CN104182976B - Fine moving object in Field Extraction - Google Patents

Fine moving object in Field Extraction Download PDF

Info

Publication number
CN104182976B
CN104182976B CN 201410395284 CN201410395284A CN104182976B CN 104182976 B CN104182976 B CN 104182976B CN 201410395284 CN201410395284 CN 201410395284 CN 201410395284 A CN201410395284 A CN 201410395284A CN 104182976 B CN104182976 B CN 104182976B
Authority
CN
Grant status
Grant
Patent type
Prior art keywords
target
confidence
area
moving object
pixel
Prior art date
Application number
CN 201410395284
Other languages
Chinese (zh)
Other versions
CN104182976A (en )
Inventor
程小六
吕伟
刘华巍
郭峰
肖世良
石君
袁晓兵
Original Assignee
中国科学院上海微系统与信息技术研究所
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Grant date

Links

Abstract

本发明涉及野外运动目标精细化提取方法,包括:获取包含运动目标的序列图像,并对序列图像进行预处理;将预处理后的序列图逐帧做差后分割为多个栅格,根据栅格的特征值确定目标所在的运动区域,并利用栅格法提取目标的运动区域进一步缩小目标范围;在目标的运动区域内对背景进行建模,通过背景减除法得到目标的二值化图,并对所述二值化图进行带反馈的像素级处理;将处理后的二值化图映射到目标所在的彩色图区域,并对所述彩色图区域进行超像素分割;将该分割结果和二值化图进行融合,根据融合结果,计算每个超像素的置信度,阈值化后最终得到精细化的运动目标。 The present invention relates to the field of fine moving object extraction method comprising: acquiring a sequence of images comprising a moving object, and the image is preprocessed sequence; raster is divided into a plurality of the sequence of FIG pretreated make a difference frame by frame, according to the gate determining a motion characteristic value of the region of the target cell is located and the target using the extracted motion area grid method further refine the target range; modeling the background area in the moving object, a target obtained by background subtraction method binarizing, and binarizing the pixel level processing with feedback; binarized color to the map of FIG target area located after the treatment and the color image is divided superpixel area; and the division result binarizing fusion the fusion results, the calculated confidence of each superpixel, finally obtained after thresholding fine moving target. 本发明可以在复杂背景下实现较实时、鲁棒、精细的野外运动目标的提取。 The present invention may be implemented in a more complicated background in real time, extracting robust, sophisticated moving target field.

Description

一种野外运动目标精细化提取方法 Fine moving object in Field Extraction

技术领域 FIELD

[0001] 本发明涉及图像处理技术领域,特别是涉及一种野外运动目标精细化提取方法。 [0001] The present invention relates to the technical field of image processing, more particularly to a moving object refinement field extraction.

背景技术 Background technique

[0002] 从真实的场景中自动提取运动目标在安防监控、精确制导及跟踪、无人地面值守系统、无线传感器网络等诸多方面有非常重要的意义。 [0002] automatically extracted from moving objects in real scene in security monitoring, tracking and precision-guided, unmanned ground duty systems, wireless sensor networks and many other aspects of a very important significance. 尤其在军事应用中,背景往往很复杂,例如野外的丛林背景、山地背景以及恶劣气象条件下的海空背景等。 Especially in military applications, the background is often very complex, such as the wild jungle background, sea and mountain backdrop backgrounds in harsh weather conditions. 在这种复杂环境下,用传统的目标提取算法难以获得较好的效果,其主要原因是运动目标的过程不是一个非线性非高斯的动态过程。 In such a complex environment, extraction algorithm is difficult to obtain good results with a conventional target, mainly because the process of moving objects is not a Gaussian non-linear dynamic process.

[0003] 运动目标的提取是目标识别的重要前提,其提取质量的好坏直接关系到后续目标分析、目标识别、目标跟踪的精确与否。 [0003] The moving object extraction target recognition is an important prerequisite, the extract is directly related to the quality of the subsequent analysis of the target, target identification, target tracking accurate or not. 常用的运动目标提取方法由帧间差法、背景减除法、 光流法。 The moving object extraction method used by the inter-frame difference method, background subtraction method, optical flow method. 帧间差法对运动区域过于敏感,但是抗噪声性能差,容易出现空洞现象;背景减除法需要建立背景模型,通常需要较多序列,运算量偏大;光流法不需要场景的任何信息但是计算量过大,难以承受。 Interframe difference method is too sensitive to motion region, but poor noise immunity, prone to the phenomenon of voids; background subtraction method requires the establishment of the background model, typically requires more sequence, too large amount of computation; information does not need any optical flow method, but the scene calculation is too large, unbearable. 另外一些较新的方法将上述方法和图像分割(graph cut)、轮廓提取等方法结合起来进行目标的提取,但是仍然存在很多问题。 Other newer methods above methods and image segmentation (graph cut), contour extraction methods are combined to extract the target, but there are still many problems. 归纳起来如下:1、在较理想背景下获得;2、很少考虑系统的实时性需求;3、提取的运动目标通常为矩形框区域,而非目标本身的轮廓。 Be summed up as follows: 1, in the ideal background; 2, little consideration real-time requirements of the system; 3, extracted moving object generally rectangular frame area, rather than the target profile itself. 4、大部分输出的是粗糙的二值化的目标图,丢失了颜色纹理等利于目标识别的信息。 4, most of the output is roughened FIG binarized target, target recognition information is lost favor color texture. 这些算法用于野外复杂的背景中时,提取性能急剧下降,而且计算复杂度较高,不能满足实时性要求。 These complex algorithms for field background, a sharp decline in extraction performance, but higher computational complexity, can not meet the real-time requirements.

