CN107590820B - Video object tracking method based on correlation filtering and intelligent device thereof - Google Patents
Video object tracking method based on correlation filtering and intelligent device thereof Download PDFInfo
- Publication number
- CN107590820B CN107590820B CN201710742685.0A CN201710742685A CN107590820B CN 107590820 B CN107590820 B CN 107590820B CN 201710742685 A CN201710742685 A CN 201710742685A CN 107590820 B CN107590820 B CN 107590820B
- Authority
- CN
- China
- Prior art keywords
- frame
- target area
- target
- obtaining
- unit
- 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.)
- Active
Links
Images
Abstract
The invention discloses a video object tracking method based on relevant filtering and an intelligent device thereof, wherein the method comprises the steps of circularly sampling a target area and the surrounding area of a target to obtain a shift sample; concatenating the shifted samples into a circulant matrix A; and performing operation updating on the circulant matrix A through discrete Fourier transform. The filtering method provided by the invention ensures the speed under the condition of ensuring higher precision by utilizing background information sampling. The method has the advantages that positive and negative samples are collected by using a circulation matrix of the area around the target, the target detector is trained by using ridge regression, and the operation of the matrix is converted into the dot multiplication of vectors through the diagonalizable property of the circulation matrix in the Fourier space, so that the operation amount is greatly reduced, and the tracking speed is greatly improved. The method has a good effect particularly under the conditions of complex extraction features and poor real-time effect. In addition, the background information is reasonably increased when the sample is adopted, the robustness of the model is enhanced, and the problems of rapid movement and shielding in the tracking process are effectively relieved.
Description
Technical Field
The invention relates to a video object tracking method based on correlation filtering and an intelligent device thereof.
Background
At present, when a tracking model is trained and updated by a traditional tracking algorithm, if features are complex, calculation overhead is high, and real-time effect is poor. When a related filtering method is adopted to realize target tracking in a traditional mode, due to the fact that the quality of samples obtained through cyclic sampling is poor, utilized background information is few, and therefore the tracking effect on rapid movement and shielding is poor.
The tracking algorithms nowadays mainly include algorithms such as machine learning and filter learning based on traditional features, and the latest tracking algorithm based on deep learning. Some template matching algorithms based on traditional characteristics need to undergo a large amount of iteration, and the calculation amount is large; although the speed of a general correlation filtering method meets the requirement, the sample quality is not high, so that the robustness of a tracking model is poor; the tracking algorithm based on deep learning is characterized in that the image passes through a convolutional neural network, so that the parameter calculation amount is extremely complex, the GPU is required to support, and the practicability is poor. Therefore, a video object tracking method based on relevant filtering and an intelligent device thereof are urgently needed to solve the technical problems that the quality of samples obtained by cyclic sampling in the traditional method is poor, and utilized background information is little, so that the tracking effect on rapid movement and shielding is poor.
Disclosure of Invention
The invention aims to solve the technical problems that the quality of a sample obtained by cyclic sampling in the traditional method is poor, and utilized background information is little, so that the tracking effect on rapid movement and shielding is poor.
The present invention provides a video object tracking method based on correlation filtering, which includes:
s101, detecting a video stream to acquire a target area frame S0;
S102, passing the target area frame S0Constructing a regression model and carrying out operation transformation;
s103, in the next frame of the video stream, a target area frame S0Obtaining a plurality of expanded search areas around;
s104, regarding the target area frame S0Extracting characteristic vectors from a plurality of the extended search areas and performing operation transformation;
s105, calculating model parameters of the filter according to the characteristic vectors;
s106, calculating a corresponding graph according to the model parameters and the characteristic vectors;
s107, acquiring a relative displacement value of the target movement through the corresponding graph;
s108, restoring a target area frame of the current frame through the relative displacement value;
s109, the target area frame in the flow S108 is used as the target area frame S of the next frame0And repeating the above steps S103 to SS109, tracking of the target object is achieved.
