CN114897938A - Improved cosine window related filtering target tracking method - Google Patents

Improved cosine window related filtering target tracking method Download PDF

Info

Publication number
CN114897938A
CN114897938A CN202210575210.8A CN202210575210A CN114897938A CN 114897938 A CN114897938 A CN 114897938A CN 202210575210 A CN202210575210 A CN 202210575210A CN 114897938 A CN114897938 A CN 114897938A
Authority
CN
China
Prior art keywords
target
cosine window
scale
image
matrix
Prior art date
Legal status (The legal status 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 status listed.)
Granted
Application number
CN202210575210.8A
Other languages
Chinese (zh)
Other versions
CN114897938B (en
Inventor
孙家伟
邓丽珍
朱虎
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
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
Application filed by Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN202210575210.8A priority Critical patent/CN114897938B/en
Publication of CN114897938A publication Critical patent/CN114897938A/en
Application granted granted Critical
Publication of CN114897938B publication Critical patent/CN114897938B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • G06T7/248—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments involving reference images or patches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/262—Analysis of motion using transform domain methods, e.g. Fourier domain methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

A relevant filtering target tracking method for improving a cosine window provides improvement of a scale self-adaptive cosine window, peak clipping processing is carried out on the cosine window by using the initial size of a target, in addition, scale change of the target is considered, and dynamic adjustment is carried out on the peak clipping position of the cosine window through a scale change factor, so that the tracking success rate of a tracking algorithm in a target scale change scene is improved; the ADTrack algorithm is improved, self-constraint and mutual constraint between a training target filter and a background filter are introduced, an objective function is further optimized, tracking drift and template rapid degradation phenomena of an existing tracker are optimized, and ADMM optimization derivation is performed on a new objective function.

Description

一种改进余弦窗的相关滤波目标跟踪方法A Correlation Filtering Target Tracking Method with Improved Cosine Window

技术领域technical field

本发明属于目标跟踪技术领域，具体涉及一种改进余弦窗的相关滤波目标跟踪方法。The invention belongs to the technical field of target tracking, and in particular relates to an improved cosine window correlation filtering target tracking method.

背景技术Background technique

图像识别与目标跟踪是计算机视觉中的一项关键工作，其具体内容是面向视频序列中的目标状态加以连续的推断处理，核心工作是定位视频在全部帧数中涉及到的跟踪目标，来获得对应目标的运动轨迹，针对每一帧图像提供跟踪目标的活动区域。目标跟踪技术目前已经被军事以及民用领域普遍推广，给人生活带来大量便利，随着人类生活水平的提高，对于目标跟踪的要求也越来越高。尽管目标跟踪相关技术每年都在革新,相关算法在性能优化方面依然存在需求，现阶段研究人员仍然面临着众多挑战。Image recognition and target tracking is a key work in computer vision. Its specific content is to continuously infer the target state in the video sequence. The core work is to locate the tracking target involved in all frames of the video to obtain Corresponding to the motion trajectory of the target, the active area of the tracking target is provided for each frame of image. Target tracking technology has been widely promoted in the military and civilian fields, bringing a lot of convenience to people's lives. With the improvement of human living standards, the requirements for target tracking are also getting higher and higher. Although the technology related to target tracking is innovated every year, there is still a need for performance optimization of related algorithms, and researchers still face many challenges at this stage.

近年来，针对目标跟踪的各种问题，许多研究者们提出了很多方法来设计一个稳定、准确的目标跟踪器。2017年Galoogahi等提出了BACF算法，为了提高样本的质量，提出了关键的“裁剪”思想，将原本的采样区域扩大，获得更多的背景信息，然后对循环样本进行中心裁剪，这样即扩大了样本数量，由提高了样本质量。再之后，2021年Li等提出的ADTrack算法，利用一种0-1掩膜提取目标信息，然后将目标和目标背景信息分开训练，进一步改进了BACF算法。In recent years, for various problems of target tracking, many researchers have proposed many methods to design a stable and accurate target tracker. In 2017, Galoogahi et al. proposed the BACF algorithm. In order to improve the quality of the samples, they proposed the key "cropping" idea, which expanded the original sampling area to obtain more background information, and then centrally cropped the loop samples, which expanded the size of the sample. The sample size is improved by the sample quality. After that, the ADTrack algorithm proposed by Li et al. in 2021 uses a 0-1 mask to extract target information, and then trains the target and target background information separately, further improving the BACF algorithm.

但是，ADTrack也存在着很多不足。首先是余弦窗函数，普通的余弦窗函数直接采用了中间极值四周零值的模型，忽略了目标的实际尺寸，若直接预处理样本，会增加目标纹理而污染目标源信息。除此之外，ADTrack的算法中对于滤波器的约束项也不理想，存在滤波器的过拟合问题和滤波器模型快速退化的情况。However, ADTrack also has many shortcomings. The first is the cosine window function. The ordinary cosine window function directly adopts the model of zero values around the middle extreme value, ignoring the actual size of the target. If the sample is directly preprocessed, the target texture will be added and the target source information will be polluted. In addition, the constraints on the filter in the ADTrack algorithm are not ideal, and there is a problem of overfitting of the filter and rapid degradation of the filter model.

发明内容SUMMARY OF THE INVENTION

本发明是一种改进余弦窗的相关滤波目标跟踪方法，并分别对目标和背景信息进行滤波训练。目前目标跟踪领域还存在着很多问题，如目标尺度变化、目标遮挡等。针对于目标尺度变化，本发明构建一种尺度自适应余弦窗，对余弦窗进行削峰，并且适时更新，再使用目标和目标背景两种滤波器，结合两者的自约束和互约束，对相关滤波目标跟踪模型进行改进，以此提高跟踪性能。The present invention is an improved cosine window correlation filtering target tracking method, and the target and background information are filtered and trained respectively. At present, there are still many problems in the field of target tracking, such as target scale change, target occlusion and so on. Aiming at the change of target scale, the present invention constructs a scale adaptive cosine window, clips the peak of the cosine window, and updates it in time, and then uses two filters of target and target background, combined with the self-constraint and mutual constraint of the two, to The correlation filtering target tracking model is improved to improve the tracking performance.

本发明采用的技术方案大致如下：The technical scheme adopted in the present invention is roughly as follows:

一种改进余弦窗的相关滤波目标跟踪方法，包括如下步骤：A correlation filtering target tracking method with improved cosine window, comprising the following steps:

步骤1，预处理阶段；判断是否处于暗场景，若判断处于暗场景，则对视频序列进行增强，并计算得到掩膜m，用于后续滤波器的训练；Step 1, preprocessing stage; judge whether it is in a dark scene, if it is judged to be in a dark scene, enhance the video sequence, and calculate the mask m, which is used for the training of subsequent filters;

步骤2，提取特征阶段；基于训练样本得到特征集xg，通过削峰处理后的余弦窗实现跟踪目标尺度更新的自适应，基于掩膜m得到目标特征集xo，得到尺度自适应余弦窗模型；Step 2, feature extraction stage; obtain feature set x g based on training samples, realize adaptive tracking target scale update through cosine window after peak clipping processing, obtain target feature set x o based on mask m, and obtain scale adaptive cosine window Model;

步骤3，训练阶段；利用步骤2得到的特征集xg和xo训练出下一帧的两个滤波器hg和ho，其中hg是用于训练背景信息的滤波器，ho是用于训练目标信息的滤波器；Step 3, training phase; use the feature sets x g and x o obtained in step 2 to train two filters h g and ho for the next frame, where h g is the filter for training background information, and ho is Filters for training target information;

步骤4，检测阶段；基于训练后的滤波器和提取的样本通道特征，得到目标响应图，通过目标响应图的最大值确定目标的位置信息。Step 4, detection stage; based on the trained filter and the extracted sample channel features, a target response map is obtained, and the position information of the target is determined by the maximum value of the target response map.