发明内容 SUMMARY

[0004] 本发明所要解决的技术问题是提供一种野外运动目标精细化提取方法,可以在复杂背景下实现较实时、鲁棒、精细的野外运动目标的提取。 [0004] The present invention solves the technical problem is to provide a moving object refinement field extraction method may be implemented more in real time, robust, fine extracted moving target field in the complex background.

[0005] 本发明解决其技术问题所采用的技术方案是:提供一种野外运动目标精细化提取方法,包括以下步骤: [0005] aspect of the present invention to solve the technical problem are: to provide a moving object refinement field extraction method, comprising the steps of:

[0006] (1)获取包含运动目标的序列图像,并对序列图像进行预处理; [0006] (1) image acquisition sequence comprising a moving object, and the image is preprocessed sequence;

[0007] (2)将预处理后的序列图逐帧做差后分割为多个栅格,根据栅格的特征值确定目标所在的运动区域,并利用栅格法提取目标的运动区域进一步缩小目标范围; [0007] (2) after dividing a sequence diagram pretreated make a difference frame by frame into a plurality of grids, where the target area is determined according to the motion characteristic value of the grid, the grid method using the objectives and extracts a motion region further refine target range;

[0008] (3)在用栅格法提取目标的运动区域内对背景进行建模,通过背景减除法得到目标的二值化图,并对所述二值化图进行带反馈的像素级处理; [0008] (3) modeling the background object in the extracted motion area by the grid method, the target obtained by background subtraction binarizing method, and the binarized pixel level processing performed with Feedback FIG. ;

[0009] (4)将处理后的二值化图映射到目标所在的彩色图区域,并对所述彩色图区域进行超像素分割;将该分割结果和二值化图进行融合,根据融合结果,计算每个超像素的置信度,阈值化后最终得到精细化的运动目标。 [0009] (4) the binarized processed color map to the map of the target area is located, and the color image region division superpixel; segmented result and the binarized FIG fused, the result of fusion calculated for each superpixel confidence, the threshold value of the finally obtained fine moving target.

[0010] 所述步骤(2)具体包括以下子步骤:将处理后的N帧序列图逐帧做差得到帧差序列图,然后对每帧分别分割为多个栅格;栅格的特征值用本栅格的均方差来表示,通过计算同一栅格在不同帧差序列中的均方差来确定该栅格是否属于目标运动区域。 [0010] The step (2) comprises the sub-steps of: a sequence of N frames processed frame by frame in FIG calculating the difference between a frame difference obtained sequence diagram, and each frame is divided into a plurality of grids; grid characteristic values are used to represent the variance of this grid, the grid is determined whether object motion area by calculating the mean square error at the same grille different frames of the sequence difference.

[0011] 所述步骤(3)具体包括以下子步骤:在目标运动区域内采用背景减除法得到目标的二值化图;然后进行形态学开闭运算,去除毛刺,填平缝隙;将帧差后灰度图中对应该二值化区域抠出作为反馈区域,对该反馈区域做平方然后归一化,再二值化并进行形态学处理后得到边缘完整的二值化图;扫描边该二值化图的最外轮廓,然后将轮廓内区域填充为目标区域。 [0011] The step (3) comprises the substeps of: binarizing using a background subtraction method to give a target region in the target motion; then the morphological opening and closing operation, deburring, fill gaps; difference frame after the grayscale image of the cutout area to be binarized as the feedback region, the feedback area and do squared normalization, and then binary morphology process to obtain a complete binary edge map; sides of the scanning the outermost contour map binarized, and then the contour of the inner region of the target area is filled.

[0012] 所述的步骤(4)中计算每个超像素的置信度时结合了对应二值化图信息及空间邻域信息,并对其进行了权重分配;其中,每个超像素的置信度为 [0012] The steps of the binarization corresponding map information and spatial information for each superpixel confidence (4) is calculated, and the right to be re-assigned; wherein each superpixel Confidence degree

Figure CN104182976BD00041

丨,a决定了超像素自身和其邻域在总置信度组成中所占的权重,a越小,目标越完整,同时带入的环境信息也较多;超像素本身的置信度 Shu, a right to determine their own super-pixel and its neighbors in the overall confidence in the composition of share of the weight, the smaller a, the more complete the goal, but also bring more environmental information; super pixels themselves confidence

