融合背景权重直方图的目标跟踪

目的 考虑到融合校正背景权重直方图(CBWH)的Mean Shift(MS)目标跟踪算法只有CBWH更新而缺少目标模板更新,以及在目标遮挡时鲁棒性欠佳的不足.方法 结合卡尔曼滤波器(KF)在目标状态预测和参数更新方面的可靠性,将两层KF框架融入融合CBWH的MS.第1层KF框架为目标位置预测层,通过KF噪声与巴氏系数之间的关系,实现跟踪结果的自适应调整,减少遮挡对跟踪结果的影响;第2层KF框架为目标模板更新层,通过KF对目标模板中的每个非零元素进行滤波,实现目标模板与CBWH的同步更新,减少目标特征变化对跟踪结果的影响.结果 在背景干扰、遮挡以及特征变化等条件下进行实验,得到本文算法、融合CBWH的MS和传统MS的平均跟踪误差分别为5.43、19.2和51.43,显示本文算法的跟踪精度最高.同时本文算法也具有良好的实时性.结论 本文算法在融合CBWH的MS基础上,加入两层KF框架,解决了原算法缺少目标模板更新和在目标遮挡时鲁棒性欠佳的不足,最后实验验证了本文算法的有效性.     

An efficient parallel algorithm for geometrically characterising drawings of a class of 3-D objects

Labelling the lines of a planar line drawing of a 3-D object in a way that reflects the geometric properties of the object is a much studied problem in computer vision, considered to be an important step towards understanding the object from its 2-D drawing. Combinatorially, the labellability problem is a Constraint Satisfaction Problem and has been shown to be NP-complete even for images of polyhedral scenes. In this paper, we examine scenes that consist of a set of objects each obtained by rotating a polygon around an arbitrary axis. The objects are allowed to arbitrarily intersect or overlay. We show that for these scenes, there is a sequential lineartime labelling algorithm. Moreover, we show that the algorithm has a fast parallel version that executes inO(log³ n) time on an Exclusive-Read-Exclusive-Write Parallel Random Access Machine withO(n ³/log³ n) processors. The algorithm not only answers the decision problem of labellability, but also produces a legal labelling, if there is one. This parallel algorithm should be contrasted with the techniques of dealing with special cases of the constraint satisfaction problem. These techniques employ an effective, but inherently sequential, relaxation procedure in order to restrict the domains of the variables.This research was partially supported by the European Community ESPRIT Basic Research Program under contracts 7141 (project ALCOM II) and 6019 (project Insight II).     

基于水动力学、分数阶微分及Steger算法的线状目标提取

目的 线状目标的检测具有非常广泛的应用领域,如车道线、道路及裂缝的检测等,而裂缝是其中最难检测的线状目标。为避免直接提取线状目标时图像分割难的问题,以裂缝和车道线为例,提出了一种新的跟踪线状目标中线的算法。方法 对图像进行高斯平滑,用一种新的分数阶微分模板增强图像中的模糊及微细线状目标;基于Steger算法提出一种提取线状目标中心线特征点的算法,避免了提取整体目标的困难;根据水动力学思想将裂隙看成溪流,通过最大熵阈值处理后,先进行特征点的连接,再基于线段之间的距离及夹角进行线段之间的连接（溪流之间的融合）。结果 对300幅裂缝图像及4种类别的其他线状目标图像进行试验,并与距离变换、最大熵阈值法+细线化Otsu阈值分割+细线化、谷底边界检测等类似算法进行比较分析,本文算法检测出的线状目标的连续性好、漏检（大间隙少）和误检（毛刺及多余线段少）率均较低。结论 本文算法能够在复杂的线状目标图像中准确快速地提取目标的中心线,一定程度上改善了复杂线状目标图像分割难的问题。     