进一步地，步骤1包括如下分步骤：Further, step 1 includes the following sub-steps:

步骤1-1，假设给定彩色图像

Figure BDA0003661864230000031
对图像的RBG三通道进行光度整合：Step 1-1, assuming a color image is given
Figure BDA0003661864230000031
Photometric integration of the RBG three channels of the image:

Figure BDA0003661864230000032
Figure BDA0003661864230000032

其中Ψm(I(x,y))代表着m通道出图像的像素值，并且αR+αG+αB＝1，将彩色图像转换为单通道图像；再对亮度进行对数平均处理：Where Ψ m (I(x, y)) represents the pixel value of the m-channel output image, and α R + α G + α B =1, convert the color image to a single-channel image; then perform logarithmic average processing on the brightness :

Figure BDA0003661864230000033
Figure BDA0003661864230000033

其中δ为常数，用于防止log(0)的出现，

Figure BDA0003661864230000034
代表当前图像的对数平均场景亮度，再通过引入阈值τ判断当前图像是否处于暗场景，
Figure BDA0003661864230000035
小于阈值即为暗场景，表示为S(I)；where δ is a constant to prevent the occurrence of log(0),
Figure BDA0003661864230000034
represents the logarithmic average scene brightness of the current image, and then judges whether the current image is in a dark scene by introducing a threshold τ,
Figure BDA0003661864230000035
Less than the threshold is a dark scene, denoted as S(I);

步骤1-2，对图像进行增强，利用之前获取的图像亮度V(x,y,I)和图像对数平均亮度

Figure BDA0003661864230000036
图像全局的增强矩阵表示为：Step 1-2, enhance the image, use the image brightness V(x, y, I) obtained before and the logarithmic average brightness of the image
Figure BDA0003661864230000036
The enhancement matrix of the image global is expressed as:

Figure BDA0003661864230000037
Figure BDA0003661864230000037

其中Vmax(I)代表图像亮度V(x,y,I)中的最大值，然后对图像的三通道进行增强处理：Where V max (I) represents the maximum value in the image brightness V(x, y, I), and then the three channels of the image are enhanced:

Figure BDA0003661864230000038
Figure BDA0003661864230000038

其中Ie表示增强过的图像，Ψm(Ie(x,y))表示增强过的图像在m通道(x,y)位置处的像素值；由上述增强过后的图像，得到图像的增强部分信息：where I e represents the enhanced image, Ψ m (I e (x, y)) represents the pixel value of the enhanced image at the position of m channel (x, y); from the enhanced image above, the enhanced image is obtained Partial information:

E(I)＝V(I)-V(Ie)E(I)=V(I)-V(I e )

步骤1-3，通过E(I)得到E(I)的平均值μ和标准差σ，由此得到全局的掩膜mg：Steps 1-3, obtain the average μ and standard deviation σ of E(I) through E(I), thereby obtaining the global mask m g :

Figure BDA0003661864230000041
Figure BDA0003661864230000041

通过裁剪矩阵P对mg进行裁剪，得到预期的掩膜m＝mg⊙P，P∈Rw×h，用于提取样本内的目标大小信息。The m g is cropped by the crop matrix P, and the expected mask m=m g ⊙P, P∈R w×h is obtained, which is used to extract the target size information within the sample.

进一步地，步骤2包括如下分步骤：Further, step 2 includes the following sub-steps:

步骤2-1，通过余弦窗预处理、循环矩阵、中心裁剪方法获取大量训练样本，余弦窗处理即为在样本上直接点乘余弦窗函数，循环矩阵和中心裁剪的操作见附图2。对获取的样本进行特征提取，包括灰度信息、颜色信息、梯度信息等，得到特征集合xg；In step 2-1, a large number of training samples are obtained through cosine window preprocessing, circulant matrix, and center clipping methods. Cosine window processing is to directly multiply the cosine window function on the samples. The operations of circulant matrix and center clipping are shown in Figure 2. Perform feature extraction on the acquired samples, including grayscale information, color information, gradient information, etc., to obtain a feature set x g ;

步骤2-2，利用目标初始大小对余弦窗削峰处理，削峰位置为参数Q∈(0,1)，Q由下面的公式得到：Step 2-2, use the initial size of the target to clip the peak of the cosine window, the clipping position is the parameter Q∈(0,1), and Q is obtained by the following formula:

Figure BDA0003661864230000042
Figure BDA0003661864230000042

Figure BDA0003661864230000043
Figure BDA0003661864230000043

其中cosWin0是最原始的余弦窗，W×H,W＝H为余弦窗尺寸，w×h,w≥h为目标大小；where cosWin 0 is the original cosine window, W×H, W=H is the cosine window size, and w×h, w≥h is the target size;

步骤2-3，每次跟踪目标尺度更新之后，获取尺度更新因子Sscale，再次更新余弦窗的削峰位置Qscale＝Q×Sscale，使模型尺度自适应；之后利用掩膜m，得到xo＝m⊙xg，代表着单纯的目标特征，得到xg和xo两种特征集；尺度自适应余弦窗的具体模型为：Step 2-3, after each tracking target scale update, obtain the scale update factor S scale , update the peak clipping position of the cosine window again Q scale = Q×S scale , so that the model scale is adaptive; then use the mask m to obtain x o = m⊙x g , which represents a simple target feature, and two feature sets x g and x o are obtained; the specific model of the scale adaptive cosine window is:

Figure BDA0003661864230000044
Figure BDA0003661864230000044

上式中cosWin0是最原始的余弦窗，Qscale为余弦窗削峰的位置，Sscale为尺度因子。In the above formula, cosWin 0 is the most primitive cosine window, Q scale is the position of the peak clipping of the cosine window, and S scale is the scale factor.

进一步地，步骤3包括如下分步骤：Further, step 3 includes the following sub-steps:

步骤3-1，滤波器的目标函数为：Step 3-1, the objective function of the filter is:

Figure BDA0003661864230000051
Figure BDA0003661864230000051

其中，P是样本裁剪矩阵；

Figure BDA0003661864230000052
代表的是第c通道的目标信息滤波器或者背景信息滤波器；cosWin是之前中提出的随着目标尺度因子变化的余弦窗函数，用于对训练样本数据进行预处理；y是理想高斯模型；ht和ht-1表示当前帧和前一帧的滤波器，M矩阵表示两种滤波器之间的关系，λ是滤波器正则化项的约束参数，μ是两个滤波器互约束项的约束参数；where P is the sample cropping matrix;
Figure BDA0003661864230000052
Represents the target information filter or background information filter of the c-th channel; cosWin is the cosine window function proposed in the previous section that changes with the target scale factor, which is used to preprocess the training sample data; y is the ideal Gaussian model; h t and h t-1 represent the filters of the current frame and the previous frame, the M matrix represents the relationship between the two filters, λ is the constraint parameter of the regularization term of the filter, and μ is the mutual constraint term of the two filters the constraint parameters;