Figure CN104182976BD00042

:bw(n)表示第n个超像素范围内的目标像素点数;S sp(n)表示第n个超像素的面积;超像素邻域的平均置信度为左、右、上、下四邻域本身置信度的均值 : Bw (n) represents the target number of pixels within the n-th superpixel range; S sp (n) indicates n-th superpixel area; average confidence super pixel neighborhood of the left and right, upper and lower neighbors domain mean own confidence

Figure CN104182976BD00043

丨为超像素左邻域本身置信度、为超像素右邻域本身置信度、为超像素上邻域本身置信度、 为超像素下邻域本身置信度,Nnelghbciur为邻域超像素的个数。 Shu number per se confidence superpixel left neighborhood of confidence superpixel itself the right neighborhood of the own confidence superpixel neighborhood, the neighborhood of the own confidence superpixel, Nnelghbciur neighborhood of superpixel .

[0013]有益效果 [0013] beneficial effects

[0014] 由于采用了上述的技术方案,本发明与现有技术相比,具有以下的优点和积极效果: [0014] By adopting the technical solution of the present invention compared to the prior art, it has the following advantages and positive effects:

[0015] 本发明使用的栅格法改变了传统方法以单像素作为最小处理单元,变为以像素块为单位进行目标运动区域检测,可以非常有效的去除风吹树木造成的一定区域范围内的较大干扰,同时逐级缩减目标区域,降低运算复杂度。 [0015] The present invention is the grid method using the conventional method of changing pixel as a minimum a single processing unit, in units of pixel block becomes the target motion area detection, can be very effectively removed the wind in a certain area of ​​trees caused by large interference, while progressively reducing the target area, to reduce the computational complexity.

[0016] 本发明采用了带反馈的像素级方法来得到较理想的二值化图,相较于传统的方法使得信噪比进行了平方级的提升。 [0016] The present invention uses a pixel-level approach with feedback to obtain ideal binarizing process compared to conventional noise ratio such that a square of magnitude improvement.

[0017] 本发明对精简的二值化图和超像素分割图进行融合,避免了二值化图的粗糙及直接利用输入源图计算量巨大的缺点,同时吸收了获得二值化图的快速性及超像素分割图精细的颜色、边界等信息,取长补短。 [0017] The present invention is of streamlined binarizing and superpixel segmentation map fusion, to avoid a great rough and direct use of a calculation input source binarizing the disadvantage, while absorbing quickly obtain binarized FIG. and FIG superpixel division of fine information of the color, border, etc., each other. 其中,超像素置信度的计算结合了对应的二值化图信息及空间邻域信息,并对其进行了权重分配,使得提取的目标更完整,同时带入的环境信息降至最小。 Wherein computing a confidence bound superpixel binarized map information and spatial information corresponding to, and subjected to a weight distribution, so that a more complete extraction target, while minimizing environmental information into.

[0018] 本发明最终得到噪声小、轮廓清晰的精细化运动目标。 [0018] The present invention finally obtained noise is small, a clear outline of the moving target fine. 在整个过程中无需背景等先验知识,可以在复杂背景下实现较实时、鲁棒、精细的野外运动目标的提取。 Without having prior knowledge background, you can achieve real-time in a complex background than in the whole process, to extract the robust, sophisticated field moving target.

附图说明 BRIEF DESCRIPTION

[0019] 图1是本发明的流程图。 [0019] FIG. 1 is a flowchart illustrating the present invention.

具体实施方式 detailed description

[0020] 下面结合具体实施例,进一步阐述本发明。 [0020] The following embodiments with reference to specific embodiments, further illustrate the present invention. 应理解,这些实施例仅用于说明本发明而不用于限制本发明的范围。 It should be understood that these embodiments are illustrative only and the present invention is not intended to limit the scope of the invention. 此外应理解,在阅读了本发明讲授的内容之后,本领域技术人员可以对本发明作各种改动或修改,这些等价形式同样落于本申请所附权利要求书所限定的范围。 Furthermore, it should be understood that, after reading the teachings of the present invention, those skilled in the art that various changes or modifications may be made to the present invention, and these equivalents also fall within the scope of the appended claims of the present application as defined.

[0021]本发明的实施方式涉及一种野外运动目标精细化提取方法,如图1所示,包括以下步骤:获取包含运动目标的序列图像,并对序列图像进行预处理;将预处理后的序列图逐帧做差后分割为多个栅格,根据栅格的特征值确定目标所在的运动区域,并利用栅格法提取目标的运动区域进一步缩小目标范围;在用栅格法提取目标的运动区域内对背景进行建模,通过背景减除法得到目标的二值化图,并对所述二值化图进行带反馈的像素级处理;将处理后的二值化图映射到目标所在的彩色图区域,并对所述彩色图区域进行超像素分割; 将该分割结果和二值化图进行融合,根据融合结果,计算每个超像素的置信度,阈值化后最终得到精细化的运动目标。 [0021] Embodiment of the present invention relates to a moving object refinement field extraction method, shown in Figure 1, comprising the steps of: acquiring an image sequence comprising a moving object, and the image is preprocessed sequence; after pretreatment FIG done frame by frame sequence of the divided plurality of raster difference, the target value determining the motion area where the feature grid, and the grid method using the extracted moving object region further refine the target range; target raster extraction method a motion model of the background area, the target obtained by background subtraction binarizing method, and the binarizing processing with pixel-level feedback; binarization processed FIGS mapped to the target is located FIG region color, and the color image region division superpixel; segmented result and the binarizing fusion, a fusion results, the calculated confidence of each superpixel, finally obtained after thresholding refinement motion aims. 具体步骤如下: Specific steps are as follows:

[0022]步骤100,获取包含运动目标的高清序列图像,做灰度化、抽样等预处理。 [0022] Step 100, acquiring high-definition image sequences containing moving targets, do graying, sample pretreatment.

[0023]步骤200,将处理后的序列图逐帧做差后分割为多个栅格,利用栅格法提取目标的运动区域,进一步缩小目标范围。 [0023] Step 200, the processing sequence frame by frame view of calculating the difference between the divided multiple grids, the grid method using the extraction target area of ​​movement, further refine the target range.

[0024] 本发明中的栅格法改变了传统方法以单像素作为最小处理单元,变为以像素块为单位进行目标运动区域检测,可以非常有效的去除风吹树木造成的一定区域范围内的较大干扰。 [0024] In the present invention, the grid method changes the traditional method of processing a single pixel as a minimum unit, in units of pixel block becomes the target motion area detection, it can be very effectively removed the wind in a certain area of ​​trees caused by greater interference.

[0025] 具体如下:栅格法需要综合考虑来确定栅格大小,栅格过密会导致计算量增大且去除干扰效果变差,栅格过疏会导致获得的目标区域过于扩大。 [0025] as follows: grid method must be considered to determine the size of the grid, the grid can cause too dense and the amount of calculation is increased to remove the interference effect deteriorates grid depopulation cause the target region obtained by expanding too. 具体计算方法如下:将处理后的N帧序列图逐帧做差得到帧差序列图(记为0〖3 = 1,2,..^1),然后对每帧分别分割为多个栅格,记栅格宽为Gw,高为Gh,栅格总数为Gn。 The calculation method is as follows: a sequence of N frames processed frame by frame in FIG calculating the difference obtained difference frame sequence diagram (referred to as 0 〖3 = 1,2, .. ^ 1), then each divided into a plurality of grids each frame , denoted raster width Gw, height of Gh, the total number of grid Gn. 栅格的特征值用本栅格的均方差来表示, 通过计算同一栅格在不同帧差序列中的均方差来确定该栅格是否属于目标运动区域。 Characterized in rasters are variance values ​​to represent this grid, this grid is determined whether object motion area by calculating the mean square error at the same grille different frames of the sequence difference. 记栅格i为Gi,t是Dt的索引值。 I is referred grid Gi, t is the index value of Dt. 第i个栅格在帧差图Dt中的平均能量值: The i-th grid average energy value of the frame difference in FIG Dt:

[0026] [0026]

Figure CN104182976BD00051

[0027] 其中,是Dt中第i个栅格的行索引。 [0027] wherein, Dt is the i-th row index grating. 是Dt中第i个栅格的列索引。 Dt is the i-th column index grating.

[0028] N-1帧帧差序列中的平均值: [0028] N-1 frames in a sequence difference average:

[0029] [0029]

Figure CN104182976BD00052

[0030] N-1帧帧差序列图中的标准方差为: [0030] The standard deviation of the difference between frames N-1 is a sequence diagram:

[0031] [0031]

Figure CN104182976BD00053

[0032]阈值如下: [0032] Threshold values ​​are as follows:

Figure CN104182976BD00054

[0033] [0033]

[0034] [0034]

[0035]其中,A是衰减增益因子,一般取经验值,取值范围0.5~2。 [0035] where, A is the attenuation factor of the gain, and generally experience value, in the range of 0.5 to 2.

[0036]当标准方差超过阈值Th时,该栅格就是运动区域,这样运动区域可以表示为: [0036] When the standard deviation exceeds the threshold value Th, the grid is a motion area, so that a motion region may be expressed as:

[0037] M={Gi},当SDi大于阈值Th。 [0037] M = {Gi}, SDi when greater than the threshold Th.

[0038]步骤300,在上述目标运动区域内对背景进行建模,通过背景减除法得到目标的二值化图;对该二值化图做了带反馈的像素级处理,相较于传统的方法,使得信噪比获得了平方级的提升,并且进一步缩小目标区域,得到较纯净、完整的二值化图。 [0038] Step 300, a background region in said target motion model, a target obtained by background subtraction binarizing method; FIG do the binarized pixel level processing with feedback, compared to conventional the method, to obtain a square level noise ratio improvement, and further refine the target region, to obtain more pure and complete binarized FIG.

