CN103366382A

CN103366382A - Active contour tracing method based on superpixel

Info

Publication number: CN103366382A
Application number: CN2013102774746A
Authority: CN
Inventors: 周雪; 邹见效; 徐红兵
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2013-07-04
Filing date: 2013-07-04
Publication date: 2013-10-23

Abstract

The invention discloses an active contour tracking method based on superpixels, which performs superpixel segmentation on a training image to obtain a training sample pool of targets and backgrounds, uses a measure learning method to obtain a projection matrix of a distance measure according to the training samples, and constructs a discriminative appearance model, perform superpixel segmentation on each frame of the test image of the sequence image, and obtain the confidence map corresponding to the test image according to the constructed discriminative apparent model, thereby obtaining the velocity field of the test image, and substituting the velocity field into the evolution equation of the level set method , to get the contour tracking result of the test image. Compared with the prior art, the invention improves the contour evolution efficiency of each frame test image, and simultaneously improves the tracking accuracy and tracking efficiency of sequence images.

Description

A Superpixel-Based Active Contour Tracking Method

技术领域 technical field

本发明属于计算机视觉技术领域，更为具体地讲，涉及一种基于超像素的主动轮廓跟踪方法。 The invention belongs to the technical field of computer vision, and more specifically relates to an active contour tracking method based on superpixels. the

背景技术 Background technique

视觉监控场景下的目标跟踪是通过对摄像机所拍摄的视频图像序列进行处理，检测、定位和跟踪其中运动的目标。由于轮廓特征能够很好地描述目标的形状信息，而这些形状信息对于后续高层的行为理解和识别提供了便利，并且相对于静止摄像机，主动轮廓跟踪方法能更好适用于移动摄像机下检测、定位和跟踪运动目标。因此近年来主动轮廓跟踪方法已逐渐成为当前学术研究的前沿和热点。在计算机视觉和模式识别领域中的国际顶级刊物TPAMI、IJCV和会议ICCV、CVPR上，主动轮廓跟踪都占据了一定的篇幅和比重。作为一个融合了计算机视觉、图像处理、模式识别、机器学习、统计分析和随机过程等多学科交叉的前沿性研究方向，其研究成果在智能视觉监控、运动分析、人机交互、智能导航、视频检索等领域具有广泛的应用潜力。 The target tracking in the visual surveillance scene is to detect, locate and track the moving target by processing the video image sequence captured by the camera. Since the contour features can well describe the shape information of the target, and these shape information facilitates the understanding and recognition of the subsequent high-level behavior, and compared with the static camera, the active contour tracking method is more suitable for the detection and positioning of the moving camera. and track sports goals. Therefore, the active contour tracking method has gradually become the frontier and hot spot of current academic research in recent years. In the top international journals TPAMI, IJCV and conferences ICCV and CVPR in the field of computer vision and pattern recognition, active contour tracking occupies a certain amount of space and proportion. As a cutting-edge research direction that integrates computer vision, image processing, pattern recognition, machine learning, statistical analysis and stochastic process, its research results are in intelligent visual monitoring, motion analysis, human-computer interaction, intelligent navigation, video Retrieval and other fields have broad application potential. the

主动轮廓跟踪方法的核心思想是根据实际问题建立一个关于轮廓的能量函数，采用变分方法最小化该能量函数，最终得到轮廓的进化方程，即将初始轮廓按照能量的负梯度方向进化，直到收敛到目标的边缘处。根据轮廓的描述方式以及所考虑的图像信息的不同，主动轮廓跟踪方法大致可以分为基于边缘和基于区域的两大类。 The core idea of the active contour tracking method is to establish an energy function about the contour according to the actual problem, use the variational method to minimize the energy function, and finally obtain the evolution equation of the contour, that is, the initial contour evolves in the direction of the negative gradient of the energy until it converges to edge of the target. According to the contour description method and the image information considered, active contour tracking methods can be roughly divided into two categories: edge-based and region-based. the

基于边缘的主动轮廓跟踪方法以Snakes模型为代表。Snakes模型采用的是参数化的轮廓描述方法，即将轮廓C直接显示地描述成参数s和时间t的一个函数。该模型考虑了轮廓边缘处图像的梯度信息，当轮廓位置越接近物体边缘处，则所建立的能量函数最小。该模型简单易用，但存在一系列的缺点：对轮廓的初始化较敏感，对于自相交和重叠等情况需要重新参数化轮廓，并且该模型不能处理拓扑变化并且具有不稳定的数值解。 The edge-based active contour tracking method is represented by the Snakes model. The Snakes model adopts a parametric contour description method, that is, the contour C is directly and explicitly described as a function of the parameter s and time t. The model considers the gradient information of the image at the edge of the contour, and the energy function established is the smallest when the contour position is closer to the edge of the object. The model is simple and easy to use, but there are a series of shortcomings: it is sensitive to the initialization of the contour, it needs to re-parameterize the contour for self-intersection and overlap, and the model cannot handle topology changes and has an unstable numerical solution. the

目前一种基于区域的主动轮廓跟踪方法——采用隐式轮廓描述的Level Sets (水平集)方法逐渐受到广泛关注。Level Sets方法是用一个n+1维的Level Sets函数的零值来表达一个n维的轮廓。常用的Level Sets函数为带符号的距离函数。基于区域的主动轮廓跟踪方法的优势是可以考虑图像的区域信息，而不仅仅局限在轮廓周围。常用的测度是一些统计特征，比如：均值、方差、纹理或所考虑区域的直方图等。Yilmaz等人对区域的颜色和纹理特征分别用核密度估计和Gabor小波建模，每个像素的Level Sets进化速度函数取决于它邻域内所有像素属于目标和背景的相似程度。具体算法可参考文献：A.Yilmaz,X.Li and M.Shah.Contour-based Object Tracking with Occlusion Handling in Video Acquired Using Mobile Cameras[J],IEEE Trans.on Pattern Analysis and Machine Intelligence,2004,26(11):1531-1536。Sun等人提出了一个有监督的Level Sets跟踪方法，该方法基于在线boosting建立单个像素的判别式模型来构建Level Sets的能量函数。具体算法可参考文献：X.Sun,H.X.Yao and S.P.Zhang.A Novel Supervised Level Set Method for Non-Rigid Object Tracking[C],IEEE Conference on Computer Vision and Pattern Recognition,2011,3393-3400。以上基于区域的主动轮廓跟踪方法在构建能量函数时，无论采用产生式模型或判别式模型，都是采用底层像素特征，并将其作为轮廓进化的基本单元，因此容易导致轮廓进化受噪音干扰、效率低等问题。 At present, a region-based active contour tracking method - the Level Sets method using implicit contour description has gradually attracted widespread attention. The Level Sets method uses the zero value of an n+1-dimensional Level Sets function to express an n-dimensional profile. The commonly used Level Sets function is a signed distance function. The advantage of the region-based active contour tracking method is that it can consider the region information of the image, not just limited to the contour. Commonly used measures are some statistical features, such as: mean, variance, texture or histogram of the area under consideration. Yilmaz et al. used kernel density estimation and Gabor wavelet to model the color and texture features of the region, and the Level Sets evolution speed function of each pixel depends on the similarity between all pixels in its neighborhood belonging to the target and the background. For the specific algorithm, please refer to the literature: A.Yilmaz, X.Li and M.Shah. Contour-based Object Tracking with Occlusion Handling in Video Acquired Using Mobile Cameras[J], IEEE Trans.on Pattern Analysis and Machine Intelligence, 2004, 26( 11): 1531-1536. Sun et al. proposed a supervised Level Sets tracking method, which is based on online boosting to establish a discriminative model of a single pixel to construct the energy function of Level Sets. For the specific algorithm, please refer to the literature: X.Sun, H.X.Yao and S.P.Zhang. A Novel Supervised Level Set Method for Non-Rigid Object Tracking[C], IEEE Conference on Computer Vision and Pattern Recognition, 2011, 3393-3400. The above region-based active contour tracking methods use the underlying pixel features as the basic unit of contour evolution no matter whether they use the generative model or the discriminative model when constructing the energy function, so it is easy to cause the contour evolution to be disturbed by noise, Low efficiency and other issues. the

