouble in the condition of nonlinear outliers and gross outliers.Aiming at this problem,we combine the Turkey M-estimator with Unscened Kalman Filter and present a new M-UKF tracking algorithms.It can not only solve the problem of estimating the nonlinear effection,but also overcome the interference of gross outliers.Simulation and experiments show its validity.(4) After conducting a comprehensive inspection of object tracking process from the perspective of biological visual cognition,we develop a robust intelligent tracking tactic based on the intrinsic and extrinsic features of frog’s visual system.It is achieved through a procession called as "fuzzy region understanding".Mean-Shift based-method is applied and the effect of scale-parameters is discussed.An adaptive local scale-parameter adjustment is presented.The proposed method has been applied in the sences which may have gross noises such as local scene break or object appearance variation or interference with movement.Experiments and comparisons with two traditional algorithms demonstrate the validity and robustness of our algorightm.更多还原