[0039] 算法如下:在目标运动区域内采用背景减除法得到目标的二值化图;然后进行形态学开闭运算,去除毛刺,填平缝隙。 [0039] The algorithm is as follows: FIG. Binarized using the background subtraction method to give a target region in the target motion; then the morphological opening and closing operation, deburring, fill the gap. 此时二值化图的最小外接矩形内可以认为是纯粹的二值化目标区域。 At this time, the smallest circumscribed rectangle binarizing may be considered purely binary target area. 但是此时的二值化目标区域在信噪比低的情况下会出现不完整的边界及不可预测的干扰。 But this time the target area of ​​the binarized be incomplete boundaries and unpredictable interference occurs at low SNR. 带反馈的方法如下:将帧差后灰度图中对应该二值化区域抠出来,对该区域做平方然后归一化,结果是拉大了原来的前景区域和背景区域的差距,此时再用二值化并进行形态学处理后得到的二值化图边缘比较完整。 The method with feedback as follows: After the grayscale image of the frame difference should pull out of the binarized region, the region is then normalized squared do, the result is a widening gap between the original foreground and background regions, in which case FIG binarized after binarized and then treating the resulting edge morphological more complete. 但是内部仍然可能会出现孔洞。 But the internal pores may still occur. 这时需要使用孔洞填充法处理,具体方法为扫描其上下左右区域的最外轮廓,然后将轮廓内区域填充为目标区域。 At this hole filling method requires the use of processing, the most specific method of scanning the outer contour of the left and right upper and lower regions, and then the filled region as a target area within the outline. 最后可得到消除噪音的、不含孔洞、边界较完整的二值化图。 Finally, the noise elimination can be obtained, free of voids, more complete boundary binarized FIG.

[0040] 步骤400,根据二值化图映射到目标所在的彩色图区域,并对该区域进行超像素分害J。 [0040] Step 400, the binarized according to the map of the target area where the color map, and the region over the pixel component damage J. 本发明对精简的二值化图和超像素分割图进行融合,避免了二值化图的粗糙及直接利用输入源图计算量巨大的缺点,同时吸收了获得二值化图的快速性及超像素分割图精细的颜色、边界等信息,取长补短,为后续得到精细化的彩色目标做铺垫。 The present invention is of streamlined binarizing and superpixel segmentation map fusion, avoiding rough and direct use huge input source in FIG calculated amount of disadvantages binarizing while absorbing rapidity obtained binarizing and super FIG finely divided pixel color information, and the like border each other, for a subsequent fine to obtain a target color pave.

[0041] 本发明中超像素的置信度计算结合了对应的二值化图信息及空间邻域信息,并对其进行了权重分配,使得提取的目标更完整,同时带入的环境信息降至最小。 [0041] Confidence super pixel calculated in conjunction with the present invention binarizing information and spatial information corresponding to, and subjected to a redistribution rights, so that a more complete extraction target, while minimizing environmental information into .

[0042] 具体如下:超像素分割法使用简单线性迭代聚类算法,该方法计算量小,分成的超像素大小较均匀。 [0042] as follows: using a simple linear division method superpixel iterative clustering algorithm, which calculates the amount of small, superpixel into more uniform size. 基于二值化图和超像素分割图的融合结果,计算每个超像素的置信度。 FIG binarization and segmentation map superpixel fusion result is calculated for each superpixel confidence. 每个超像素的置信度为: Each super pixel confidence level:

[0043] [0043]

Figure CN104182976BD00061

[0044] 参数a决定了超像素自身和其邻域在总置信度组成中所占的权重,权重对结果有较大影响。 [0044] The parameter a determines the superpixel itself and its neighbors in the overall confidence composition weight share of the weight, the weight has a great influence on the results. a越小,目标越完整,同时带入的环境信息也较多。 a smaller target more complete, while incorporating environmental information is greater. 相反,a越大,目标越不完整,同时带入的环境信息也较少。 In contrast, a larger target the less complete, while environmental information into fewer.

[0045] 其中该超像素本身的置信度: [0045] wherein the superpixel itself confidence:

[0046] [0046]

Figure CN104182976BD00062

[0047] 其中,Sbw(n)表示第n个超像素范围内的目标像素(指对应二值化图内目标)点数。 [0047] wherein, Sbw (n) represents the target pixel within the n-th superpixel range (refer to the corresponding binarizing the target) points. Ssp(n)表示第n个超像素的面积。 SSP (n) indicates n-th superpixel area.

[0048]该超像素邻域的平均置信度为左、右、上、下四邻域本身置信度的均值: [0048] The super-pixel neighborhood average confidence of left, right, upper, lower mean confidence neighbors domain itself:

[0049] [0049]

Figure CN104182976BD00063

[0050] 其中,为超像素左邻域本身置信度、为超像素右邻域本身置信度、(7,^»为超像素上邻域本身置信度、Cf»(«)为超像素下邻域本身置信度,Nnelghb_为邻域超像素的个数。 [0050] wherein, per se confidence superpixel left neighborhood of superpixel the right neighborhood itself confidence level (7, ^ »is ultra pixel neighborhood itself confidence level of Cf >> (<<) of the superpixel subjacent confidence domain itself, Nnelghb_ superpixel is the number of neighborhood.