近年来，由于富含语义信息以及灵活的处理方式，超像素(Superpixel)已经作为一种非常有效的图像描述的工具，被广泛应用于图像分割和目标识别领域。它将图像划分为像素的集合，每个集合中的像素具有某种相似的特性，比如颜色、亮度或纹理相似等。超像素具有计算效率高、富含语义、保持边界等优点。因此将这些超像素作为图像处理的基本单元，进行后续的建模和挖掘，比只考虑底层视觉特征像素更为有效。在文献S.Wang,H.C.Lu,F.Yang and M.H.Yang.Superpixel Tracking[C].IEEE International Conference on Computer Vision.2011.1323-1330.中，Wang等人提出了一种基于超像素的bounding box跟踪方法，利用均值漂移算法来建立判别式表观模型，判断每个超像素属于目标还是背景。在判别式表观模型中，所采用的距离测度对其性能起着极其重要的作用。目前常用的欧式距离由于忽略了数据的统计规律，对所有情况都采用同一测度，缺乏针对性，因此比较难获得满意的效果。特别是在大多数实际跟踪场景中，目标和背景都具有多种颜色或纹理的表观形式，即使来自于同一类别，根据欧式距离测度计算得到的差异性还是很明显的，因此采用欧式距离进行度量并不可靠。事实上，符合数据分布规律的距离测度可以预先从标定好的数据中学习得到。 In recent years, due to its rich semantic information and flexible processing methods, superpixels have been widely used in the fields of image segmentation and object recognition as a very effective image description tool. It divides the image into collections of pixels, and the pixels in each collection have some similar characteristics, such as similar color, brightness or texture. Superpixels have the advantages of high computational efficiency, rich semantics, and boundary preservation. Therefore, it is more effective to use these superpixels as the basic unit of image processing for subsequent modeling and mining than to only consider the underlying visual feature pixels. In the literature S.Wang, H.C.Lu, F.Yang and M.H.Yang.Superpixel Tracking[C].IEEE International Conference on Computer Vision.2011.1323-1330. In, Wang et al proposed a superpixel-based bounding box tracking method , using the mean shift algorithm to establish a discriminative appearance model to judge whether each superpixel belongs to the target or the background. In a discriminative appearance model, the adopted distance measure plays an extremely important role in its performance. The currently commonly used Euclidean distance ignores the statistical laws of the data and uses the same measure for all situations, which is not pertinent, so it is difficult to obtain satisfactory results. Especially in most actual tracking scenarios, the target and the background have multiple colors or textures. Even if they come from the same category, the difference calculated according to the Euclidean distance measure is still obvious, so the Euclidean distance is used for Metrics are not reliable. In fact, the distance measure that conforms to the data distribution law can be learned from the calibrated data in advance. the

需要强调的是，虽然基于超像素的图像描述已经成功应用于目标跟踪领域，但仅限于传统的bounding box跟踪，如何将其引入主动轮廓跟踪框架，建立更加有效的表观模型仍然存在诸多难点，因此仍是一个挑战。 It should be emphasized that although superpixel-based image description has been successfully applied to the field of object tracking, it is limited to traditional bounding box tracking. There are still many difficulties in how to introduce it into the framework of active contour tracking and establish a more effective appearance model. So it remains a challenge. the

发明内容 Contents of the invention

本发明的目的在于克服现有技术的不足，提供一种基于超像素的主动轮廓跟踪方法，采用测度学习方法建立以超像素为基本单元的判别式表观模型，提高轮廓跟踪的准确性和鲁棒性。 The purpose of the present invention is to overcome the deficiencies of the prior art, to provide a superpixel-based active contour tracking method, using the measure learning method to establish a discriminative appearance model with superpixels as the basic unit, to improve the accuracy and robustness of contour tracking. Stickiness. the

为实现上述发明目的，本发明基于超像素的主动轮廓跟踪方法，其特征在于，包括以下步骤： In order to achieve the above-mentioned purpose of the invention, the present invention is based on the superpixel active contour tracking method, which is characterized in that it comprises the following steps:

S1：将训练图像分为目标和背景两部分，进行超像素分割，提取每个超像素的特征向量，构建目标训练样本池T_obj和背景训练样本池T_bac； S1: Divide the training image into two parts, the target and the background, perform superpixel segmentation, extract the feature vector of each superpixel, and construct the target training sample pool T _obj and the background training sample pool T _bac ;

S2：根据训练样本采用测度学习方法得到距离测度的投影矩阵L，投影矩阵L每隔m,m≥1帧测试图像更新一次； S2: According to the training sample, the projection matrix L of the distance measure is obtained by using the measure learning method, and the projection matrix L is updated every m, m≥1 frame test image;

S3：根据目标训练样本池T_obj、背景训练样本池T_bac和距离测度的投影矩阵L，构建基于超像素的判别式表观模型，其中每个超像素的置信分数的计算公式为： S3: According to the target training sample pool T _obj , the background training sample pool T _bac and the projection matrix L of the distance measure, construct a discriminative appearance model based on superpixels, where the confidence score of each superpixel The calculation formula is:

${S S}_{c c}^{sp sp} = = \frac{11 - - P P ((sp sp | | bac bac)) / / P P ((sp sp | | obj obj))}{11 + + P P ((sp sp | | bac bac)) / / P P ((sp sp | | obj obj))}$

其中P(sp|obj)和P(sp|bac)分别表示超像素sp属于目标类obj和背景类bac的似然概率，采用非参数的核密度估计方法得到； Among them, P(sp|obj) and P(sp|bac) represent the likelihood probability that the superpixel sp belongs to the object class obj and the background class bac respectively, and are obtained by using a non-parametric kernel density estimation method;