步骤3-2，对于整体目标函数来说，因为k∈{g,o}，再加上最后的两者互约束，所以目标函数可以看成7部分累加，第1、2和第4、5部分是多了裁剪矩阵的常规线性模型，最小二乘加上正则项，正则项的目的是为了防止滤波器过拟合；第3、6部分是两种滤波器的自约束，可以有效防止滤波器的快速退化；第7部分是两种滤波器的互约束，在训练期间相互绑定，使得两种滤波器的判别能力更加强大；Step 3-2, for the overall objective function, because k∈{g,o}, plus the last two mutual constraints, the objective function can be regarded as the accumulation of 7 parts, the 1st, 2nd and the 4th, 5th. Part of it is a conventional linear model with more clipping matrices, least squares plus a regular term, the purpose of the regular term is to prevent the filter from overfitting; Parts 3 and 6 are the self-constraints of the two filters, which can effectively prevent filtering The fast degradation of the filter; the seventh part is the mutual constraint of the two filters, which are bound to each other during training, making the discriminative ability of the two filters more powerful;

步骤3-3，通过ADMM迭代算法进行求解；因为cosWin是对样本的预处理，所以迭代时可以将其忽略。对于已知的ho和M，对hg进行ADMM迭代最优解；使用增广拉格朗日法，引入松弛变量

Figure BDA0003661864230000053
PT是裁剪矩阵P的转置，IN是N×N的单位矩阵；目标函数的增广拉格朗日形式表示为：Step 3-3, solve by ADMM iterative algorithm; because cosWin is the preprocessing of samples, it can be ignored during iteration. For known h o and M, perform ADMM iterative optimal solution for h g ; use the augmented Lagrangian method to introduce slack variables
Figure BDA0003661864230000053
P T is the transpose of the clipping matrix P, and I N is an N×N identity matrix; the augmented Lagrangian form of the objective function is expressed as:

Figure BDA0003661864230000061
Figure BDA0003661864230000061

其中

Figure BDA0003661864230000062
是拉格朗日向量，γ是惩罚因子；采用ADMM方法，通过迭代的方式将上式转化为下面三个子问题：in
Figure BDA0003661864230000062
is the Lagrangian vector, and γ is the penalty factor; using the ADMM method, the above formula is transformed into the following three sub-problems through iteration:

Figure BDA0003661864230000063
Figure BDA0003661864230000063

对于

Figure BDA0003661864230000064
子问题，通过一阶导数求得hg的闭式解：for
Figure BDA0003661864230000064
Subproblem, the closed-form solution of h g is obtained by the first derivative:

Figure BDA0003661864230000065
Figure BDA0003661864230000065

对于ξ*子问题，需要先将其转换到频域进行进一步的计算：For the ξ * subproblem, it needs to be converted to the frequency domain for further computation:

Figure BDA0003661864230000066
Figure BDA0003661864230000066

上式分解为T个子问题,T＝42表示特征的维度，设每个子问题为

Figure BDA0003661864230000067
得：The above formula is decomposed into T sub-problems, T=42 represents the dimension of the feature, and each sub-problem is set as
Figure BDA0003661864230000067
have to:

Figure BDA0003661864230000068
Figure BDA0003661864230000068

对上式求导得：Derive the above formula to get:

Figure BDA0003661864230000071
Figure BDA0003661864230000071

使用Sherman-Morrison方程来优化求解逆矩阵，得：Using the Sherman-Morrison equation to optimally solve the inverse matrix, we get:

Figure BDA0003661864230000072
Figure BDA0003661864230000072

其中

Figure BDA0003661864230000073
为标量；in
Figure BDA0003661864230000073
is a scalar;

hg和ho的迭代过程相同，M矩阵的迭代为：

Figure BDA0003661864230000074
The iterative process of h g and h o is the same, and the iteration of the M matrix is:
Figure BDA0003661864230000074

进一步地，步骤4中，目标响应图

Figure BDA0003661864230000075
表示为：Further, in step 4, the target response graph
Figure BDA0003661864230000075
Expressed as:

Figure BDA0003661864230000076
Figure BDA0003661864230000076

其中，

Figure BDA0003661864230000077
代表给定数据的傅里叶域的对应量，
Figure BDA0003661864230000078
代表反傅里叶变换之后的响应图，D代表滤波器的维数，
Figure BDA0003661864230000079
和
Figure BDA00036618642300000710
是第f帧的两个滤波器，
Figure BDA00036618642300000711
表示从f+1帧中提取的搜索区域样本的第c通道特征，
Figure BDA00036618642300000712
是经过掩膜处理的
Figure BDA00036618642300000713
ρ是一个控制由
Figure BDA00036618642300000714
和
Figure BDA00036618642300000715
产生的两个响应图的权重参数。in,
Figure BDA0003661864230000077
represents the corresponding quantity in the Fourier domain of the given data,
Figure BDA0003661864230000078
represents the response map after inverse Fourier transform, D represents the dimension of the filter,
Figure BDA0003661864230000079
and
Figure BDA00036618642300000710
are the two filters of frame f,
Figure BDA00036618642300000711
represents the c-th channel feature of the search area sample extracted from frame f+1,
Figure BDA00036618642300000712
is masked
Figure BDA00036618642300000713
ρ is a control by
Figure BDA00036618642300000714
and
Figure BDA00036618642300000715
Weight parameter for the two response graphs produced.

与现有技术相比，本发明具有以下有益效果：Compared with the prior art, the present invention has the following beneficial effects:

(1)发明了一种尺度自适应余弦窗。分析了现存余弦窗的不足，利用目标初始大小对余弦窗进行削峰处理。另外加上考虑目标的尺度变化，通过尺度变化因子，对余弦窗削峰的位置进行动态调整，提高了跟踪算法在目标发生尺度变化场景下的跟踪成功率。(1) A scale-adaptive cosine window is invented. The insufficiency of the existing cosine window is analyzed, and the cosine window is clipped by the initial size of the target. In addition, considering the scale change of the target, the position of the peak clipping of the cosine window is dynamically adjusted through the scale change factor, which improves the tracking success rate of the tracking algorithm in the scene where the target scale changes.

(2)改进ADTrack算法，引入了训练目标和背景两种滤波器之间的自约束和互约束，进一步优化了目标函数，优化了现存跟踪器的跟踪漂移和模板快速退化现象，并且对新的目标函数进行了ADMM的优化推导。(2) Improve the ADTrack algorithm, introduce the self-constraint and mutual constraint between the training target and the background filter, further optimize the objective function, optimize the tracking drift of the existing tracker and the rapid template degradation phenomenon, and for the new The objective function is optimized and deduced by ADMM.

附图说明Description of drawings

图1为本发明实施例中的尺度自适应余弦窗的构建。FIG. 1 shows the construction of a scale-adaptive cosine window in an embodiment of the present invention.

图2为循环矩阵和裁剪矩阵的示范。Figure 2 is an example of a circulant matrix and a clipping matrix.

图3为本发明实施例中的算法模型总体框架。FIG. 3 is an overall framework of an algorithm model in an embodiment of the present invention.

图4为本发明实施例中的TC128数据集tiger1视频序列的跟踪结果。FIG. 4 is the tracking result of the tiger1 video sequence of the TC128 data set in the embodiment of the present invention.

图5为本发明实施例中的基于TC128数据集对比近几年算法和本发明方法的总精确率结果示意图。FIG. 5 is a schematic diagram of the overall accuracy result of comparing the algorithm in recent years and the method of the present invention based on the TC128 data set in the embodiment of the present invention.

图6为本发明实施例中的基于TC128数据集对比近几年算法和本发明方法的总成功率结果示意图。FIG. 6 is a schematic diagram showing the result of comparing the total success rate of the algorithm in recent years and the method of the present invention based on the TC128 data set in the embodiment of the present invention.