[0051]确定置信度阈值,置信度高于阈值的超像素保留,低于阈值的超像素抛弃,最终得到精细化的彩色运动目标。 [0051] determining a confidence threshold, the confidence is higher than a threshold retained super pixel, the pixel is below the threshold value over discarded to finally obtain a color moving object refinement.

[0052]下面以一个具体的实施例进一步说明本发明。 [0052] The following further illustrate the present invention in a specific embodiment.

[0053] (1)获取包含运动目标的高清序列图像,输入的序列图像为高清彩色图,建议每张图像素在300万像素以上。 [0053] (a) acquiring an image sequence comprising a moving object definition, the definition of a sequence of images input color map, recommended for each pixel in FIG. 300 million pixels. 抽样后的灰度图像素在20万~50万像素之间较合适。 Grayscale pixels after sampling more appropriate between 200,000 to 500,000 pixels. 对其做灰度化、抽样等预处理。 Its done graying, sample pretreatment.

[0054] (2)将处理后的帧序列逐帧做差得到帧差序列图(记为0\七=1,2,.』-1),然后对每帧分别分割为多个栅格,栅格宽Gw为原图像的1/32,高Gh为原图像的1/6,这样栅格总数Gn为192。 [0054] (2) the frame-processed frame sequence to obtain the frame difference calculating the difference between a sequence diagram (referred to as 0 \ seven = 1,2, '-. 1), and then each frame is divided into a plurality of grids, grid width Gw is a 1/32 of the original image, the original image is high 1/6 Gh, Gn so that the total number of grid 192. 首先,某帧中某栅格的特征用本栅格内所有点的能量均值来表示,然后计算同一栅格特征在不同帧差序列中的均方差来表示该栅格的时间能量,然后计算同一帧(一般为中间帧)中所有栅格的时间能量的均方差,该均方差及其均值的和衰减a(a经验值范围0.5 ~2,典型值为1)倍作为阈值,超过阈值的栅格区集合即为目标运动区域。 First, features in a certain frame grid with all points in this lattice energy represented the mean and variance are calculated in the same grid features different frame difference shows the time sequence of the power grid, then the same calculation a frame (usually an intermediate frame) energy in time all the rasters are variance, and the mean square error and the mean attenuation a (a empirical value range of 0.5 to 2, typically 1) times as a threshold value, exceeds the threshold value of the gate grid target area is the collection of sports area.

[0055] (3)在上述目标运动区域内采用背景减除法得到目标的二值化图;由于获得的帧图像序列数目不够多,所以参数建模(如高斯背景建模)会出现偏差且需要不断更新,计算量也较大。 [0055] (3) The binarizing the background subtraction method to obtain the target region in said target motion; Since the number of frames of the image sequence obtained enough, the model parameters (e.g., Gaussian background model) and need go awry constantly updated, the calculation amount is large. 由于在运动目标在不同帧中同一地点出现的时间是很少的,所以大部分时间背景处于显露区域,所以简单的非参数中值建模法比较适用,该方法计算量较小且效果较好。 Since the target temporal motion occurs in the same location in different frames is very small, so most of the time in the background area exposed, so that a simple non-parametric modeling method is more suitable value, the smaller and better method of calculating the amount of . 之后采用Otsu阈值法对背景减除后的图像二值化。 After using the Otsu threshold method for image binarization after background subtraction.

[0056] (4)对该二值化图首先做结构元素为3X3的中值滤波,可以去除大部分的孤立噪点;然后进行形态学开闭运算,去除毛刺,填平缝隙。 [0056] (4) First the binarizing as structural elements of the 3X3 median filter, may remove most of the isolated noise; then the morphological opening and closing operation, deburring, fill the gap. 此时二值化图的最小外接矩形内可以认为是纯粹的二值化目标区域。 At this time, the smallest circumscribed rectangle binarizing may be considered purely binary target area. 但是此时的二值化目标区域在信噪比低的情况下会出现不完整的边界及不可预测的干扰。 But this time the target area of ​​the binarized be incomplete boundaries and unpredictable interference occurs at low SNR. 去除方法如下:将帧差后灰度图中对应该二值化区域抠出来,对该区域做平方然后归一化,结果是拉大了原来的前景区域和背景区域的差距,此时再用二值化并进行形态学处理得到的二值化图边缘比较完整。 Removed as follows: The difference between the grayscale image of the subsequent frame should pull out of the binarized region, the region is then normalized squared do, the result is a widening gap between the original foreground and background regions, in which case then binarization and binarizing the edge morphological obtained more complete. 但是内部仍然可能会出现孔洞。 But the internal pores may still occur. 这时需要使用孔洞填充法处理,具体方法为扫描其上下左右区域的最外轮廓,然后将轮廓内区域填充为目标区域。 At this hole filling method requires the use of processing, the most specific method of scanning the outer contour of the left and right upper and lower regions, and then the filled region as a target area within the outline. 最后可得到消除噪音的、不含孔洞、边界较完整的二值化图; [0057] (4)上述二值化区域回到原始彩色图像,需要经过两次坐标转换,才能找到目标所在的彩色图区域。 Finally, the noise elimination can be obtained, free of voids, more complete boundary binarizing; [0057] (4) The area of ​​the binarized image back to the original color, after two coordinate conversion, to find the target color is located FIG area. 对上述彩色目标区域进行基于简单线性迭代聚类算法超像素分割,该方法计算量小,分成的超像素大小较均匀。 The color of the target area is divided based on a simple iterative clustering algorithm linear superpixel, the method calculates a small amount into the superpixel more uniform size.