S4：在当前帧测试图像中选定包括目标在内的一个局部区域，对该局部区域进行超像素分割，超像素数量记为N，提取得到每个超像素sp_k,1≤k≤N的特征向量f_k；根据步骤S3中的置信分数计算公式计算每个超像素的置信分数

得到测试图像的置信图； S4: Select a local area including the target in the current frame test image, perform superpixel segmentation on the local area, record the number of superpixels as N, and extract each superpixel sp _k , 1≤k≤N Feature vector f _k ; calculate the confidence score of each superpixel according to the confidence score calculation formula in step S3

Get the confidence map of the test image;

S5：根据步骤S4中得到的置信图构建测试图像的速度场

S5: Construct the velocity field of the test image according to the confidence map obtained in step S4

${F f}_{data data}^{i i,, j j} = = \{\begin{matrix} {S S}_{c c}^{{sp sp}_{k k}} & if if & {x x}_{i i,, j j} &Element; &Element; {{{sp sp}_{k k}}}_{k k = = 11}^{N N} \\ - - 11 & if if & {x x}_{i i,, j j} &NotElement; &NotElement; {{{sp sp}_{k k}}}_{k k = = 11}^{N N} \end{matrix}$

其中，(i,j)表示测试图像中像素的坐标； Among them, (i, j) represents the coordinates of the pixels in the test image;

S6：将步骤S5中得到的测试图像的速度场

代入水平集方法的进化方程，将上一帧测试图像的轮廓跟踪结果作为初始值进行轮廓进化，得到目标的轮廓跟踪结果； S6: the velocity field of the test image obtained in step S5

Substitute into the evolution equation of the level set method, use the contour tracking result of the previous frame test image as the initial value for contour evolution, and obtain the contour tracking result of the target;

S7：根据步骤S6中的得到的轮廓跟踪结果，将目标和背景的超像素分别放入对应的训练样本池中对训练样本池进行更新，返回步骤S2对序列图像中下一帧测试图像进行轮廓跟踪。 S7: According to the contour tracking results obtained in step S6, put the superpixels of the target and the background into the corresponding training sample pools to update the training sample pools, and return to step S2 to contour the next frame of the test image in the sequence image track. the

其中，步骤S4中局部区域的选定方法为：在第一帧测试图像选定局部区域时，手动指定目标的初始轮廓，根据初始轮廓确定局部区域；后续每一帧测试图像根据上一帧测试图像的轮廓跟踪结果确定局部区域。 Wherein, the selection method of the local area in step S4 is: when selecting the local area in the first frame test image, manually specify the initial outline of the target, and determine the local area according to the initial outline; The contour tracking results of the image determine the local area. the

其中，步骤S6中进化方程为： Wherein, in step S6, evolutionary equation is:

$\frac{{Φ Φ}_{t t} - - {Φ Φ}_{t t - - 11}}{Δt Δt} + + (({F f}_{data data}^{i i,, j j} + + {F f}_{curv curv})) \cdot &Center Dot; | | {&dtri; &dtri; Φ Φ}_{t t - - 11} | | = = 00$

其中Φ_t是第t次迭代的水平集函数，Φ_t-1是第t-1次迭代的水平集函数，水平集函数的初始值Φ₀是上一帧测试图像的轮廓跟踪结果的水平集函数，Δt是预设的迭代步长，F_curv＝εκ是只跟轮廓曲率κ相关的内部进化速度，ε是预设的常数，

是Φ_t-1的梯度范数。 where _Φt is the level set function of the tth iteration, Φt _-1 is the level set function of the t-1th iteration, and the initial value of the level set function, _Φ0, is the level set of the contour tracking result of the previous frame test image function, Δt is the preset iteration step size, F _curv = εκ is the internal evolution rate only related to the contour curvature κ, ε is a preset constant,

is the gradient norm of Φt _-1 .

其中，步骤S7中训练样本池在更新时采用队列方式进行更新，新增样本排在队列末端，当样本数量超过预设的队列长度，删除队列前端的旧样本。 Wherein, in step S7, the training sample pool is updated in a queue mode, and the newly added samples are arranged at the end of the queue. When the number of samples exceeds the preset queue length, the old samples at the front of the queue are deleted. the

本发明基于超像素的主动轮廓跟踪方法，对训练图像进行超像素分割得到目标和背景的训练样本池，根据训练样本采用测度学习方法得到距离测度的投影矩阵，构建判别式表观模型，将序列图像的每帧测试图像进行超像素分割，根据构建好的判别式表观模型得到测试图像对应的置信图，从而得到测试图像的速度场，将速度场代入水平集方法的进化方程，得到测试图像的轮廓跟踪结果。 The present invention is based on the superpixel active contour tracking method, performs superpixel segmentation on the training image to obtain the training sample pool of the target and the background, uses the measurement learning method to obtain the projection matrix of the distance measure according to the training sample, constructs the discriminant appearance model, and converts the sequence Superpixel segmentation is performed on each frame of the test image of the image, and the confidence map corresponding to the test image is obtained according to the constructed discriminant apparent model, so as to obtain the velocity field of the test image, and the velocity field is substituted into the evolution equation of the level set method to obtain the test image contour tracking results. the

本发明基于超像素的主动轮廓跟踪方法具有以下有益效果： The active contour tracking method based on superpixels of the present invention has the following beneficial effects:

①、建立以超像素为图像描述的基本单元，提取具备一定语义描述的中层视觉特征，该过程能为后续的建模和轮廓进化提供便利； ①. Establish superpixels as the basic unit of image description, and extract middle-level visual features with certain semantic descriptions. This process can facilitate subsequent modeling and contour evolution;

②、距离测度的投影矩阵采用测度学习方法得到，将原始特征空间投影到另一个更能反映数据本质特性的空间，在该空间下计算出的距离测度更为真实可靠； ②. The projection matrix of the distance measure is obtained by using the measure learning method, and the original feature space is projected to another space that can better reflect the essential characteristics of the data. The distance measure calculated in this space is more real and reliable;

③、提出以超像素为基本单元的轮廓进化方法。由于超像素内所有像素具有表观相似性，直接反映到Level Sets速度场中，即同一个超像素内所有像素在进化速度的方向和大小上保持一致，比直接考虑单个像素更能提高轮廓的进化效率。 ③. A contour evolution method with superpixels as the basic unit is proposed. Due to the apparent similarity of all pixels in a superpixel, it is directly reflected in the velocity field of Level Sets, that is, all pixels in the same superpixel are consistent in the direction and magnitude of the evolution velocity, which can improve the contour accuracy more than directly considering a single pixel. evolutionary efficiency. the

附图说明 Description of drawings

图1是本发明基于超像素的主动轮廓跟踪方法的一种具体实施方式流程图； Fig. 1 is a kind of specific embodiment flowchart of the active contour tracking method based on superpixel of the present invention;