具体实施方式Detailed ways

下面结合说明书附图对本发明的技术方案做进一步的详细说明。The technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.

本发明是一种改进余弦窗的相关滤波目标跟踪方法，并分别对目标和背景信息进行滤波训练。目前目标跟踪领域还存在着很多问题，如目标尺度变化、目标遮挡等。针对于目标尺度变化，本发明构建一种尺度自适应余弦窗，对余弦窗进行削峰，并且适时更新，再使用目标和目标背景两种滤波器，结合两者的自约束和互约束，对相关滤波目标跟踪模型进行改进，以此提高跟踪性能。The present invention is an improved cosine window correlation filtering target tracking method, and the target and background information are filtered and trained respectively. At present, there are still many problems in the field of target tracking, such as target scale change, target occlusion and so on. Aiming at the change of target scale, the present invention constructs a scale adaptive cosine window, clips the peak of the cosine window, and updates it in time, and then uses two filters of target and target background, combined with the self-constraint and mutual constraint of the two, to The correlation filtering target tracking model is improved to improve the tracking performance.

本发明的技术方案主要内容如下：针对相关滤波目标跟踪算法中余弦窗会增加样本的纹理、污染样本的问题，构建一种尺度自适应余弦窗函数，构建方法为先基于目标的基准大小，对余弦窗模型进行削峰，然后利用DSST算法尺度检测模块中的尺度因子Sscale，适时更新削峰的余弦窗模型，使之尺度自适应，具体效果见附图1。本发明总的跟踪模型基于ADTrack算法，具体效果见附图3，除了上述的尺度自适应余弦窗外，还在ADTrack基础模型上优化了滤波器的自约束

Figure BDA0003661864230000091
和互约束
Figure BDA0003661864230000092
并重新推导模型的ADMM迭代。The main contents of the technical solution of the present invention are as follows: in view of the problem that the cosine window in the correlation filtering target tracking algorithm will increase the texture of the sample and contaminate the sample, a scale adaptive cosine window function is constructed, and the construction method is based on the benchmark size of the target. Peak clipping is performed on the cosine window model, and then the scale factor S scale in the scale detection module of the DSST algorithm is used to update the peak clipping cosine window model in a timely manner to make it scale adaptive. The specific effect is shown in Figure 1. The overall tracking model of the present invention is based on the ADTrack algorithm, and the specific effect is shown in Figure 3. In addition to the above-mentioned scale adaptive cosine window, the self-constraint of the filter is optimized on the ADTrack basic model.
Figure BDA0003661864230000091
and mutual restraint
Figure BDA0003661864230000092
And re-derive the ADMM iteration of the model.

本发明具体步骤如下：The concrete steps of the present invention are as follows:

Step1：预处理阶段，利用阈值τ来判断是否处于暗场景，τ的最佳值通过调参获取。若判断处于暗场景，则对视频序列进行增强，目的是提高黑夜状态下的目标跟踪的准确性和鲁棒性，并且得到掩膜m，目的是用于后续两个滤波器的训练。Step1: In the preprocessing stage, the threshold τ is used to judge whether it is in a dark scene, and the optimal value of τ is obtained by adjusting the parameters. If it is judged that it is in a dark scene, the video sequence is enhanced to improve the accuracy and robustness of target tracking in the dark state, and a mask m is obtained, which is used for the training of the subsequent two filters.

假设给定彩色图像

Figure BDA0003661864230000093
对图像的RBG三通道进行光度整合：Assume given a color image
Figure BDA0003661864230000093
Photometric integration of the RBG three channels of the image:

Figure BDA0003661864230000094
Figure BDA0003661864230000094

其中Ψm(I(x,y))代表着m通道出图像的像素值，并且αR+αG+αB＝1，可以理解为将彩色图像转换为单通道图像。再对亮度进行对数平均处理：where Ψ m (I(x,y)) represents the pixel value of the m-channel output image, and α R + α G + α B =1, which can be understood as converting a color image into a single-channel image. Then logarithmically average the brightness:

Figure BDA0003661864230000095
Figure BDA0003661864230000095

其中δ是一个很小的值，目的是防出现log(0)的错误情况，这个

Figure BDA0003661864230000096
就代表了当前图像的对数平均场景亮度，再通过引入阈值τ，就可以判断当前图像是否处于暗场景，
Figure BDA0003661864230000097
小于阈值即为暗场景，表示为S(I)。where δ is a small value, the purpose is to prevent log(0) error conditions, this
Figure BDA0003661864230000096
It represents the logarithmic average scene brightness of the current image, and then by introducing a threshold τ, it can be judged whether the current image is in a dark scene,
Figure BDA0003661864230000097
Less than the threshold is a dark scene, denoted as S(I).

之后对图像进行增强，利用之前获取的图像亮度V(x,y,I)和图像对数平均亮度

Figure BDA0003661864230000098
图像全局的增强矩阵可以表示为：After that, the image is enhanced, using the previously acquired image brightness V(x,y,I) and the image logarithmic average brightness
Figure BDA0003661864230000098
The enhancement matrix of the image global can be expressed as:

Figure BDA0003661864230000101
Figure BDA0003661864230000101

其中Vmax(I)代表图像亮度V(x,y,I)中的最大值，然后可以对图像的三通道进行增强处理：Where V max (I) represents the maximum value in the image brightness V(x, y, I), and then the three channels of the image can be enhanced:

Figure BDA0003661864230000102
Figure BDA0003661864230000102

其中Ie表示增强过的图像，Ψm(Ie(x,y))也就表示增强过的图像在m通道(x,y)位置处的像素值。由上述增强过后的图像，也就很容易得到图像的增强部分信息：where I e represents the enhanced image, and Ψ m (I e (x, y)) also represents the pixel value of the enhanced image at the position of m channel (x, y). From the enhanced image above, it is easy to get the enhanced part of the image:

E(I)＝V(I)-V(Ie)E(I)=V(I)-V(I e )

最后通过E(I)得到E(I)的平均值μ和标准差σ，由此就可以得到全局的掩膜mg：Finally, the average μ and standard deviation σ of E(I) are obtained through E(I), so that the global mask m g can be obtained:

Figure BDA0003661864230000103
Figure BDA0003661864230000103

参考BACF算法中的裁剪矩阵P，对mg进行裁剪，就可以得到预期的掩膜m＝mg⊙P,P∈Rw×h用于提取样本内的目标大小信息。Referring to the cropping matrix P in the BACF algorithm, by cropping m g , the expected mask m=m g ⊙P, P∈R w×h is used to extract the target size information in the sample.

Step2：提取特征阶段。通过余弦窗预处理、循环矩阵、中心裁剪方法获取大量训练样本，余弦窗处理即为在样本上直接点乘余弦窗函数，循环矩阵和中心裁剪的操作见附图2。再对获取的样本进行特征提取，包括灰度信息、颜色信息、梯度信息等，得到特征集合xg。考虑到原始余弦窗的不足，本发明对余弦窗进行改进，具体步骤见附图1，先利用目标初始大小对余弦窗削峰处理，削峰位置为参数Q∈(0,1)，Q可由下面的公式得到：Step2: Extract feature stage. A large number of training samples are obtained through the cosine window preprocessing, circulant matrix, and center cropping methods. The cosine window processing is to directly multiply the cosine window function on the sample. The operations of the circulant matrix and the center cropping are shown in Figure 2. Then perform feature extraction on the obtained samples, including grayscale information, color information, gradient information, etc., to obtain a feature set x g . Taking into account the shortcomings of the original cosine window, the present invention improves the cosine window. The specific steps are shown in Figure 1. First, the initial size of the target is used to clip the peak of the cosine window. The peak clipping position is the parameter Q∈(0,1), and Q can be The following formula is obtained:

Figure BDA0003661864230000111
Figure BDA0003661864230000111

Figure BDA0003661864230000112
Figure BDA0003661864230000112

其中cosWin0是最原始的余弦窗，W×H,W＝H为余弦窗尺寸，w×h,w≥h为目标大小。where cosWin 0 is the original cosine window, W×H, W=H is the cosine window size, and w×h, w≥h is the target size.