[0058] (5)融合二值化图和超像素分割图之前,需要先统一两者的坐标。 [0058] prior to (5) and fusion binarizing superpixel segmentation map, both the need to coordinate uniform. 因为之前的彩色高清源图是经过抽样后参与后续计算的,所以此时得到的二值化图需要先以同样的比例插值回去,这样才能保证插值后的二值化图和彩色目标源图一一对应。 Because the source image before the color definition is involved in the sample after a subsequent calculation, the binarizing obtained at this time need to be interpolated back at the same rate, so as to ensure binarizing the interpolated target source color and a FIG. a correspondence. 参数a决定了超像素自身和其邻域在总置信度组成中所占的权重,权重对结果有较大影响。 Determine the parameters of a super-pixel itself and its neighbors in the composition of total confidence in the share of heavy weights, weights have a greater impact on the results. 经过多次实验,取经验值0.5比较合适。 After several experiments, taking appropriate empirical value 0.5. 置信度的阈值经验值为0.5。 Confidence threshold empirical value 0.5. 小于阈值的抛弃区域用白色代替。 Less than a threshold region was replaced with a white disposable.

Claims (3)

  1. 1. 一种野外运动目标精细化提取方法,其特征在于,包括以下步骤: (1) 获取包含运动目标的序列图像,并对序列图像进行预处理; (2) 将预处理后的序列图逐帧做差后分割为多个栅格,根据栅格的特征值确定目标所在的运动区域,并利用栅格法提取目标的运动区域进一步缩小目标范围;具体为,将处理后的N帧序列图逐帧做差得到帧差序列图,然后对每帧分别分割为多个栅格;栅格的特征值用本栅格的均方差来表示,通过计算同一栅格在不同帧差序列中的均方差来确定该栅格是否属于目标运动区域; (3) 在用栅格法提取目标的运动区域内对背景进行建模,通过背景减除法得到目标的二值化图,并对所述二值化图进行带反馈的像素级处理; (4) 将处理后的二值化图映射到目标所在的彩色图区域,并对所述彩色图区域进行超像素分割;将该分割结果和二值 A moving object refinement field extraction method comprising the steps of: (1) image acquisition sequence comprising a moving object, and the image is preprocessed sequences; (2) a sequence pretreated by FIG. calculating the difference between the divided frame into a plurality of grids, where the target area is determined according to the motion characteristic value of the grid, and the grid method using the extracted moving object region further refine the target range; specifically, the N frames processed sequence diagram calculating the difference between frames frame by frame to obtain a difference sequence diagram, respectively, and each frame is divided into a plurality of grids; a mean square error of this grid raster feature value represented by the same calculation grid are in different frames of a sequence difference determining the variance of the motion of the target region belongs grid; (3) model of the background object in the extracted motion area by the grid method, the target obtained by background subtraction binarizing method, and the binary FIG pixel level processing of feedback band; and (4) the binarizing color image mapped to the target area is located after the treatment, and the color image region division superpixel; segmented result and the binary 图进行融合,根据融合结果,计算每个超像素的置信度, 阈值化后最终得到精细化的运动目标。 FIG fusion The fusion results, the calculated confidence of each superpixel, finally obtained after thresholding fine moving target.
  2. 2. 根据权利要求1所述的野外运动目标精细化提取方法,其特征在于,所述步骤(3)具体包括以下子步骤:在目标运动区域内采用背景减除法得到目标的二值化图;然后进行形态学开闭运算,去除毛刺,填平缝隙;将帧差后灰度图中对应该二值化区域抠出作为反馈区域,对该反馈区域做平方然后归一化,再二值化并进行形态学处理后得到边缘完整的二值化图;扫描该二值化图的最外轮廓,然后将轮廓内区域填充为目标区域。 The field of the moving object extraction method of a refined, wherein said step (3) comprises the substeps claim: background subtraction method using a target obtained binarized FIG motion in the target region; then morphological opening and closing operation, deburring, fill the gap; frame difference after the grayscale image of the cutout area to be binarized as the feedback region, the feedback area and do squared normalization, and then binarization and the morphology process to obtain a complete binary edge map; scanning the outermost contour of the binarizing, and then filled with the inner contour of the target area region.
  3. 3. 根据权利要求1所述的野外运动目标精细化提取方法,其特征在于,所述的步骤(4) 中计算每个超像素的置信度时结合了对应二倌化图信息及空间邻域信息,并对其进行了权重分配;其中,每个超像素的置信度# The field of the moving object as claimed in claim 1, fine extraction method, wherein the binding of the map information corresponding to groom and two spatial neighborhood each super pixel calculated confidence (4) in said step information, and the right to re-assign them; where each pixel super confidence #
    Figure CN104182976BC00021
    决定了超像素自身和其邻域在总置信度组成中所占的权重,α越小,目标越完整,同时带入的环境信息也较多;超像素本身的置信) The right to decide their own super-pixel and its neighbors in the overall confidence in the composition of share of the weight, the smaller α, the more complete the goal, but also bring more environmental information; confidence super pixel itself)
    Figure CN104182976BC00022
    良示第η个超像素范围内的目标像素点数;Ssp(n)衷示第η个超像素的面积:超像素邻域的平询詈信度为左、右、h、下四邻域本身置信度的均1 Liang shows a first η target number of pixels within a super pixel range; Ssp (n) co shown of η th superpixel area: flat Inquiry Dirty reliability super pixel neighborhood of the left and right, h, the neighbors domain itself Confidence degrees are 1
    Figure CN104182976BC00023
    与超像素左邻域本身置信度 And ultra-left pixel neighborhood itself Confidence
    Figure CN104182976BC00024
    为超像素右邻域本身置信ϋ Confidence ϋ is the right super-pixel neighborhood itself
    Figure CN104182976BC00025
    3超像素上邻域本身置信ί 3 super pixel neighborhood itself confidence ί
    Figure CN104182976BC00026
    丨为超像素下邻域本身置信度,Nneighboui·为邻域超像素的个数。 Shu is the neighborhood itself under ultra-confidence pixels, Nneighboui · is the number of pixels of super neighborhood.
CN 201410395284 2014-08-12 2014-08-12 Fine moving object in Field Extraction CN104182976B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 201410395284 CN104182976B (en) 2014-08-12 2014-08-12 Fine moving object in Field Extraction