图2是本发明基于超像素的主动轮廓跟踪方法中基于超像素的表观模型的置信图获得示例图； Fig. 2 is an example diagram of obtaining a confidence map based on a superpixel-based apparent model in the superpixel-based active contour tracking method of the present invention;

图3是本发明基于超像素和像素得到的速度场对比示意图； Fig. 3 is a comparative schematic diagram of the velocity field obtained based on superpixels and pixels in the present invention;

图4是本发明与现有技术迭代次数的对比示例图； Fig. 4 is the comparative example figure of the present invention and the number of iterations of the prior art;

图5是本发明与现有技术跟踪准确率的对比示例图。 Fig. 5 is an example graph comparing tracking accuracy between the present invention and the prior art. the

具体实施方式 Detailed ways

下面结合附图对本发明的具体实施方式进行描述，以便本领域的技术人员更好地理解本发明。需要特别提醒注意的是，在以下的描述中，当已知功能和设计的详细描述也许会淡化本发明的主要内容时，这些描述在这里将被忽略。 Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here. the

图1是本发明基于超像素的主动轮廓跟踪方法的一种具体实施方式流程图。如图1所示，本发明基于超像素的主动轮廓跟踪方法包括以下步骤： FIG. 1 is a flow chart of a specific embodiment of the superpixel-based active contour tracking method of the present invention. As shown in Figure 1, the active contour tracking method based on superpixels of the present invention comprises the following steps:

S101：构建训练样本池： S101: Build a training sample pool:

将训练图像分为目标和背景两部分，进行超像素分割，提取每个超像素的特征向量，构建目标和背景的训练样本池，其中目标训练样本池记为T_obj，背景训练样本池记为T_bac。在实际应用中，一般选用前几帧的测试图像作为训练图像。 Divide the training image into two parts, the target and the background, perform superpixel segmentation, extract the feature vector of each superpixel, and construct the training sample pool of the target and the background, where the target training sample pool is denoted as T _obj , and the background training sample pool is denoted as T _bac . In practical applications, the test images of the first few frames are generally selected as the training images.

S102：计算距离测度的投影矩阵： S102: Calculate the projection matrix of the distance measure:

根据训练样本采用测度学习方法得到距离测度的投影矩阵L，投影矩阵L每隔m,m≥1帧测试图像更新一次。 According to the training samples, the projection matrix L of the distance measure is obtained by using the measure learning method, and the projection matrix L is updated every m, m≥1 frame test images. the

由于传统欧式距离缺乏对数据统计规律的认知，不能如实反映数据的本质特征，特别是针对同一类别中有多种表观模式(多种颜色或纹理)存在的情况，基于欧式距离的判别式表观模型更难获得满意的结果。因此，在本发明引入测度学习方法来计算距离测度，以此建立多模态的判别式表观模型。测度学习实际上就是求解一个投影矩阵L，通过该投影矩阵将原始特征空间投影到另外一个更能反映数据本质特性的特征空间。于是，欧式距离就转化成马氏距离(Mahalanobis distance)。 Because the traditional Euclidean distance lacks the cognition of the statistical laws of data, it cannot truly reflect the essential characteristics of the data, especially for the situation where there are multiple appearance patterns (multiple colors or textures) in the same category, the discriminant formula based on Euclidean distance Appearance models are more difficult to obtain satisfactory results. Therefore, in the present invention, a measure learning method is introduced to calculate the distance measure, so as to establish a multimodal discriminant appearance model. Measure learning is actually to solve a projection matrix L, through which the original feature space is projected to another feature space that can better reflect the essential characteristics of the data. Thus, the Euclidean distance is transformed into Mahalanobis distance. the

本实施方式中采用LMNN(Large Margin Nearest Neighbor，大边缘最近邻)测度学习方法，该距离测度学习能从带有标签的数据集中学习一个投影，将原始特征空间投影到一个新的特征空间，最终目标是使得投影后样本只和属于同一类别中的有限个样本保持近邻关系，而和不同标签的样本至少保持一个单位长度的距离(即margin，边缘)。在本发明中，即是从带有“目标”和“背景”两个标签的训练样本池中学习一个投影，更好地判别测试图像中每个超像素。 In this embodiment, the LMNN (Large Margin Nearest Neighbor) measure learning method is adopted. The distance measure learning can learn a projection from the labeled data set, project the original feature space to a new feature space, and finally The goal is to make the projected samples only maintain a neighbor relationship with a limited number of samples belonging to the same category, and keep at least a unit length distance (ie margin, edge) with samples of different labels. In the present invention, a projection is learned from the training sample pool with two labels of "target" and "background", so as to better distinguish each superpixel in the test image. the

在本发明中，由于训练样本池会根据每帧测试图像的轮廓跟踪结果进行更新，考虑到计算效率，可以不用每次训练样本池更新的时候都重新计算一次距离测度，而是采用每隔几帧测试图像更新一次距离测度的策略。每次更新时，将上一次得到的投影矩阵作为更新的初始值输入，第一次计算投影矩阵采用的初始值为单位矩阵。 In the present invention, since the training sample pool will be updated according to the contour tracking results of each frame of the test image, considering the calculation efficiency, it is not necessary to recalculate the distance measure every time the training sample pool is updated, but to use every few A strategy for updating the distance measure once for the frame test image. For each update, the projection matrix obtained last time is input as the initial value of the update, and the initial value used for the first calculation of the projection matrix is the identity matrix. the

S103：构建基于超像素的判别式表观模型： S103: Construct a discriminative appearance model based on superpixels:

根据目标训练样本池T_obj、背景训练样本池T_bac和投影矩阵L，构建测试图像基于超像素的判别式表观模型。由于投影矩阵L是通过测度学习方法得到的，因此本发明中得到的判别式表观模型更能体现图像的特征。 According to the target training sample pool T _obj , the background training sample pool T _bac and the projection matrix L, a superpixel-based discriminative appearance model of the test image is constructed. Since the projection matrix L is obtained through a measure learning method, the discriminative appearance model obtained in the present invention can better reflect the characteristics of the image.