结合连续卷积算子的自适应加权目标跟踪算法

目的 在视觉跟踪领域中,特征的高效表达是鲁棒跟踪的关键,观察到在相关滤波跟踪中,不同卷积层表达了目标的不同方面特征,提出了一种结合连续卷积算子的自适应加权目标跟踪算法。方法 针对目标定位不准确的问题,提出连续卷积算子方法,将离散的位置估计转换成连续位置估计,使得位置定位更加准确;利用不同卷积层的特征表达,提高跟踪效果。首先利用深度卷积网络结构提取多层卷积特征,通过计算相关卷积响应大小,决定在下一帧特征融合时各层特征所占的权重,凸显优势特征,然后使用从不同层训练得到的相关滤波器与提取得到的特征进行相关运算,得到最终的响应图,响应图中最大值所在的位置便是目标所在的位置和尺度。结果 与目前较流行的3种目标跟踪算法在目标跟踪基准数据库（OTB-2013）中的50组视频序列进行测试,本文算法平均跟踪成功率达到85.4%。结论 本文算法在光照变化、尺度变化、背景杂波、目标旋转、遮挡和复杂环境下的跟踪具有较高的鲁棒性。     

Minimizing the total projection of a set of vectors, with applications to layered manufacturing

In layered manufacturing (LM), a three-dimensional polyhedral solid is built as a stack of two-dimensional slices. Each slice (a polygon) is built by filling its interior with a sequence of parallel line segments (of some small non-zero width), in a process called hatching. A critical step in hatching is choosing a direction which minimizes the number of segments. In this paper, this problem is approximated as the problem of finding a direction which minimizes the total projected length of a certain set of vectors. Efficient algorithms are proposed for the latter problem, using techniques from computational geometry. Experimental and theoretical analyses show that this approach yields results that approximate closely the optimal solution to the hatching problem. Extensions of these results to several related problems are also discussed.     

全局多极团的分层关联多目标跟踪

目的 由于背景的复杂性,光照的多变性以及目标的相关性等因素的影响,使得多目标跟踪算法的鲁棒性相对较差。目前,在多目标跟踪问题中面临的主要挑战包括：遮挡、误检、目标运动的复杂性以及由于目标具有相似的外观特征所引起的模糊性。针对以上问题,提出一种基于全局多极团的分层关联多目标跟踪算法。方法 该方法以数据关联中的全局关联为依托,基于分层和网络流思想,跟踪采用两层框架,每一层中均利用较短的轨迹片段形成更长的轨迹,根据网络流思想,首先构建网络的无向图,其中无向图的结点是由几个轨迹片段构成的,无向图权值的确定是利用目标的运动模型和外观模型的线性组合得到,然后借助聚合虚拟结点处理目标之间的遮挡问题,接着重点加入空间约束以解决身份转换的问题。最后利用最大二值整数规划在叠加片段上求解无向图,同时得到多个极大团。结果 实验在公共数据集上进行,通过在TUD-Stadmitte、TUD-Crossing、PETS2009、Parking Lot 1、Parking Lot 2、Town Center这6个数据集上验证,该方法对各个数据集跟踪准确度均有提高,其中针对数据集TUD-Stadmitte提高了5%以上,针对数据集Town Center处理的身份转换数量减少了12个。结论 本文依据数据关联思想,提出一种全局多极团的分层关联多目标跟踪算法,其中重点加入的空间约束能有效地处理多目标跟踪问题,尤其涉及遮挡问题,效果更佳。在智能视频监控领域中该方法具有实际应用价值。     

自适应卷积特征选择的实时跟踪算法

目的 针对深度卷积特征相关滤波跟踪算法因特征维度多造成的跟踪速度慢及其在目标发生形变、遮挡等情况时存在跟踪失败的问题,提出了一种自适应卷积特征选择的实时跟踪算法。方法 该算法先分析结合深度卷积特征的相关滤波跟踪算法定位目标的特性,然后提出使用目标区域和搜索区域的特征均值比来评估卷积操作,选取满足均值比大于阈值的特征通道数最多的卷积层,减少卷积特征的层数及维度,并提取该卷积层的有效卷积特征来训练相关滤波分类器,最后采用稀疏的模型更新策略提高跟踪速度。结果 在OTB-100标准数据集上进行算法测试,本文算法的平均距离精度值达86.4%,平均跟踪速度达29.9帧/s,比分层卷积相关滤波跟踪算法平均距离精度值提高了2.7个百分点,速度快将近3倍。实验结果表明,本文自适应特征选择的方式在保证跟踪精度的同时有效地提升了跟踪的速度,且优于当前使用主成分分析降维的方式;与现有前沿跟踪算法对比,本文算法的整体性能优于实验中对比的9种算法。结论 该算法采用自适应卷积通道和卷积层选择的方式有效地减少了卷积层数和特征维度,降低了模型的复杂度,提升了跟踪速度,利用稀疏模型更新策略进一步提升了跟踪的速度,减少了模型漂移现象,当目标发生快速运动、遇到遮挡、光照变化等复杂场景时,仍可实时跟踪到目标,具有较强的鲁棒性和适应性。     

