【作者】 张波；

【导师】 田蔚风；

【作者基本信息】 上海交通大学 ， 精密仪器及机械， 2007， 博士

【摘要】 以探测搜索、捕获跟踪、激光瞄准为技术依托的机载光电稳定跟踪系统已成为发展新一代光电装备的战略基点之一,对于国防安全、社会稳定、现代化生产和生活有着非常重要的意义。在机载光电稳定跟踪系统中,图像跟踪是一个关键技术,直接决定了跟踪系统的性能。同时,图像跟踪也是计算机视觉领域的一个热点研究问题,在视觉监控、人机互动、机器人导航等领域有着广泛的应用。图像跟踪方法大致可分为两大类:概率跟踪方法和确定性跟踪方法。概率跟踪方法由于跟踪性能稳定、可靠,已成为图像跟踪的主流方法,卡尔曼滤波和粒子滤波是这类方法的典型代表。卡尔曼滤波对系统模型和后验分布有严格限制,只能处理线性、高斯、单模态的情况,而在图像跟踪应用中,后验概率的分布往往是非线性、非高斯、多模态的,因此卡尔曼滤波的应用受到一定的限制。与卡尔曼滤波不同,粒子滤波对于系统模型没有特殊要求,而且能够保持状态的多模态分布,不易受杂波的影响,在跟踪领域得到了很大发展。但常规粒子滤波跟踪算法存在计算量大、采样效率低等问题,而且由于实际跟踪场景的复杂性,给常规粒子滤波跟踪算法提出了很大的挑战。为了提高粒子滤波跟踪算法的鲁棒性,进而为科学研究和工程应用提供理论和算法支持,针对粒子滤波跟踪算法中存在的问题,本文提出了改进的粒子滤波跟踪算法。改进工作主要从两方面入手:一方面设计有效的采样算法,使采样粒子尽可能集中在目标附近,更好地描述目标状态后验概率,提高粒子滤波的采样效率,减小计算量;另一个方面提高粒子滤波似然评价的区分性和精确性,设计了颜色和形状双重信息融合的似然模型、自适应似然模型和混合似然模型,并且将这些似然模型结合到粒子滤波中实施跟踪。具体的研究工作主要有:第一,针对单一视觉信息在动态变化环境下描述目标不够充分、跟踪目标不够稳定的缺点,设计了一种基于粒子滤波框架的自适应信息融合跟踪算法,研究中利用颜色和形状双重信息描述目标,通过民主融合策略将两种信息融合在一起,使得跟踪算法能根据当前跟踪形势自适应调整两种信息的权重,实现信息间的优势互补。由于这种算法能够始终利用对当前跟踪场景稳定的信息跟踪目标,解决了动态变化场景下单一信息跟踪失效的问题。在设计粒子滤波跟踪算法时,利用自适应信息融合策略构建似然模型,提高了粒子滤波跟踪算法在复杂场景下的稳健性。实验结果表明,自适应信息融合跟踪算法在目标平动、转动、部分遮挡等情形下都能成功地跟踪目标。当跟踪场景中出现颜色相似目标时,采用多块颜色模型能够很好的区分目标,解决了相似目标冲突问题。第二,针对粒子滤波采样效率低,计算量大的问题,提出了两种改进的粒子滤波跟踪算法:辅助核粒子滤波跟踪算法和层次采样跟踪算法。辅助核粒子滤波算法首先采用辅助粒子滤波采样粒子,然后利用均值偏移移动粒子到后验密度模式处。与粒子滤波和核粒子滤波相比,辅助核粒子滤波算法在采样时充分考虑观测值,解决了粒子滤波和核粒子滤波采样时由于没有利用观测值而造成粒子不能完全覆盖在目标位置附近的问题;辅助核粒子滤波算法减少了均值偏移迭代次数,与核粒子滤波相比,降低了计算量,同时跟踪性能还可能更高。层次式采样策略在采样粒子时,既强调了粒子的集中性又考虑粒子的多样性,所得粒子集能更好地表达后验概率密度,对于目标发生突然快速运动的情况,其跟踪性能优于粒子滤波和核粒子滤波。实验表明,本文提出的两种跟踪算法在目标快速运动、平动、转动、遮挡、复杂背景下均能较好地跟踪目标。第三,为了反映跟踪过程中目标表观的变化,提高跟踪算法在动态场景下的适应性和稳健性,提出了一种基于自适应表观模型的跟踪算法。利用递增核密度近似方法在线更新灰度表观模型,方便了跟踪算法的实时处理。在实施跟踪算法时,采用自适应表观模型设计粒子滤波的似然模型。另外,采用遮挡处理策略减少异常像素对状态估计的影响。实验结果表明,与固定表观模型跟踪算法相比,基于自适应表观模型的跟踪算法在光照变化、姿态变化、表情变化以及部分或完全遮挡等情形下,均具有较好的跟踪效果。第四,针对跟踪场景的复杂性,为了使跟踪算法能“辨别”目标,提高跟踪算法区分目标和背景的能力,提出了一种混合表观模型跟踪算法。混合表观模型由固定表观模型、快速变化表观模型和特征基表观模型组成,能充分反映目标表观变化。其中特征基表观模型通过加权递增主成分分析方法在线更新,获得的特征基能更好地描述目标表观,提高跟踪算法“辨别”目标的能力。采用上述混合表观模型设计粒子滤波似然模型,使跟踪算法在复杂场景下更稳定,更精确。另外,在整个跟踪算法中,采用遮挡处理策略,进一步提高了跟踪算法在遮挡发生时的有效性。实验表明,所提算法在光照变化、姿态变化、部分遮挡、全部遮挡等环境下均能稳定可靠地跟踪目标,性能优于利用单一表观模型的跟踪算法。本论文的研究主要集中在单目标跟踪的应用中,但由于图像跟踪在计算机视觉领域有着广泛的应用背景,因此,本文的研究成果对于计算机视觉的其它相关应用同样具有重要的意义。更多还原