之后每次跟踪目标尺度更新之后，获取尺度更新因子Sscale，再次更新余弦窗的削峰位置Qscale＝Q×Sscale，使模型尺度自适应。之后利用掩膜m，得到xo＝m⊙xg，代表着单纯的目标特征，至此就得到了xg和xo两种特征集。尺度自适应余弦窗的具体模型为：After each tracking target scale update, the scale update factor S scale is obtained, and the peak clipping position Q scale =Q×S scale of the cosine window is updated again to make the model scale adaptive. Then, using the mask m, x o =m⊙x g is obtained, which represents the pure target feature, and thus two feature sets of x g and x o are obtained. The specific model of the scale-adaptive cosine window is:

Figure BDA0003661864230000113
Figure BDA0003661864230000113

上式中cosWin0是最原始的余弦窗，Qscale为余弦窗削峰的位置，Sscale为尺度因子，W×H,W＝H为余弦窗尺寸，w×h,w≥h为目标大小。In the above formula, cosWin 0 is the most original cosine window, Q scale is the position of the cosine window clipping peak, S scale is the scale factor, W×H, W=H is the cosine window size, w×h, w≥h is the target size .

Step3：训练阶段。利用Step2得到的xg和xo特征集训练出下一帧的两个滤波器：hg和ho，hg是用于训练背景信息的滤波器，ho是用于训练目标信息的滤波器。本算法模型基于ADTrack模型改良，具体模型见附图3，模型的目标函数可以表示为：Step3: Training phase. Use the x g and x o feature sets obtained in Step2 to train two filters for the next frame: h g and h o , h g is the filter for training background information, and h o is the filter for training target information device. This algorithm model is improved based on the ADTrack model. The specific model is shown in Figure 3. The objective function of the model can be expressed as:

Figure BDA0003661864230000114
Figure BDA0003661864230000114

其中，P是ADTrack基准算法BACF算法中的样本裁剪矩阵；hk c代表的是第c通道的目标信息或者背景信息滤波器；cosWin是之前中提出的随着目标尺度因子变化的余弦窗函数，用于对训练样本数据进行预处理；y是理想高斯模型；ht和ht-1表示当前帧和前一帧的滤波器，M矩阵表示两种滤波器之间的关系，λ是滤波器正则化项的约束参数，μ是两个滤波器互约束项的约束参数。Among them, P is the sample cropping matrix in the ADTrack benchmark algorithm BACF algorithm; h k c represents the target information or background information filter of the c-th channel; cosWin is the cosine window function proposed in the previous section with the change of the target scale factor, It is used to preprocess the training sample data; y is the ideal Gaussian model; h t and h t-1 represent the filters of the current frame and the previous frame, the M matrix represents the relationship between the two filters, and λ is the filter The constraint parameter of the regularization term, μ is the constraint parameter of the mutual constraint terms of the two filters.

对于整体目标函数来说，因为k∈{g,o}，再加上最后的两者互约束，所以目标函数可以看成7部分累加，第1、2和第4、5部分是多了裁剪矩阵的常规线性模型，最小二乘加上正则项，正则项的目的是为了防止滤波器过拟合；第3、6部分是两种滤波器的自约束，可以有效防止滤波器的快速退化；第7部分是两种滤波器的互约束，在训练期间相互绑定，使得两种滤波器的判别能力更加强大。For the overall objective function, because k∈{g,o}, plus the last two mutual constraints, the objective function can be regarded as a 7-part accumulation, and the 1st, 2nd, 4th, and 5th parts are more clipped The conventional linear model of the matrix, the least squares plus the regular term, the purpose of the regular term is to prevent the filter from overfitting; Parts 3 and 6 are the self-constraints of the two filters, which can effectively prevent the rapid degradation of the filter; Part 7 is the mutual constraint of the two filters, which are bound to each other during training, making the discriminative power of the two filters more powerful.

接下来通过ADMM迭代算法进行求解。在迭代的过程中，可以先假定ho和M是已知的，对hg进行ADMM迭代最优解。由于W余弦窗函数仅仅是对样本的预处理，所以进行迭代的时候可以将其忽略。使用增广拉格朗日法，引入松弛变量

Figure BDA0003661864230000121
PT是裁剪矩阵P的转置，IN是N×N的单位矩阵；目标函数的增广拉格朗日形式表示为：Next, it is solved by the ADMM iterative algorithm. In the iterative process, it can be assumed that h o and M are known, and the optimal solution of ADMM iteration is performed on h g . Since the W cosine window function is only a preprocessing of the samples, it can be ignored when iterating. Using Augmented Lagrangian Method, Introducing Slack Variables
Figure BDA0003661864230000121
P T is the transpose of the clipping matrix P, and I N is an N×N identity matrix; the augmented Lagrangian form of the objective function is expressed as:

Figure BDA0003661864230000122
Figure BDA0003661864230000122

其中

Figure BDA0003661864230000123
是拉格朗日向量，γ是惩罚因子。采用ADMM方法，可以通过迭代的方式将上式转化为下面三个子问题：in
Figure BDA0003661864230000123
is the Lagrangian vector and γ is the penalty factor. Using the ADMM method, the above equation can be transformed into the following three sub-problems in an iterative manner:

Figure BDA0003661864230000131
Figure BDA0003661864230000131

对于

Figure BDA0003661864230000132
子问题，可以直接通过一阶导数求得hg的闭式解：for
Figure BDA0003661864230000132
subproblem, the closed-form solution of h g can be obtained directly through the first derivative:

Figure BDA0003661864230000133
Figure BDA0003661864230000133

对于子问题ξ*，需要先将其转换到频域进行进一步的计算：For the subproblem ξ * , it needs to be converted to the frequency domain for further computation:

Figure BDA0003661864230000134
Figure BDA0003661864230000134

上式可以分解为T个子问题,T＝42表示特征的维度，设每个子问题为

Figure BDA0003661864230000135
可得：The above formula can be decomposed into T sub-problems, T=42 represents the dimension of the feature, and each sub-problem is set as
Figure BDA0003661864230000135
Available:

Figure BDA0003661864230000136
Figure BDA0003661864230000136

对上式求导可得：Derivation of the above formula can be obtained:

Figure BDA0003661864230000137
Figure BDA0003661864230000137

由于出现了矩阵除法，计算量大，需要用到Sherman-Morrison方程来优化求解逆矩阵，得：Due to the occurrence of matrix division, the amount of calculation is large, and the Sherman-Morrison equation needs to be used to optimize the solution of the inverse matrix, so that:

Figure BDA0003661864230000138
Figure BDA0003661864230000138

其中

Figure BDA0003661864230000139
为标量。in
Figure BDA0003661864230000139
is a scalar.

hg和ho的迭代过程大致相同，不再赘述，M矩阵的迭代为：

Figure BDA0003661864230000141
The iterative processes of h g and h o are roughly the same, and will not be repeated here. The iteration of the M matrix is:
Figure BDA0003661864230000141