A Fixed-Parameter Approach to 2-Layer Planarization

A bipartite graph is biplanar if the vertices can be placed on two parallel lines (layers) in the plane such that there are no edge crossings when edges are drawn as line segments between the layers. In this paper we study the 2-Layer Planarization problem: Can k edges be deleted from a given graph G so that the remaining graph is biplanar? This problem is NP-complete, and remains so if the permutation of the vertices in one layer is fixed (the 1-Layer Planarization problem). We prove that these problems are fixed-parameter tractable by giving linear-time algorithms for their solution (for fixed k). In particular, we solve the 2-Layer Planarization problem in O(k · 6^k + |G|) time and the 1-Layer Planarization problem in O(3^k · |G|) time. We also show that there are polynomial-time constant-approximation algorithms for both problems.     

Digitization scheme that assures faithful reconstruction of plane figures

In the present extensive work, we propose a digitization scheme for a broad class of plane figures. It includes arbitrary polygons (possibly, non-convex and with holes) and plane sets whose boundary consists of smooth curves or straight segments. Our approach is based on an appropriate scaling of the original continuous real object so that the obtained magnified object and its (appropriately constructed) digitization feature analogous geometric properties. As a bi-product of the presented theory we prove the strong NP-hardness of the problem of obtaining an optimal (i.e., with a minimal number of facets) polyhedral reconstruction in which the facets are trapezoids or triangles. This result implies the strong NP-hardness of the general polyhedral reconstruction problem, which was a long-standing open problem. As another major result, we show that the proposed digitization scheme allows faithful reconstruction of plane figures from the considered general class. The reconstructed set features the basic properties of the original object, such as location of curve and straight segments and inflection points of its boundary.     

Camera Autocalibration and the Calibration Pencil

We study the geometric object given by the set of lines incident with the absolute conic. We see that this object is given by a pencil of quadrics of P ⁵, which is characterized. We describe some of its most relevant properties for the camera autocalibration problem. Finally, we illustrate the applicability of the theory proposing a linear algorithm for the metric upgrading of a projective calibration of a set of ten or more cameras with varying parameters and known skew and aspect ratio.This revised version was published online in June 2005 with correction to CoverDate     

融合深度模型和传统模型的显著性检测

目的 显著性检测是图像和视觉领域一个基础问题,传统模型对于显著性物体的边界保留较好,但是对显著性目标的自信度不够高,召回率低,而深度学习模型对于显著性物体的自信度高,但是其结果边界粗糙,准确率较低。针对这两种模型各自的优缺点,提出一种显著性模型以综合利用两种方法的优点并抑制各自的不足。方法 首先改进最新的密集卷积网络,训练了一个基于该网络的全卷积网络（FCN）显著性模型,同时选取一个现有的基于超像素的显著性回归模型,在得到两种模型的显著性结果图后,提出一种融合算法,融合两种方法的结果以得到最终优化结果,该算法通过显著性结果Hadamard积和像素间显著性值的一对一非线性映射,将FCN结果与传统模型的结果相融合。结果 实验在4个数据集上与最新的10种方法进行了比较,在HKU-IS数据集中,相比于性能第2的模型,F值提高了2.6%;在MSRA数据集中,相比于性能第2的模型,F值提高了2.2%,MAE降低了5.6%;在DUT-OMRON数据集中,相比于性能第2的模型,F值提高了5.6%,MAE降低了17.4%。同时也在MSRA数据集中进行了对比实验以验证融合算法的有效性,对比实验结果表明提出的融合算法改善了显著性检测的效果。结论 本文所提出的显著性模型,综合了传统模型和深度学习模型的优点,使显著性检测结果更加准确。     