Furthermore, after detecting the first frame of the video, the target area frame S is obtained0;
According to the target area frame S0Size of the first, constructing the bandwidth and S0A size-proportional Gaussian shape regression model y, and performing discrete Fourier transform operation on the y to obtain
With S0As the basic search area of the next frame, obtaining 4 extended search areas S with the same rectangular size around the basic search area1,S2,S3,S4;
According to the basic search area S0And the extended search area S1,S2,S3,S4Respectively extracting the directional gradient histogram features to obtain the shift sample x of the corresponding feature vector0,x1,x2,x3,x4。
Further, the method comprises obtaining a shifted sample x of the corresponding feature vector0,x1,x2,x3,x4For each of said shifted samples, connecting by cyclic shift a circulant matrix A into a respective region0,A1,A2,A3,A4。
Further, for the circulant matrix A0,A1,A2,A3,A4Performing discrete Fourier transform to obtain the frequency domain
According toCalculate out the corresponding S0,S1,S2,S3,S4Model parameters α ═ α for region dependent filters0,α1,α2,α3,α4};
Wherein i respectively corresponds to S0,S1,S2,S3,S4Target area of λ1The learning rate value is 0.015 for the fixed parameter according to the aboveSubjecting it to complex conjugate transformation to obtain corresponding
wherein Z represents that the current frame is based on the previous frame target area frame S0Obtained shifted samples, λ2The penalty factor value is 25 for the fixed parameter;
according toObtaining the maximum point of the corresponding graph r so as to obtain the relative displacement value of the target movement;
reducing the target area frame S of the current frame according to the relative shift value0Search region S as a basis for the next frame0。
The application also provides an intelligent device, which is characterized in that the intelligent device comprises:
a detection unit for detecting a video stream acquisition target region frame S0;
A configuration arithmetic unit for passing through the target area frame S0Constructing a regression model and carrying out operation transformation;
an acquisition unit for framing S in the target area in the next frame of the video stream0Obtaining multiple extended searches aroundA cable region;
an extraction unit for extracting a target region frame S0Extracting characteristic vectors from a plurality of the extended search areas and performing operation transformation;
the calculation unit is used for calculating the model parameters of the filter according to the characteristic vectors and calculating the corresponding graph according to the model parameters and the characteristic vectors;
the acquisition unit is further used for acquiring a relative displacement value of the target movement through the corresponding graph;
the restoring unit is used for restoring a target area of the current frame through the relative displacement value;
and the repeated circulation unit is used for realizing the tracking of the target object by repeated circulation.
Furthermore, the detection unit is further configured to obtain a target area frame S after detecting the first frame of the video0;
The construction unit is also used for constructing a frame S according to the target area0Size of (d), constructing bandwidth and S0A size-proportional gaussian shape regression model y;
the computing unit is also used for carrying out discrete Fourier transform operation on the y to obtain
The acquisition unit is also used for acquiring the data by S0As the basic search area of the next frame, obtaining 4 extended search areas S with the same rectangular size around the basic search area1,S2,S3,S4;
The extraction unit is used for searching the area S according to the basis0And the extended search area S1,S2,S3,S4Respectively extracting the directional gradient histogram features to obtain the shift sample x of the corresponding feature vector0,x1,x2,x3,x4。
Furthermore, the connection unit is further configured to obtain the shift samples of the corresponding feature vectors according to the obtained shift samplesThis x0,x1,x2,x3,x4For each of said shifted samples, connecting by cyclic shift a circulant matrix A into a respective region0,A1,A2,A3,A4。
Furthermore, the computing unit is also used for aligning the circulant matrix A0,A1,A2,A3,A4Performing discrete Fourier transform to obtain the frequency domain
And is also used according toCalculating model parameters α of the relevant filter corresponding to the regions S0, S1, S2, S3 and S4 to be { α%0,α1,α2,α3,α4Wherein i corresponds to the target areas, λ, representing S0, S1, S2, S3, S4, respectively1The learning rate value is 0.015 for the fixed parameter according to the above Subjecting it to complex conjugate transformation to obtain corresponding
And also for passingAccording toA corresponding map r of the current frame is obtained, where Z denotes a shifted sample, λ, of the current frame obtained on the basis of the previous frame target area block S02The penalty factor value is 25 for the fixed parameter;
the above-mentionedAn acquisition unit further for obtainingObtaining the maximum point of the corresponding graph r so as to obtain the relative displacement value of the target movement;
the restoring unit is used for restoring the target area frame S0 of the current frame as the basic search area S of the next frame according to the relative shift value0。
The invention has the beneficial effects that:
1. the filtering method provided by the invention ensures the speed under the condition of ensuring higher precision by utilizing background information sampling. When samples are taken, background information is reasonably increased, the robustness of the model is enhanced, and the problems of rapid movement and shielding in the tracking process are effectively solved.