在判别式表观模型中，针对每一个超像素sp，需要给出一个置信分数(Confidence Score)来反映其与目标或背景的相似程度。本发明中，定义的置信分数

计算公式为： In the discriminative appearance model, for each superpixel sp, a confidence score (Confidence Score) needs to be given to reflect its similarity with the target or background. In the present invention, the defined confidence score

The calculation formula is:

${S S}_{c c}^{sp sp} = = \frac{11 - - P P ((sp sp | | bac bac)) / / P P ((sp sp | | obj obj))}{11 + + P P ((sp sp | | bac bac)) / / P P ((sp sp | | obj obj))} - - - - - - ((11))$

其中P(sp|obj)和P(sp|bac)分别表示超像素sp属于目标类obj和背景类bac的似然概率，采用非参数的核密度估计方法得到。 Among them, P(sp|obj) and P(sp|bac) represent the likelihood probability that the superpixel sp belongs to the object class obj and the background class bac, respectively, and are obtained by using a non-parametric kernel density estimation method. the

置信分数

的取值范围介于-1和1之间，具有如下对称的判别属性： confidence score

The value range of is between -1 and 1, and has the following symmetric discriminant properties:

$\{\begin{matrix} - - 11 < < {S S}_{c c}^{sp sp} < < 00 & P P ((sp sp | | bac bac)) > > P P ((sp sp | | obj obj)) \\ 00 & P P ((sp sp | | bac bac)) = = P P ((sp sp | | obj obj)) \\ 00 < < {S S}_{c c}^{sp sp} < < 11 & P P ((sp sp | | bac bac)) < < P P ((sp sp | | obj obj)) \end{matrix} - - - - - - ((22))$

在本发明中，采用非参数的核密度估计方法获得似然概率P(sp|obj)和P(sp|bac)，即根据超像素sp离训练样本池中的其他样本的距离来近似逼近似然概率P(sp|obj)和P(sp|bac)。本实施方式中，选用高斯核函数，并采用K近邻假设，似然概率比P(sp|bac)P(sp|obj)可以由下式近似逼近： In the present invention, the likelihood probabilities P(sp|obj) and P(sp|bac) are obtained by using a non-parametric kernel density estimation method, that is, to approximate the approximation according to the distance of the superpixel sp from other samples in the training sample pool Natural probabilities P(sp|obj) and P(sp|bac). In this embodiment, the Gaussian kernel function is selected, and the K-nearest neighbor assumption is adopted, and the likelihood probability ratio P(sp|bac)P(sp|obj) can be approximated by the following formula:

$\frac{P P ((sp sp | | bac bac))}{P P ((sp sp | | obj obj))} = = \frac{\frac{11}{| | {T T}_{bac bac} | |} {Σ Σ}_{n no = = 11}^{| | {T T}_{bac bac} | |} exp exp ((- - \frac{{D D.}^{22} (({f f}_{sp sp},, {f f}_{n no}))}{{22 σ σ}^{22}}))}{\frac{11}{| | {T T}_{obj obj} | |} {Σ Σ}_{n no = = 11}^{| | {T T}_{obj obj} | |} exp exp ((- - \frac{{D D.}^{22} (({f f}_{sp sp},, {f f}_{n no}))}{{22 σ σ}^{22}}))} \approx \approx \frac{\frac{11}{| | {T T}_{bac bac}^{NN NN} | |} {Σ Σ}_{j j = = 11}^{| | {T T}_{bac bac}^{NN NN} | |} exp exp ((- - \frac{{D D.}^{22} (({f f}_{sp sp},, {f f}_{n no}^{NN NN}))}{{22 σ σ}^{22}}))}{\frac{11}{| | {T T}_{obj obj}^{NN NN} | |} {Σ Σ}_{j j = = 11}^{| | {T T}_{obj obj}^{NN NN} | |} exp exp ((- - \frac{{D D.}^{22} (({f f}_{sp sp},, {f f}_{n no}^{NN NN}))}{{22 σ σ}^{22}}))} - - - - - - ((33))$

其中

和表示来自目标训练样本池T_obj和背景训练样本池T_bac的子集，分别包括的是超像素sp_k的前K个近邻样本，符号|·|代表训练样本池中的样本个数。D(f_sp,f_n)是超像素特征向量f_sp与样本特征向量f_n之间的距离测度，σ是预设的计算参数。 in

and Indicates the subsets from the target training sample pool T _obj and the background training sample pool T _bac , respectively including the first K neighbor samples of the superpixel sp _k , and the symbol |·| represents the number of samples in the training sample pool. D(f _sp , f _n ) is the distance measure between the superpixel feature vector f _sp and the sample feature vector f _n , and σ is a preset calculation parameter.

本发明中距离测度D(f_sp,f_n)为马氏距离，计算公式为： In the present invention, the distance measurement D (f _sp , f _n ) is the Mahalanobis distance, and the calculation formula is:

D(f_sp,f_n)＝||L(f_sp-f_n)||²＝(f_sp-f_n)^TM(f_sp-f_n) （4） D(f _sp ,f _n )＝||L(f _sp -f _n )|| ² ＝(f _sp -f _n ) ^T M(f _sp -f _n ) (4)

其中，M＝L^TL。 Wherein, M=L ^T L.

S104：对测试图像进行超像素分割和特征提取，得到置信图： S104: Perform superpixel segmentation and feature extraction on the test image to obtain a confidence map:

本步骤主要是对测试图像进行预处理，首先在图像中选定包括目标在内的一个局部区域，对该局部区域进行超像素分割，超像素数量记为N，提取得到每个超像素sp_k,1≤k≤N的特征向量f_k。其具体步骤包括： This step is mainly to preprocess the test image. First, a local area including the target is selected in the image, and superpixel segmentation is performed on the local area. The number of superpixels is recorded as N, and each superpixel sp _k is extracted , the eigenvector f _k of 1≤k≤N. Its specific steps include:

3.1、选定目标周围的一个局部区域。 3.1. Select a local area around the target. the

在第一帧测试图像选定局部区域时，可以手动指定目标的初始轮廓，根据初始轮廓确定局部区域，即该局部区域将初始轮廓包括在内，该局部区域采用现有技术可以快速得到，一般是通过对初始轮廓进行一定比例的扩展得到。在后续每一帧测试图像选择目标的局部区域时，可以采用上一帧的轮廓跟踪结果作为局部区域选择依据。 When the local area is selected in the first frame of the test image, the initial contour of the target can be manually specified, and the local area is determined according to the initial contour, that is, the local area includes the initial contour, and the local area can be obtained quickly by using existing technologies. It is obtained by expanding the initial contour by a certain proportion. When selecting the local area of the target in each subsequent frame of the test image, the contour tracking result of the previous frame can be used as the basis for selecting the local area. the

3.2、对局部区域进行超像素分割，超像素个数记为N。 3.2. Perform superpixel segmentation on the local area, and the number of superpixels is denoted as N. the

本实施方式中，采用的是SLIC(Simple Linear Iterative Clustering，简单线性迭代聚类)超像素分割方法。该方法能以较低的计算复杂度获得指定数量的规则的超像素分割。具体算法参考文献：R.Achanta,A.Shaji,K.Smith and A.Lucchi.SLIC Superpixels Compared to State-of-the-Art Superpixel Methods[J].IEEE Trans.on Pattern Analysis and Machine Intelligence,2012,34(11):2274-2282。 In this embodiment, the SLIC (Simple Linear Iterative Clustering, Simple Linear Iterative Clustering) superpixel segmentation method is adopted. This method can obtain a specified number of regular superpixel segmentations with low computational complexity. Specific algorithm references: R.Achanta, A.Shaji, K.Smith and A.Lucchi.SLIC Superpixels Compared to State-of-the-Art Superpixel Methods[J].IEEE Trans.on Pattern Analysis and Machine Intelligence,2012, 34(11):2274-2282. the

3.3、针对每个分割得到的超像素进行特征提取，得到每个超像素sp_k,1≤k≤N的特征向量f_k。 3.3. Perform feature extraction for each segmented superpixel to obtain a feature vector f _k of each superpixel sp _k , 1≤k≤N.