OPTICS聚类与目标区域概率模型的多运动目标跟踪

目的 针对多运动目标在移动背景情况下跟踪性能下降和准确度不高的问题,本文提出了一种基于OPTICS聚类与目标区域概率模型的方法。方法 首先引入了Harris-Sift特征点检测,完成相邻帧特征点匹配,提高了特征点跟踪精度和鲁棒性;再根据各运动目标与背景运动向量不同这一点,引入了改进后的OPTICS加注算法,在构建的光流图上聚类,从而准确的分离出背景,得到各运动目标的估计区域;对每个运动目标建立一个独立的目标区域概率模型(OPM),随着检测帧数的迭代更新,以得到运动目标的准确区域。结果 多运动目标在移动背景情况下跟踪性能下降和准确度不高的问题通过本文方法得到了很好地解决,Harris-Sift特征点提取、匹配时间仅为Sift特征的17%。在室外复杂环境下,本文方法的平均准确率比传统背景补偿方法高出14%,本文方法能从移动背景中准确分离出运动目标。结论 实验结果表明,该算法能满足实时要求,能够准确分离出运动目标区域和背景区域,且对相机运动、旋转,场景亮度变化等影响因素具有较强的鲁棒性。     

Determining the Castability of Simple Polyhedra

A polyhedron P is castable if its boundary can be partitioned by a plane into two polyhedral terrains. Castable polyhedra can be manufactured easily using two cast parts, where each cast part can be removed from the object without breaking the cast part or the object. If we assume that the cast parts are each removed by a single translation, it is shown that for a simple polyhedron with n vertices, castability can be decided in time and linear space using a simple algorithm. A more complicated algorithm solves the problem in time and space, for any fixed ε > 0. In the case where the cast parts are to be removed in opposite directions, a simple O(n ² )-time algorithm is presented. Finally, if the object is a convex polyhedron and the cast parts are to be removed in opposite directions, a simple algorithm is presented. Received June 1, 1994; revised May 25, 1995.     

Hierarchical channel router

The channel routing problem is a special case of the wire routing problem when interconnections have to be performed within a rectangular strip having no obstructions between terminals located on opposite sides of the rectangle.We present here a new channel routing algorithm, based on reduction of the problem to the case of a ($\text{2 × n}$) grid and on consistent utilization of a ‘divide-and-conquer’ approach. For the current implementation of the algorithm, the running time is proportional to $\text{N}{}^1\text{n}\text{log}\text{(m)}$ where $\text{N}$ is the number of nets, $\text{n}$ the length of the channel (number of columns) and $\text{m}$ the width of the channel (number of tracks).Traditional technological restrictions are assumed, ie, net terminals are located on vertical grid lines, two wiring layers are available for interconnections. One layer is used exclusively for vertical segments and another for horizontal. Vias are introduced for each layer change.This algorithm consistently outperforms several known routers in quality of wiring. We tested the algorithm on several benchmark problems. One of them, Deutsch's ‘difficult example’, was routed with only 19 horizontal wiring tracks (the absolute minimum for this case), whereas all other known routers required 20 or more tracks.     