【Abstract】 Airborne electro-optical stabilizing and tracking system dependent on detection, search and infrared aiming has become one of keypoint of electro-optical equipments and has very important significance to national defence, social stabilization and production. In the tracking system, image tracking is a key technology and directly determines the performance of the tracking system. Meanwhile, image tracking is an active topic in computer vision field and has wide applications in surveillance, human-machine interaction and robot navigation, etc.Image tracking methods are roughly divided into two main categories: probabilistic tracking and deterministic tracking. Probabilistic tracking methods have beome dominant due to their stable and robust performance. Kalman filter and particle filer are their two classical representatives. Kalman filter has rigorous limitation to system models and posterior distribution and thus is merely capable of dealing with linear, Gaussian and mono-modal situations. However, in image tracking application, the posterior density is often non-linear, non-Gaussian and multi-modal and thus the application of Kalman filter is limited. Different from Kalman filter, Particle filter is greatly developed in tracking field because of its ability of maintaining multi-modal distribution of the state and robustness to noise. However, the conventional particle filter based tracking algorithm has some defciencies, such as high computational cost and low sampling efficiency. In addition, the complexity of tracking scenes poses great challengs on tracking algorithms.To improve the robustness of particle filter based tracking algorithms and futher provide theory and algorithm supports for science research and engineering applications, the thesis proposes some enhanced tracking algorithms which mainly improve the tracking performance from two aspects. One is to design effective sampling methods, cluster the particles around the target as possible as, thus better representing the posterior density and improving the sampling efficency and decreasing the computational load. The other is to design better discriminant and accurate likelihood models including the two cue fusion based likelihood model, the adaptive likelihood model and hybrid likelihood model. Then these likelihood models are embedded into particle filter framework for tracking. The research contents of the thesis are as follows:Firstly, to overcome the shortcoming of single visual cue in complex enviroments, a tracking algirhtm based on adaptive cue fusion mechanism is proposed. The color cue and shape cue are ulitlized to represent the target and democratic integration is applied to fusion these two cues, thus facilating the tracking algorithm on-line adjusting the weight of two cues and utilizing their strongpoints. Due to the use of reliable cues for tracking, the failure of single cue in complex scenes is solved. During designing particle filter based tracking algorithm, the likelihood model is constructed dependent on adaptive cue fusion mechanism, thus enhancing the robustness of tracking algorithm. The tracking results demonstrate that the tracking algorithm based on adaptive cue fusion is able to successfully track target in presence of move, rotation and partial occlusion. When the target with similar color appears, the tracking algorithm with multi-part color model can distinguish the target, thus addressing the collision problem of similar targets.Secondly, to deal with low sampling efficency and huge computational load, two enhanced particle filter based tracking algorithms are proposed: auxiliary kernel particle filter based tracking algorithm and hierarchical sample tracking algorithm. Firstly, auxiliary particle filter is applied for sampling particles and then mean shift is invoked to move particles to their local maximum of the posterior density. Compared with particle filter and kernel particle filter, auxiliary kernel particle filter takes recent observations into account, thereby alleviating the incapability of completely covering around the true target in particle filter and kernel particle filter and decreasing the computational cost. Hierarchical sample tracking algorithm considers not only the importance of particles but also the diversity, thus being capable of better representing the posterior density. The performance of hierarchical sample tracking algorithm is superior to that of particle filter and kernel particle filter when the tracked target moves suddenly. The experiments show two tracking algorithms proposed can better track targets with sudden move, rotation, occlusions in complex enviroments.Thirdly, to reflect the changes of target appearance during tracking and improve the robustness and reliabity of tracking algorithms in dynamic scenes, an adaptive appearance model based tracking algorithm is proposed. Incremental kernel density approximate is used to on-line update grey appearance model, thus facilating real-time processing. During implementing tracking algorithm, the likelihood model of particle filter is construted by adaptive appearance model and furthermore occlusion handling strategy is invoked to declare outlier pixels and deal with occlusion events, thus decreasing their influence on the appearance model. Numours experiments demonstrate that compared with the fixed appearance model based tracking algoirhtm, the tracking algoirhtm based on adaptive appearance model achieves more better tracking performance in presence of lighting changes, pose variations, expression changes, partial occlusions and even full occlusions.Fourthly, to improve the ability of distinguishing the target and background, a hybrid appearance model based tracking algorithm is developed. The hybrid appearance models are composed of a fixed appearance model, a fast change appearance model and an eigenbasis appearance model. Weighted incremental principal component analysis is invoked to learn and update the eigenbasis of object appearance, thus effectively increasing the ability of discriminating the object. The hybrid appearance model is further applied to design the likelihood model, thereby making the tracking algorithm more stable and accuate under complex environments. In addition, occlusion handing strategy is also used to provent the eigenbais from erroneously absorbing the appearance of occluding object. Experimental results show the propsed tracking algorithm is superior to the tracking algoirhtm dependent on single appearance model under the variations of lighting and pose, partial and even full occlusions.The purpose of the thesis mainly focuses on the application of tracking single target. However, due to the wide applications of image tracking in computer vision field, the research of the thesis has important significance to the related application of computer vision.更多还原