Step4：检测阶段。目标响应图

Figure BDA0003661864230000142
可以表示为:Step4: detection stage. target response graph
Figure BDA0003661864230000142
It can be expressed as:

Figure BDA0003661864230000143
Figure BDA0003661864230000143

其中，

Figure BDA0003661864230000144
代表着给定数据的傅里叶域的对应量，
Figure BDA0003661864230000145
代表着反傅里叶变换之后的响应图，D对应着滤波器的维数，
Figure BDA0003661864230000146
和
Figure BDA0003661864230000147
是第f帧的两个滤波器，
Figure BDA0003661864230000148
表示从f+1帧中提取的搜索区域样本的第c通道特征，
Figure BDA0003661864230000149
是经过掩膜处理的
Figure BDA00036618642300001410
ρ是一个控制由
Figure BDA00036618642300001411
和
Figure BDA00036618642300001412
产生的两个响应图的权重参数。最后根据响应图
Figure BDA00036618642300001413
最大值就可以确定目标的位置信息。in,
Figure BDA0003661864230000144
represents the corresponding quantity in the Fourier domain of the given data,
Figure BDA0003661864230000145
represents the response map after inverse Fourier transform, D corresponds to the dimension of the filter,
Figure BDA0003661864230000146
and
Figure BDA0003661864230000147
are the two filters of frame f,
Figure BDA0003661864230000148
represents the c-th channel feature of the search area sample extracted from frame f+1,
Figure BDA0003661864230000149
is masked
Figure BDA00036618642300001410
ρ is a control by
Figure BDA00036618642300001411
and
Figure BDA00036618642300001412
Weight parameter for the two response graphs produced. Finally according to the response graph
Figure BDA00036618642300001413
The maximum value can determine the location information of the target.

本发明通过尺度自适应余弦窗、滤波器的自约束和互约束，提高了目标跟踪算法在尺度变化场景下的跟踪性能，减少了跟踪过程中的模板漂移问题，大大提高了算法的准确率和成功率，如附图5-6和表1所示，本方法跟踪成功率排名第一，精确率和AutoTrack算法并列第一。The invention improves the tracking performance of the target tracking algorithm in the scale changing scene through the self-constraint and mutual constraint of the scale-adaptive cosine window and the filter, reduces the template drift problem in the tracking process, and greatly improves the accuracy and efficiency of the algorithm. Success rate, as shown in Figures 5-6 and Table 1, the tracking success rate of this method ranks first, and the accuracy rate is tied for the first place with the AutoTrack algorithm.

表1 各算法在TC128数据集下的总精确率和总成功率Table 1 The total accuracy and total success rate of each algorithm under the TC128 dataset

OursOurs CPCFCPCF ADTrackADTrack AutoTrackAutoTrack BACFBACF KCFKCF SRDCFSRDCF BiCFBiCF PrecisionPrecision 0.7020.702 0.6970.697 0.6890.689 0.7020.702 0.6440.644 0.5440.544 0.6440.644 0.6410.641 SuccessSuccess 0.6490.649 0.6170.617 0.6190.619 0.6290.629 0.6100.610 0.4540.454 0.5840.584 0.5590.559

以上所述仅为本发明的较佳实施方式，本发明的保护范围并不以上述实施方式为限，但凡本领域普通技术人员根据本发明所揭示内容所作的等效修饰或变化，皆应纳入权利要求书中记载的保护范围内。The above descriptions are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, but any equivalent modifications or changes made by those of ordinary skill in the art based on the contents disclosed in the present invention should be included in the within the scope of protection described in the claims.

Claims (6)