Airborne hyperspectral mapping of trees in an urban area

The management of trees in urban areas requires accurate maps, which are difficult to build in the dense patchwork of numerous material properties. Remote sensing is a useful technique that measures the response of all vegetation occurrences, including trees, when high spatial resolution is available. The continuous narrow spectral bands of hyperspectral images enable the detection of the oxygen and water content, which ensures a perfect correction of the atmospheric effect. When calibrated, sunlight reflectance images can be used to map surface chemical compositions by the detection of diagnostic and sharp absorption features. In the visible and near- infrared, the vegetation is detected by chlorophyll-a absorption features that are characteristic of the pigment content. The reflectance intensities due to the texture of leaves occur between 450 and 920 nm while the water content imprint is detectable beyond 920 nm. The sharp spectral feature intensities of the main associated pigments, not only chlorophylls, are well quantified by indices measuring a normalized difference of reflectance in a spectral interval between two bounding wavelengths. A regression line calculated on all bands within that interval ensures a low sensitivity of the indices to the smaller variations in reflectance intensity. Such unbiased indices may be combined, using successive index thresholds deduced from a training spectral library, to divide the spectra into subsets, minimizing the confusion between the numerous vegetation types with almost identical compositions. Therefore, for each subset of the spectra, a classic spectral angle mapping (SAM) method can be used on the corresponding sub-selection of the spectral library to measure angles at full spectral resolution and map tree types with great accuracy, grouped according to their spectral similarity. In this study, chemical and physical information is carefully separated. The tree crown physical properties are studied by comparing the local juxtaposition of pixel sets to a characteristic texture identifiable by image segmentation into objects. Instead of looking for objects in the reflectance image or any statistical compression of its information, a 25 channel co-image, built from 11 information layers of chemical sharp spectral feature indices and 14 information layers of SAM indices matching a spectral library of reference vegetation groups, was used. Tree canopies also present wide internal variations due to (i) a complex mixture with a background in the case of sparse foliage, or (ii) pigment content adaptation to light exposure intensity from one side to another. Both effects are minimized by using the mean spectrum of each object, assuming that less significant spectra, being at plus or minus one or two standard deviations from an object mean spectrum, would be less affected by anomalous pixel data. Thus, two overlapping hierarchic layers at the pixel scale and the object scale are available to describe the main chemistry or pigment content that identifies the vegetation types. The final classification is given by the upper layer at the object scale but in such an organization, the pixel scale layers can be used to analyse the data further and reorder them to obtain other parameters potentially useful for management purposes.     

基于偏最小二乘分析的双模粒子滤波目标跟踪

针对在复杂背景下,基于主成分分析(PCA)的目标跟踪方法准确率较低的问题,使用偏最小二乘分析,提出一种双模粒子滤波的跟踪算法.首先采用偏最小二乘分析对目标区域建模,作为观测模型;然后利用仿射变换描述目标的形变过程,分别在李群及其切向量空间上建立双模的动态模型;最后结合特征空间更新策略,使用粒子滤波实现目标跟踪.实验表明,所提出的算法能够有效滤除背景噪声,跟踪结果稳定且准确.     

融合分割先验的多图像目标语义分割

目的 在序列图像或多视角图像的目标分割中，传统的协同分割算法对复杂的多图像分割鲁棒性不强，而现有的深度学习算法在前景和背景存在较大歧义时容易导致目标分割错误和分割不一致。为此，提出一种基于深度特征的融合分割先验的多图像分割算法。方法 首先，为了使模型更好地学习复杂场景下多视角图像的细节特征，通过融合浅层网络高分辨率的细节特征来改进PSPNet-50网络模型，减小随着网络的加深导致空间信息的丢失对分割边缘细节的影响。然后通过交互分割算法获取一至两幅图像的分割先验，将少量分割先验融合到新的模型中，通过网络的再学习来解决前景/背景的分割歧义以及多图像的分割一致性。最后通过构建全连接条件随机场模型，将深度卷积神经网络的识别能力和全连接条件随机场优化的定位精度耦合在一起，更好地处理边界定位问题。结果 本文采用公共数据集的多图像集进行了分割测试。实验结果表明本文算法不但可以更好地分割出经过大量数据预训练过的目标类，而且对于没有预训练过的目标类，也能有效避免歧义的区域分割。本文算法不论是对前景与背景区别明显的较简单图像集，还是对前景与背景颜色相似的较复杂图像集，平均像素准确度（PA）和交并比（IOU）均大于95%。结论 本文算法对各种场景的多图像分割都具有较强的鲁棒性，同时通过融入少量先验，使模型更有效地区分目标与背景，获得了分割目标的一致性。     