测试图像的特征与训练样本池中样本的特征是一致的。本实施方式中，对每个超像素统计其中所有像素的颜色和纹理特征。其中颜色特征采用HIS（Hue-Saturation-Intensity，色调-饱和度-强度）空间的归一化的颜色直方图。纹理采用LBP(Local Binary Pattern，局部二值模式)方法，在一个3×3邻域区域内，将邻域像素灰度值与中心像素进行比较，若周围像素值大于中心像素值，则对应的邻域像素点的位置被标记为1，否则为0。这样，3×3邻域内的8个点可产生8bit的无符号数，将这8位无符号数转化成对应的十进制数(0～255)，则每个像素点都可以得到一个LBP值。用一个直方图统计同一超像素内所有像素的LBP值，并进行归一化。将归一化后的颜色直方图和LBP统计直方图合并，得到超像素sp_k最终的特征向量f_k。 The characteristics of the test image are consistent with the characteristics of the samples in the training sample pool. In this embodiment, the color and texture features of all pixels in each superpixel are counted. The color feature adopts the normalized color histogram of HIS (Hue-Saturation-Intensity, Hue-Saturation-Intensity) space. The texture adopts the LBP (Local Binary Pattern, local binary pattern) method. In a 3×3 neighborhood area, the gray value of the neighborhood pixel is compared with the center pixel. If the surrounding pixel value is greater than the center pixel value, the corresponding The location of the neighboring pixel is marked as 1, otherwise it is 0. In this way, 8 points in the 3×3 neighborhood can generate 8-bit unsigned numbers, and convert these 8-bit unsigned numbers into corresponding decimal numbers (0-255), and each pixel can get an LBP value. Use a histogram to count the LBP values of all pixels in the same superpixel and normalize them. The normalized color histogram and the LBP statistical histogram are combined to obtain the final feature vector f _k of the superpixel sp _k .

3.4、将得到的每个超像素的特征向量f_k代入基于超像素的表观模型，得到测试图像的每个超像素的置信分数

得到测试图像的置信图。 3.4. Substitute the obtained feature vector f _k of each superpixel into the superpixel-based apparent model to obtain the confidence score of each superpixel of the test image

Get a confidence map of the test image.

图2是本发明基于超像素的主动轮廓跟踪方法中基于超像素的表观模型的置信图获得示例图。如图2所示，图2（a）是LMNN测度学习方法的原理示意图，(b)是目标周围的局部区域，(c)是超像素分割结果，(d)是得到的置信图。可以看出，由于采用了测度学习方法，超像素sp和它的同标签的近邻超像素聚集的更加紧密，而和不同标签的样本则保持了一个单位长度的边缘距离(Margin)。可以看出，基于测度学习的判别式表观模型，能有效提高对多模态目标的判别能力。有关LMNN测度学习方法的具体细节参考文献：K.Q.Weinberger and L.K.Saul.Distance Metric Learning for Large Margin Nearest Neighbor Classification[J].Journal of Machine Learning Research.2009.10:207-244. Fig. 2 is an example diagram of obtaining a confidence map of a superpixel-based appearance model in the superpixel-based active contour tracking method of the present invention. As shown in Figure 2, Figure 2 (a) is a schematic diagram of the principle of the LMNN measure learning method, (b) is the local area around the target, (c) is the superpixel segmentation result, and (d) is the obtained confidence map. It can be seen that due to the use of the measure learning method, the superpixel sp and its neighboring superpixels with the same label are gathered more closely, and a unit-length edge distance (Margin) is maintained with samples of different labels. It can be seen that the discriminative appearance model based on measure learning can effectively improve the discriminative ability of multi-modal targets. For the specific details of the LMNN measure learning method reference: K.Q.Weinberger and L.K.Saul.Distance Metric Learning for Large Margin Nearest Neighbor Classification[J].Journal of Machine Learning Research.2009.10:207-244.

S105：Level Sets轮廓进化，得到轮廓跟踪结果。具体步骤包括： S105: Level Sets contour evolution, obtain contour tracking results. Specific steps include:

5.1、构建速度场： 5.1. Construct velocity field:

Level Sets轮廓进化过程中，引导轮廓进化的速度场对进化的准确率和效率起着至关重要的作用。而基于判别式表观模型得到的置信图正好可以看作是该速度场，因为置信图(介于-1到1之间)中的数值符号正好对应于Level Sets依照轮廓法线进化的方向(向内或向外)，绝对值则指定了速度的大小。比如，当P(sp_k|bac)＞P(sp_k|obj)时，超像素sp_k的置信分数

的符号为负，假设此时超像素在轮廓内部，则速度项有把轮廓沿着法线方向向内推的作用，反之如果超像素在轮廓外部，则有相反向外拉的作用力。但是由于Level Sets进化的实施是基于像素的，因此需要把基于超像素的测试图像的置信图进一步扩展成像素速度场，即属于同一超像素内的所有像素，具有和超像素相同的进化速度，目标区域之外的所有像素由于属于背景，因此速度赋值为-1，从而形成整幅测试图像的速度场，用

表示，即： In the process of Level Sets contour evolution, the velocity field guiding contour evolution plays a vital role in the accuracy and efficiency of evolution. The confidence map obtained based on the discriminant appearance model can be regarded as the velocity field, because the numerical sign in the confidence map (between -1 and 1) corresponds to the evolution direction of the Level Sets according to the contour normal ( inward or outward), the absolute value specifies the magnitude of the velocity. For example, when P(sp _k |bac) > P(sp _k |obj), the confidence score of the superpixel sp _k

The sign of is negative, assuming that the superpixel is inside the contour at this time, the velocity term has the effect of pushing the contour inward along the normal direction, otherwise, if the superpixel is outside the contour, it has the opposite force of pulling outward. However, since the implementation of Level Sets evolution is based on pixels, it is necessary to further expand the confidence map of the superpixel-based test image into a pixel velocity field, that is, all pixels belonging to the same superpixel have the same evolution speed as the superpixel, All pixels outside the target area belong to the background, so the velocity is assigned as -1, thus forming the velocity field of the entire test image, using

means, that is:

${F f}_{data data}^{i i,, j j} = = \{\begin{matrix} {S S}_{c c}^{{sp sp}_{k k}} & if if & {x x}_{i i,, j j} &Element; &Element; {{{sp sp}_{k k}}}_{k k = = 11}^{N N} \\ - - 11 & if if & {x x}_{i i,, j j} &NotElement; &NotElement; {{{sp sp}_{k k}}}_{k k = = 11}^{N N} \end{matrix} - - - - - - ((55))$

其中，(i,j)表示测试图像中像素的坐标。 where (i,j) denote the coordinates of pixels in the test image. the