1.一种改进余弦窗的相关滤波目标跟踪方法，其特征在于：所述方法包括如下步骤：1. a correlation filtering target tracking method of improving cosine window, is characterized in that: described method comprises the steps: 步骤1，预处理阶段；判断是否处于暗场景，若判断处于暗场景，则对视频序列进行增强，并计算得到掩膜m，用于后续滤波器的训练；Step 1, preprocessing stage; judge whether it is in a dark scene, if it is judged to be in a dark scene, enhance the video sequence, and calculate the mask m, which is used for the training of subsequent filters; 步骤2，提取特征阶段；基于训练样本得到特征集xg，通过削峰处理后的余弦窗实现跟踪目标尺度更新的自适应，基于掩膜m得到目标特征集xo，得到尺度自适应余弦窗模型；Step 2, feature extraction stage; obtain feature set x g based on training samples, realize adaptive tracking target scale update through cosine window after peak clipping processing, obtain target feature set x o based on mask m, and obtain scale adaptive cosine window Model; 步骤3，训练阶段；利用步骤2得到的特征集xg和xo训练出下一帧的两个滤波器hg和ho，其中hg是用于训练背景信息的滤波器，ho是用于训练目标信息的滤波器；Step 3, training phase; use the feature sets x g and x o obtained in step 2 to train two filters h g and ho for the next frame, where h g is the filter for training background information, and ho is Filters for training target information; 步骤4，检测阶段；基于训练后的滤波器和提取的样本通道特征，得到目标响应图，通过目标响应图的最大值确定目标的位置信息。Step 4, detection stage; based on the trained filter and the extracted sample channel features, a target response map is obtained, and the position information of the target is determined by the maximum value of the target response map. 2.根据权利要求1所述的一种改进余弦窗的相关滤波目标跟踪方法，其特征在于：步骤1包括如下分步骤：2. the correlation filtering target tracking method of a kind of improved cosine window according to claim 1 is characterized in that: step 1 comprises the following steps: 步骤1-1，假设给定彩色图像
Figure FDA0003661864220000011
对图像的RBG三通道进行光度整合：Step 1-1, assuming a color image is given
Figure FDA0003661864220000011
Photometric integration of the RBG three channels of the image:
Figure FDA0003661864220000012
Figure FDA0003661864220000012
 其中Ψm(I(x,y))代表着m通道出图像的像素值，并且αR+αG+αB＝1，将彩色图像转换为单通道图像；再对亮度进行对数平均处理：Where Ψ m (I(x, y)) represents the pixel value of the m-channel output image, and α R + α G + α B =1, convert the color image to a single-channel image; then perform logarithmic average processing on the brightness :
Figure FDA0003661864220000013
Figure FDA0003661864220000013
 其中δ为常数，用于防止log(0)的出现，
Figure FDA0003661864220000014
代表当前图像的对数平均场景亮度，再通过引入阈值τ判断当前图像是否处于暗场景，
Figure FDA0003661864220000015
小于阈值即为暗场景，表示为S(I)；where δ is a constant to prevent the occurrence of log(0),
Figure FDA0003661864220000014
represents the logarithmic average scene brightness of the current image, and then judges whether the current image is in a dark scene by introducing a threshold τ,
Figure FDA0003661864220000015
Less than the threshold is a dark scene, denoted as S(I); 步骤1-2，对图像进行增强，利用之前获取的图像亮度V(x,y,I)和图像对数平均亮度
Figure FDA0003661864220000021
图像全局的增强矩阵表示为：Step 1-2, enhance the image, use the image brightness V(x, y, I) obtained before and the logarithmic average brightness of the image
Figure FDA0003661864220000021
The enhancement matrix of the image global is expressed as:
Figure FDA0003661864220000022
Figure FDA0003661864220000022
 其中Vmax(I)代表图像亮度V(x,y,I)中的最大值，然后对图像的三通道进行增强处理：Where V max (I) represents the maximum value in the image brightness V(x, y, I), and then the three channels of the image are enhanced:
Figure FDA0003661864220000023
Figure FDA0003661864220000023
 其中Ie表示增强过的图像，Ψm(Ie(x,y))表示增强过的图像在m通道(x,y)位置处的像素值；由上述增强过后的图像，得到图像的增强部分信息：where I e represents the enhanced image, Ψ m (I e (x, y)) represents the pixel value of the enhanced image at the position of m channel (x, y); from the enhanced image above, the enhanced image is obtained Partial information: E(I)＝V(I)-V(Ie)E(I)=V(I)-V(I e ) 步骤1-3，通过E(I)得到E(I)的平均值μ和标准差σ，由此得到全局的掩膜mg：Steps 1-3, obtain the average μ and standard deviation σ of E(I) through E(I), thereby obtaining the global mask m g :
Figure FDA0003661864220000024
Figure FDA0003661864220000024
 通过裁剪矩阵P对mg进行裁剪，得到预期的掩膜m＝mg⊙P，P∈Rw×h，用于提取样本内的目标大小信息。The m g is cropped by the crop matrix P, and the expected mask m=m g ⊙P, P∈R w×h is obtained, which is used to extract the target size information within the sample. 3.根据权利要求1所述的一种改进余弦窗的相关滤波目标跟踪方法，其特征在于：步骤2包括如下分步骤：3. the correlation filtering target tracking method of a kind of improved cosine window according to claim 1, is characterized in that: step 2 comprises the following steps: 步骤2-1，通过余弦窗预处理、循环矩阵、中心裁剪方法获取大量训练样本，余弦窗处理即为在样本上直接点乘余弦窗函数，对获取的样本进行特征提取，包括灰度信息、颜色信息、梯度信息等，得到特征集合xg；Step 2-1: Obtain a large number of training samples through cosine window preprocessing, circulant matrix, and center cropping methods. Cosine window processing is to directly multiply the cosine window function on the samples, and perform feature extraction on the obtained samples, including grayscale information, Color information, gradient information, etc., to obtain a feature set x g ; 步骤2-2，利用目标初始大小对余弦窗削峰处理，削峰位置为参数Q∈(0,1)，Q由下面的公式得到：Step 2-2, use the initial size of the target to clip the peak of the cosine window, the clipping position is the parameter Q∈(0,1), and Q is obtained by the following formula:
Figure FDA0003661864220000031
Figure FDA0003661864220000031
Figure FDA0003661864220000032
Figure FDA0003661864220000032
 其中cosWin0是最原始的余弦窗，W×H,W＝H为余弦窗尺寸，w×h,w≥h为目标大小；where cosWin 0 is the original cosine window, W×H, W=H is the cosine window size, and w×h, w≥h is the target size; 步骤2-3，每次跟踪目标尺度更新之后，获取尺度更新因子Sscale，再次更新余弦窗的削峰位置Qscale＝Q×Sscale，使模型尺度自适应；之后利用掩膜m，得到xo＝m⊙xg，代表着单纯的目标特征，得到xg和xo两种特征集；尺度自适应余弦窗的具体模型为：Step 2-3, after each tracking target scale update, obtain the scale update factor S scale , update the peak clipping position of the cosine window again Q scale = Q×S scale , so that the model scale is adaptive; then use the mask m to obtain x o = m⊙x g , which represents a simple target feature, and two feature sets x g and x o are obtained; the specific model of the scale adaptive cosine window is:
Figure FDA0003661864220000033
Figure FDA0003661864220000033
 上式中cosWin0是最原始的余弦窗，Qscale为余弦窗削峰的位置，Sscale为尺度因子。In the above formula, cosWin 0 is the most primitive cosine window, Q scale is the position of the peak clipping of the cosine window, and S scale is the scale factor. 4.根据权利要求1所述的一种改进余弦窗的相关滤波目标跟踪方法，其特征在于：步骤3包括如下分步骤：4. the correlation filtering target tracking method of a kind of improved cosine window according to claim 1, is characterized in that: step 3 comprises the following steps: 步骤3-1，滤波器的目标函数为：Step 3-1, the objective function of the filter is:
Figure FDA0003661864220000034
Figure FDA0003661864220000034
 其中，P是样本裁剪矩阵；
Figure FDA0003661864220000035
代表的是第c通道的目标信息滤波器或者背景信息滤波器；cosWin是之前中提出的随着目标尺度因子变化的余弦窗函数，用于对训练样本数据进行预处理；y是理想高斯模型；ht和ht-1表示当前帧和前一帧的滤波器，M矩阵表示两种滤波器之间的关系，λ是滤波器正则化项的约束参数，μ是两个滤波器互约束项的约束参数；where P is the sample cropping matrix;
Figure FDA0003661864220000035
Represents the target information filter or background information filter of the c-th channel; cosWin is the cosine window function proposed in the previous section that changes with the target scale factor, which is used to preprocess the training sample data; y is the ideal Gaussian model; h t and h t-1 represent the filters of the current frame and the previous frame, the M matrix represents the relationship between the two filters, λ is the constraint parameter of the regularization term of the filter, and μ is the mutual constraint term of the two filters the constraint parameters; 步骤3-2，对于整体目标函数来说，因为k∈{g,o}，再加上最后的两者互约束，所以目标函数看成7部分累加，第1、2和第4、5部分是多了裁剪矩阵的常规线性模型，最小二乘加上正则项，防止滤波器过拟合；第3、6部分是两种滤波器的自约束，防止滤波器的快速退化；第7部分是两种滤波器的互约束，在训练期间相互绑定，使得两种滤波器的判别能力更加强大；Step 3-2, for the overall objective function, because k∈{g,o}, plus the last two mutual constraints, the objective function is regarded as the accumulation of 7 parts, the 1st and 2nd and the 4th and 5th parts. It is a conventional linear model with more clipping matrices, least squares plus a regular term to prevent the filter from overfitting; Parts 3 and 6 are the self-constraints of the two filters to prevent rapid degradation of the filter; Part 7 is The mutual constraints of the two filters are bound to each other during training, which makes the discriminative ability of the two filters more powerful; 步骤3-3，通过ADMM迭代算法进行求解；因为cosWin是对样本的预处理，所以迭代时可以将其忽略。对于已知的ho和M，对hg进行ADMM迭代最优解；使用增广拉格朗日法，引入松弛变量
Figure FDA0003661864220000041
PT是裁剪矩阵P的转置，IN是N×N的单位矩阵。Step 3-3, solve by ADMM iterative algorithm; because cosWin is the preprocessing of samples, it can be ignored during iteration. For known h o and M, perform ADMM iterative optimal solution for h g ; use the augmented Lagrangian method to introduce slack variables
Figure FDA0003661864220000041
P T is the transpose of the clipping matrix P, and I N is an N×N identity matrix. 5.根据权利要求4所述的一种改进余弦窗的相关滤波目标跟踪方法，其特征在于：步骤3-3中，目标函数的增广拉格朗日形式表示为：5. the correlation filtering target tracking method of a kind of improved cosine window according to claim 4, is characterized in that: in step 3-3, the augmented Lagrangian form of objective function is expressed as:
Figure FDA0003661864220000042
Figure FDA0003661864220000042
 其中
Figure FDA0003661864220000043
是拉格朗日向量，γ是惩罚因子，IN是N×N的单位矩阵，FN是N×N的傅里叶矩阵；采用ADMM方法，通过迭代的方式将上式转化为下面三个子问题：in
Figure FDA0003661864220000043
is a Lagrangian vector, γ is a penalty factor, I N is an N×N unit matrix, and F N is an N×N Fourier matrix; ADMM method is used to iteratively convert the above formula into the following three subsections question:
Figure FDA0003661864220000051
Figure FDA0003661864220000051
 对于
Figure FDA0003661864220000052
子问题，通过一阶导数求得hg的闭式解：for
Figure FDA0003661864220000052
Subproblem, the closed-form solution of h g is obtained by the first derivative:
Figure FDA0003661864220000053
Figure FDA0003661864220000053
 对于ξ*子问题，需要先将其转换到频域进行进一步的计算：For the ξ * subproblem, it needs to be converted to the frequency domain for further computation:
Figure FDA0003661864220000054
Figure FDA0003661864220000054
 上式分解为T个子问题,T＝42表示特征的维度，设每个子问题为
Figure FDA0003661864220000055
得：The above formula is decomposed into T sub-problems, T=42 represents the dimension of the feature, and each sub-problem is set as
Figure FDA0003661864220000055
have to:
Figure FDA0003661864220000056
Figure FDA0003661864220000056
 对上式求导得：Derive the above formula to get:
Figure FDA0003661864220000057
Figure FDA0003661864220000057
 使用Sherman-Morrison方程来优化求解逆矩阵，得：Using the Sherman-Morrison equation to optimally solve the inverse matrix, we get:
Figure FDA0003661864220000058
Figure FDA0003661864220000058
 其中
Figure FDA0003661864220000059
为标量；in
Figure FDA0003661864220000059
is a scalar; hg和ho的迭代过程相同，M矩阵的迭代为：
Figure FDA0003661864220000061
The iterative process of h g and h o is the same, and the iteration of the M matrix is:
Figure FDA0003661864220000061
 6.根据权利要求1所述的一种改进余弦窗的相关滤波目标跟踪方法，其特征在于：步骤4中，目标响应图
Figure FDA0003661864220000062
表示为：6. the correlation filtering target tracking method of a kind of improved cosine window according to claim 1, is characterized in that: in step 4, the target response graph
Figure FDA0003661864220000062
Expressed as:
Figure FDA0003661864220000063
Figure FDA0003661864220000063
 其中，
Figure FDA0003661864220000064
代表给定数据的傅里叶域的对应量，
Figure FDA0003661864220000065
代表反傅里叶变换之后的响应图，D代表滤波器的维数，
Figure FDA0003661864220000066
和
Figure FDA0003661864220000067
是第f帧的两个滤波器，
Figure FDA0003661864220000068
表示从f+1帧中提取的搜索区域样本的第c通道特征，
Figure FDA0003661864220000069
是经过掩膜处理的
Figure FDA00036618642200000610
ρ是一个控制由
Figure FDA00036618642200000611
和
Figure FDA00036618642200000612
产生的两个响应图的权重参数。in,
Figure FDA0003661864220000064
represents the corresponding quantity in the Fourier domain of the given data,
Figure FDA0003661864220000065
represents the response map after inverse Fourier transform, D represents the dimension of the filter,
Figure FDA0003661864220000066
and
Figure FDA0003661864220000067
are the two filters of frame f,
Figure FDA0003661864220000068
represents the c-th channel feature of the search area sample extracted from frame f+1,
Figure FDA0003661864220000069
is masked
Figure FDA00036618642200000610
ρ is a control by
Figure FDA00036618642200000611
and
Figure FDA00036618642200000612
Weight parameter for the two response graphs produced.
CN202210575210.8A 2022-05-25 2022-05-25 An improved cosine window correlation filter target tracking method Active CN114897938B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210575210.8A CN114897938B (en) 2022-05-25 2022-05-25 An improved cosine window correlation filter target tracking method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210575210.8A CN114897938B (en) 2022-05-25 2022-05-25 An improved cosine window correlation filter target tracking method