From the zonotope construction to the Minkowski addition of convex polytopes

A zonotope is the Minkowski addition of line segments in R^d. The zonotope construction problem is to list all extreme points of a zonotope given by its line segments. By duality, it is equivalent to the arrangement construction problem—that is, to generate all regions of an arrangement of hyperplanes.By replacing line segments with convex V-polytopes, we obtain a natural generalization of the zonotope construction problem: the construction of the Minkowski addition of k polytopes. Gritzmann and Sturmfels studied this general problem in various aspects and presented polynomial algorithms for the problem when one of the parameters k or d is fixed. The main objective of the present work is to introduce an efficient algorithm for variable d and k. Here we call an algorithm efficient or polynomial if it runs in time bounded by a polynomial function of both the input size and the output size. The algorithm is a natural extension of a known algorithm for the zonotope construction, based on linear programming and reverse search. It is compact, highly parallelizable and very easy to implement.This work has been motivated by the use of polyhedral computation for optimal tolerance determination in mechanical engineering.     

定位图像匹配尺度与区域的摄像机位姿实时跟踪

目的 提出一种定位图像匹配尺度及区域的有效算法,通过实现当前屏幕图像特征点与模板图像中对应尺度下部分区域中的特征点匹配,实现摄像机对模板图像的实时跟踪,解决3维跟踪算法中匹配精度与效率问题。方法 在预处理阶段,算法对模板图像建立多尺度表示,各尺度下的图像进行区域划分,在每个区域内采用ORB（oriented FAST and rotated BRIEF）方法提取特征点并生成描述子,由此构建图像特征点的分级分区管理模式。在实时跟踪阶段,对于当前摄像机获得的图像,首先定位该图像所对应的尺度范围,在相应尺度范围内确定与当前图像重叠度大的图像区域,然后将当前图像与模板图像对应的尺度与区域中的特征点集进行匹配,最后根据匹配点对计算摄像机的位姿。结果 利用公开图像数据库（stanford mobile visual search dataset）中不同分辨率的模板图像及更多图像进行实验,结果表明,本文算法性能稳定,配准误差在1个像素左右;系统运行帧率总体稳定在2030 帧/s。结论 与多种经典算法对比,新方法能够更好地定位图像匹配尺度与区域,采用这种局部特征点匹配的方法在配准精度与计算效率方面比现有方法有明显提升,并且当模板图像分辨率较高时性能更好,特别适合移动增强现实应用。     

融合多重注意力机制的人眼注视点预测

目的 经典的人眼注视点预测模型通常采用跳跃连接的方式融合高、低层次特征,容易导致不同层级之间特征的重要性难以权衡,且没有考虑人眼在观察图像时偏向中心区域的问题。对此,本文提出一种融合注意力机制的图像特征提取方法,并利用高斯学习模块对提取的特征进行优化,提高了人眼注视点预测的精度。方法 提出一种新的基于多重注意力机制(multiple attention mechanism,MAM)的人眼注视点预测模型,综合利用3种不同的注意力机制,对添加空洞卷积的ResNet-50模型提取的特征信息分别在空间、通道和层级上进行加权。该网络主要由特征提取模块、多重注意力模块和高斯学习优化模块组成。其中,空洞卷积能够有效获取不同大小的感受野信息,保证特征图分辨率大小的不变性;多重注意力模块旨在自动优化获得的低层丰富的细节信息和高层的全局语义信息,并充分提取特征图通道和空间信息,防止过度依赖模型中的高层特征;高斯学习模块用来自动选择合适的高斯模糊核来模糊显著性图像,解决人眼观察图像时的中心偏置问题。结果 在公开数据集SALICON(saliency in context)上的实验表明,提出的方法相较于同结构的SAM-Res(saliency attention modal)模型以及DINet(dilated inception network)模型在相对熵(Kullback-Leibler divergence,KLD)、sAUC(shuffled area under ROC curve)和信息增益(information gain,IG)评价标准上分别提高了33%、0.3%和6%;53%、0.5%和192%。结论 实验结果表明,提出的人眼注视点预测模型能通过加权的方式分别提取空间、通道、层之间的特征,在多数人眼注视点预测指标上超过了主流模型。