图3是本发明基于超像素和像素得到的速度场对比示意图。如图3所示，每一个网格点代表一个像素，图3(a)是基于超像素得到的像素速度场示意图，图3(b)是采用普通方法得到的像素速度场示意图。经过对比可以看出，属于同一超像素内的像素(同一种符号表示)在速度的大小和方向上都保持一致性，而基于像素得到的速度场在大小和方向上缺乏规律，会降低轮廓的进化效率和准确率。在本发明中，以超像素为轮廓进化的基本单元，即同一超像素内所有像素在进化速度的方向和大小上保持一致性，比直接考虑单个像素更能提高轮廓的进化效率。 Fig. 3 is a schematic diagram of the comparison of velocity fields obtained based on superpixels and pixels in the present invention. As shown in Figure 3, each grid point represents a pixel, Figure 3(a) is a schematic diagram of the pixel velocity field obtained based on superpixels, and Figure 3(b) is a schematic diagram of the pixel velocity field obtained by a common method. After comparison, it can be seen that the pixels belonging to the same superpixel (the same symbol) maintain consistency in the magnitude and direction of the velocity, while the velocity field obtained based on pixels lacks regularity in magnitude and direction, which will reduce the accuracy of the contour. Evolutionary efficiency and accuracy. In the present invention, superpixels are used as the basic unit of contour evolution, that is, all pixels in the same superpixel maintain consistency in the direction and size of evolution speed, which can improve the evolution efficiency of contours more than directly considering a single pixel. the

5.2、将速度场代入Level Sets方法的进化方程，本实施方式中采用的进化方程为： 5.2. Substituting the velocity field into the evolution equation of the Level Sets method, the evolution equation adopted in this embodiment is:

$\frac{{Φ Φ}_{t t} - - {Φ Φ}_{t t - - 11}}{Δt Δt} + + (({F f}_{data data}^{i i,, j j} + + {F f}_{curv curv})) \cdot &Center Dot; | | {&dtri; &dtri; Φ Φ}_{t t - - 11} | | = = 00 - - - - - - ((66))$

其中Φ_t是第t次迭代的Level Sets函数，Φ_t-1是第t-1次迭代的Level Sets函数，水平集函数的初始函数Φ₀是上一帧测试图像的轮廓跟踪结果的Level Sets函数。Δt是预设的迭代步长。F_curv＝εκ是只跟轮廓曲率κ相关的内部进化速度，ε是预设的常数，F_curv起着平滑轮廓的作用。

是Φ_t-1的梯度范数。 Where Φ _t is the Level Sets function of the t-th iteration, Φ _t-1 is the Level Sets function of the t-1 iteration, and the initial function Φ ₀ of the level set function is the Level Sets of the contour tracking result of the previous frame test image function. Δt is the preset iteration step size. F _curv = εκ is the internal evolution rate only related to the contour curvature κ, ε is a preset constant, and F _curv plays the role of smoothing the contour.

is the gradient norm of Φt _-1 .

在每帧测试图像进行轮廓进化时，将上一帧测试图像的轮廓跟踪结果作为初始轮廓，在公式(6)的作用下，通过不断迭代更新Φ，从而得到测试图像的轮廓跟踪结果。迭代结束条件一般采用两种方式确定：预先设定迭代次数，将最后一次的轮廓作为轮廓跟踪结果；或在每次迭代后计算Φ_t和Φ_t-1的差值，预先设定差值阈值，当Φ_t和Φ_t-1的差值小于阈值时，将此时Φ_t表示的轮廓作为轮廓跟踪结果。在第一帧图像进行轮廓进化时，其初始轮廓可以通过手动指定。 When the contour evolution of each frame of test image is performed, the contour tracking result of the previous frame of test image is used as the initial contour, and under the action of formula (6), the contour tracking result of the test image is obtained by continuously iteratively updating Φ. The iteration end condition is generally determined in two ways: preset the number of iterations, and use the last contour as the contour tracking result; or calculate the difference between Φ _t and Φ _t-1 after each iteration, and pre-set the difference threshold , when the difference between Φ _t and Φ _t-1 is less than the threshold, the contour represented by Φ _t at this time is taken as the contour tracking result. When performing contour evolution on the first frame image, its initial contour can be specified manually.

S106：在线更新训练样本池。 S106: Update the training sample pool online. the

在获得一帧测试图像的轮廓跟踪结果后，在轮廓内部（目标）和轮廓外部（背景）的超像素可以分别放到目标和背景的训练样本池中，用于下一帧测试图像的轮廓跟踪。考虑到计算复杂度以及为了防止旧样本被一次性替代掉，本实施方式训练样本池在更新时采用队列方式进行更新，新增样本排在队列末端，当样本数量超过预设的队列长度，即删除队列前端的旧样本，保持队列长度不变。返回步骤S101根据更新重新构建训练样本池。 After obtaining the contour tracking result of a test image, the superpixels inside the contour (target) and outside the contour (background) can be put into the training sample pools of the target and background, respectively, for the contour tracking of the next test image . Considering the computational complexity and in order to prevent the old samples from being replaced at one time, the training sample pool in this embodiment is updated in a queue mode when updating, and the new samples are arranged at the end of the queue. When the number of samples exceeds the preset queue length, that is Delete old samples at the front of the queue, keeping the queue length constant. Return to step S101 to rebuild the training sample pool according to the update. the

实施例 Example

为了实施本发明的具体思想，在多个视频序列上做了比较仿真实验。为便于定量地比较，定义了跟踪准确率

来反映真实标定数据C_gt与跟踪结果C_t的相似程度。 In order to implement the specific idea of the present invention, comparative simulation experiments are done on multiple video sequences. For the convenience of quantitative comparison, the tracking accuracy is defined

To reflect the similarity between the real calibration data C _gt and the tracking result C _t .

图4是本发明与现有技术进化方程迭代次数的对比示例图。如图3所示，将本发明基于超像素的主动轮廓跟踪方法，即基于超像素并考虑测度学习(SP-based with DML)与超像素未考虑测度学习(SP-based without DML)、基于像素的方法(Pixel-based)采用同样视频序列进行对比仿真。可以看出，本发明所提出的基于超像素的主动轮廓跟踪方法能更快地收敛。 Fig. 4 is an example diagram comparing the number of iterations of the evolution equation between the present invention and the prior art. As shown in Figure 3, the active contour tracking method based on superpixels of the present invention, that is, based on superpixels and considering measure learning (SP-based with DML) and superpixels without considering measure learning (SP-based without DML), based on pixel The method (Pixel-based) uses the same video sequence for comparative simulation. It can be seen that the superpixel-based active contour tracking method proposed by the present invention can converge faster. the