Publications (2)

Publication Number Publication Date
CN114897938A true CN114897938A (en) 2022-08-12
CN114897938B CN114897938B (en) 2025-02-11

Family

ID=82725509

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210575210.8A Active CN114897938B (en) 2022-05-25 2022-05-25 An improved cosine window correlation filter target tracking method

Country Status (1)

Country Link
CN (1) CN114897938B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102496016A (en) * 2011-11-22 2012-06-13 武汉大学 Infrared target detection method based on space-time cooperation framework
CN111951298A (en) * 2020-06-25 2020-11-17 湖南大学 A target tracking method fused with time series information
US20210227132A1 (en) * 2018-05-30 2021-07-22 Arashi Vision Inc. Method for tracking target in panoramic video, and panoramic camera

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102496016A (en) * 2011-11-22 2012-06-13 武汉大学 Infrared target detection method based on space-time cooperation framework
US20210227132A1 (en) * 2018-05-30 2021-07-22 Arashi Vision Inc. Method for tracking target in panoramic video, and panoramic camera
CN111951298A (en) * 2020-06-25 2020-11-17 湖南大学 A target tracking method fused with time series information

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
王永雄等: "基于背景感知与快速尺寸判别的相关滤波跟踪算法", 数据采集与处理, no. 02, 15 March 2020 (2020-03-15) *

Also Published As

Publication number Publication date
CN114897938B (en) 2025-02-11

Similar Documents

Publication Publication Date Title
CN110210551B (en) A Visual Object Tracking Method Based on Adaptive Subject Sensitivity
CN108550161B (en) Scale self-adaptive kernel-dependent filtering rapid target tracking method
CN114037930B (en) Video action recognition method based on spatiotemporal enhancement network
CN110084221B (en) Serialized human face key point detection method with relay supervision based on deep learning
CN111192292A (en) Target tracking method based on attention mechanism and twin network and related equipment
CN109410247A (en) A kind of video tracking algorithm of multi-template and adaptive features select
CN109993137A (en) A fast face correction method based on convolutional neural network
CN111489364A (en) Medical image segmentation method based on lightweight full convolution neural network
CN111612024B (en) Feature extraction method, device, electronic equipment and computer readable storage medium
CN111639524A (en) Automatic driving image semantic segmentation optimization method
CN110866938B (en) A fully automatic video moving object segmentation method
CN113298036B (en) Method for dividing unsupervised video target
CN111460915B (en) Light weight neural network-based finger vein verification method and system
CN113420794A (en) Binaryzation Faster R-CNN citrus disease and pest identification method based on deep learning
CN113449671B (en) Pedestrian re-recognition method and device based on multi-scale multi-feature fusion
CN114445665B (en) Hyperspectral image classification method based on non-local U-shaped network enhanced by Transformer
CN111860278A (en) A human action recognition algorithm based on deep learning
CN111931722B (en) Correlated filtering tracking method combining color ratio characteristics
US12307746B1 (en) Nighttime unmanned aerial vehicle object tracking method fusing hybrid attention mechanism
CN105046202A (en) Adaptive face identification illumination processing method
CN116229154A (en) Class increment image classification method based on dynamic hybrid model
CN111612802A (en) A re-optimization training method and application based on existing image semantic segmentation model
CN114972435A (en) Object Tracking Method Based on Long-Short-Time Integrated Appearance Update Mechanism
CN114897938A (en) Improved cosine window related filtering target tracking method
CN109543684B (en) Real-time target tracking detection method and system based on full convolution neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
CB03 Change of inventor or designer information
CB03 Change of inventor or designer information

Inventor after: Deng Lizhen

Inventor after: Sun Jiawei

Inventor after: Zhu Hu

Inventor before: Sun Jiawei

Inventor before: Deng Lizhen

Inventor before: Zhu Hu

GR01 Patent grant
GR01 Patent grant