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 201410395284 CN104182976B (en) 2014-08-12 2014-08-12 Fine moving object in Field Extraction

Publications (2)

Publication Number Publication Date
CN104182976A true CN104182976A (en) 2014-12-03
CN104182976B true CN104182976B (en) 2017-02-01

Family

ID=51963994

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 201410395284 CN104182976B (en) 2014-08-12 2014-08-12 Fine moving object in Field Extraction

Country Status (1)

Country Link
CN (1) CN104182976B (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6233007B1 (en) * 1998-06-22 2001-05-15 Lucent Technologies Inc. Method and apparatus for tracking position of a ball in real time
CN103237228A (en) * 2013-04-28 2013-08-07 清华大学 Time-space consistency segmentation method for binocular stereoscopic video

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6233007B1 (en) * 1998-06-22 2001-05-15 Lucent Technologies Inc. Method and apparatus for tracking position of a ball in real time
CN103237228A (en) * 2013-04-28 2013-08-07 清华大学 Time-space consistency segmentation method for binocular stereoscopic video

Also Published As

Publication number Publication date Type
CN104182976A (en) 2014-12-03 application

Similar Documents

Publication Publication Date Title
Kim et al. Optimized contrast enhancement for real-time image and video dehazing
Nouar et al. Improved object tracking with camshift algorithm
CN101916370A (en) Method for processing non-feature regional images in face detection
CN101246551A (en) Fast license plate locating method
CN101551863A (en) Method for extracting roads from remote sensing image based on non-sub-sampled contourlet transform
CN101916448A (en) Moving object detecting method based on Bayesian frame and LBP (Local Binary Pattern)
Chen et al. A new process for the segmentation of high resolution remote sensing imagery
Zhang et al. Moving vehicles detection based on adaptive motion histogram
CN101872546A (en) Video-based method for rapidly detecting transit vehicles
CN102592268A (en) Method for segmenting foreground image
Sharma et al. Colour image segmentation techniques and issues: an approach
Shanmugavadivu et al. Particle swarm optimized multi-objective histogram equalization for image enhancement
CN101087413A (en) Division method of motive object in video sequence
CN104156704A (en) Novel license plate identification method and system
CN101854467A (en) Method for adaptively detecting and eliminating shadow in video segmentation
CN101216943A (en) A method for video moving object subdivision
CN103116985A (en) Detection method and device of parking against rules
Zhu et al. An adaptive histogram equalization algorithm on the image gray level mapping
CN101533515A (en) Background modeling method based on block facing video monitoring
CN101556693A (en) Division method for extracted watershed SAR image with threshold method and marking
CN102968782A (en) Automatic digging method for remarkable objects of color images
CN102147861A (en) Moving target detection method for carrying out Bayes judgment based on color-texture dual characteristic vectors
CN102298773A (en) One shape adaptive non-local means denoising
Liu et al. Cast shadow removal with gmm for surface reflectance component
CN101916446A (en) Gray level target tracking algorithm based on marginal information and mean shift

Legal Events

Date Code Title Description
C06 Publication
C10 Entry into substantive examination
C14 Grant of patent or utility model