图5是本发明与现有技术跟踪准确率的对比示例图。如图5所示，本实施例分别对4种视频序列进行仿真，其中图5(a)是小丑鱼的视频序列，图5(b)是篮球运动员的视频序列，图5(c)是猴子的视频序列，图5(d)是单人滑雪的视频序列。采用的对比现有技术包括ADL(Adaboost-based Level Set Method),参见文献：X.Sun,H.X.Yao and S.P.Zhang.A Novel Supervised Level Set Method for Non-Rigid Object Tracking[C],IEEE Conference on Computer Vision and Pattern Recognition,2011,3393-3400)、SPDL(Superpixel Driven Level Set Method)，参见文献：X.Zhou,X.Li,T.J.Chin and D.Suter.Superpixel-Driven Level Set Tracking[C].IEEE International Conference on Image Processing.2012.409-412.)和SPT(Superpixel Tracking Method),参见文献：S.Wang,H.C.Lu,F.Yang and M.H.Yang.Superpixel Tracking[C].IEEE International Conference on Computer Vision.2011.1323-1330.)，其中ADL是基于Adaboost的像素级别的Level Sets跟踪方法；SPDL是超像素驱动的Level Sets方法，但在表观模型构建过程中没有考虑测度学习；而SPT是基于超像素的bounding box跟踪方法，该方法采用均值漂移聚类算法得到超像素的置信图，基于该置信图进行Level Sets跟踪得到轮廓结果。从图5中可以看出，本发明基于超像素的主动轮廓跟踪方法具有更高的准确率和鲁棒性。 Fig. 5 is an example graph comparing tracking accuracy between the present invention and the prior art. As shown in Figure 5, the present embodiment carries out simulation to 4 kinds of video sequences respectively, and wherein Fig. 5 (a) is the video sequence of clown fish, and Fig. 5 (b) is the video sequence of basketball player, and Fig. 5 (c) is The video sequence of the monkey, Fig. 5(d) is the video sequence of the single skier. The comparative existing technologies used include ADL (Adaboost-based Level Set Method), see literature: X.Sun, H.X.Yao and S.P.Zhang. A Novel Supervised Level Set Method for Non-Rigid Object Tracking[C], IEEE Conference on Computer Vision and Pattern Recognition, 2011, 3393-3400), SPDL (Superpixel Driven Level Set Method), see literature: X.Zhou, X.Li, T.J.Chin and D.Suter.Superpixel-Driven Level Set Tracking[C].IEEE International Conference on Image Processing.2012.409-412.) and SPT (Superpixel Tracking Method), see literature: S.Wang, H.C.Lu, F.Yang and M.H.Yang. Superpixel Tracking[C].IEEE International Conference on Computer Vision.2011.1323 -1330.), where ADL is a pixel-level Level Sets tracking method based on Adaboost; SPDL is a superpixel-driven Level Sets method, but does not consider measure learning in the process of building the apparent model; and SPT is based on superpixel bounding The box tracking method uses the mean shift clustering algorithm to obtain a confidence map of superpixels, and performs Level Sets tracking based on the confidence map to obtain the contour result. It can be seen from FIG. 5 that the superpixel-based active contour tracking method of the present invention has higher accuracy and robustness. the

尽管上面对本发明说明性的具体实施方式进行了描述，以便于本技术领域的技术人员理解本发明，但应该清楚，本发明不限于具体实施方式的范围，对本技术领域的普通技术人员来讲，只要各种变化在所附的权利要求限定和确定的本发明的精神和范围内，这些变化是显而易见的，一切利用本发明构思的发明创造均在保护之列。 Although the illustrative specific embodiments of the present invention have been described above, so that those skilled in the art can understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, As long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list. the

Claims

1. an active contour tracking method based on superpixels, comprising the following steps:

S1: Divide the training image into two parts, the target and the background, perform superpixel segmentation, extract the feature vector of each superpixel, and construct the target training sample pool T _obj and the background training sample pool T _bac ;

S2: According to the training sample, the projection matrix L of the distance measure is obtained by using the measure learning method, and the projection matrix L is updated every m, m≥1 frame test image;

S3: According to the projection matrix L of the training sample pool and the distance measure, construct a discriminative appearance model based on superpixels, in which the confidence score of each superpixel

The calculation formula is:

{S S}_{c c}^{sp sp} = = \frac{11 - - P P ((sp sp | | bac bac)) / / P P ((sp sp | | obj obj))}{11 + + P P ((sp sp | | bac bac)) / / P P ((sp sp | | obj obj))}

Among them, P(sp|obj) and P(sp|bac) represent the likelihood probability that the superpixel sp belongs to the object class obj and the background class bac, respectively, and are obtained by using a non-parametric kernel density estimation method;

S4: Select a local area including the target in the current frame test image, perform superpixel segmentation on the local area, record the number of superpixels as N, and extract each superpixel sp _k , 1≤k≤N Feature vector f _k ; calculate the confidence score of each superpixel according to the confidence score calculation formula in step S3

Get the confidence map of the test image;

{F f}_{data data}^{i i,, j j} = = \{\begin{matrix} {S S}_{c c}^{{sp sp}_{k k}} & if if & {x x}_{i i,, j j} &Element; &Element; {{{sp sp}_{k k}}}_{k k = = 11}^{N N} \\ - - 11 & if if & {x x}_{i i,, j j} &NotElement; &NotElement; {{{sp sp}_{k k}}}_{k k = = 11}^{N N} \end{matrix}

Among them, (i, j) represents the coordinates of the pixels in the test image;

S6: the velocity field of the test image obtained in step S5

S7: According to the contour tracking results obtained in step S6, put the superpixels of the target and the background into the corresponding training sample pools to update the training sample pools, return to step S1 to rebuild the target training sample pool T _obj and background training Sample pool T _bac .

2. The active contour tracking method according to claim 1, characterized in that, the measure learning method in the step S2 is a large edge nearest neighbor LMNN measure learning method.

3. The active contour tracking method according to claim 1, wherein the method for selecting the local area in the step S4 is: when the first frame test image selects the local area, manually specify the initial contour of the target, The local area is determined according to the initial contour; each subsequent frame of the test image is determined according to the contour tracking result of the previous frame of the test image.

4. The active contour tracking method according to claim 1, wherein the evolution equation in the step S6 is:

\frac{{Φ Φ}_{t t} - - {Φ Φ}_{t t - - 11}}{Δt Δt} + + (({F f}_{data data}^{i i,, j j} + + {F f}_{curv curv})) \cdot \cdot | | {&dtri; &dtri; Φ Φ}_{t t - - 11} | | = = 00

where _Φt is the level set function of the tth iteration, Φt _-1 is the level set function of the t-1th iteration, and the initial function of the level set function _Φ0 is the level set of the contour tracking result of the previous frame test image function, Δt is the preset iteration step size, F _curv = εκ is the internal evolution rate only related to the contour curvature κ, and ε is a preset constant.

5. According to the active contour tracking method described in any one of claims 1 to 4, it is characterized in that, in the step S7, the training sample pool adopts a queue mode to update when updating, and newly added samples are arranged at the end of the queue, when the number of samples If the preset queue length is exceeded, the old samples at the front of the queue are deleted.