2. The method has the advantages that positive and negative samples are collected by using a circulation matrix of the area around the target, the target detector is trained by using ridge regression, and the operation of the matrix is converted into the dot multiplication of vectors by using the diagonalizable property of the circulation matrix in the Fourier space, so that the operation amount is greatly reduced, and the tracking speed is greatly improved. The problem of if extract the characteristic more complicated, real-time effect is poor is solved.
3. The method solves the problems of insufficient background information and poor quality of training samples when the general related filtering method realizes target tracking, obtains 4 extended search areas by expanding the basic search area in the image when extracting the sample, increases the background information for the sample and improves the sample quality. On the premise of not influencing the high-speed calculation of the filtering method in the Fourier space, the robustness of the tracking model is improved, and the method has high practicability.
4. The method ensures high efficiency in speed, enhances the sample quality by utilizing background information through a measure of expanding a search area, has higher model robustness compared with a common filtering algorithm, and can quickly and accurately track the object in the video.
Drawings
FIG. 1 is a flow chart of a method for tracking video objects based on correlation filtering according to an embodiment of the present application;
FIG. 2 is an architecture diagram of a smart device according to another embodiment of the present application;
FIG. 3 is a schematic diagram of an embodiment of the present application;
FIG. 4 is a schematic flow chart of the present application as a whole;
FIG. 5 is a diagram illustrating a first state effect of an embodiment of the present application;
FIG. 6 is a diagram illustrating a second state effect of an embodiment of the present application;
FIG. 7 is a diagram illustrating a third state effect of an embodiment of the present application;
the specific implementation mode is as follows:
the following examples are given for the purpose of clarity of the invention and are not intended to limit the embodiments of the invention. It will be apparent to those skilled in the art that other variations and modifications can be made in the invention without departing from the spirit of the invention, and it is intended to cover all such modifications and variations as fall within the true spirit of the invention.
The steps illustrated in the flow charts of the figures may be performed in a computer system such as a set of computer-executable instructions. Also, while a logical order is shown in the flow diagrams, in some cases, the steps shown or described may be performed in an order different than here.
The noun explains:
robustness means that the control system maintains stable and robust characteristics under the perturbation of certain parameters.
Ridge regression: the method is a special biased estimation regression method for collinear data analysis, and is an improved least square estimation method.
Discrete Fourier transform: is a method of signal analysis by transforming a signal from the time domain to the frequency domain.
Description of the symbols:
for example:a Discrete Fourier Transform (DFT) representing x,to representThe complex conjugate transformation of (2).Represents the Inverse Discrete Fourier Transform (IDFT) on x.
The related filtering method is to obtain all possible shifted samples through cyclic sampling in the target area frame, and connect all samples into a cyclic matrix a (a cyclic matrix a is formed after the cyclic shift of the eigenvector x obtained after the target area image sampling). The parameters of the filter (i.e. the parameters of the tracking model) are continuously updated in the tracking process, and the updating process is essentially the discrete Fourier transform of the circulant matrix, so that the calculation amount is reduced. However, the traditional correlation filtering method adopts a limited number of negative samples around the target, but does not fully acquire the background information of the target.
The following describes a single-target tracking process, and can also detect a plurality of targets by combining with a detection module to respectively complete the tracking process of each target, and the specific flow of single-target tracking is realized as follows;
as shown in fig. 1 and 4, the present application provides a video object tracking method based on correlation filtering, including:
s101, detecting a video stream to acquire a target area frame S0;
S102, passing the target area frame S0Constructing a regression model and carrying out operation transformation;
s103, in the next frame of the video stream, a target area frame S0Obtaining a plurality of expanded search areas around;
s104, regarding the target area frame S0Extracting characteristic vectors from a plurality of the extended search areas and performing operation transformation;
s105, calculating model parameters of the filter according to the characteristic vectors;
s106, calculating a corresponding graph according to the model parameters and the characteristic vectors;
s107, acquiring a relative displacement value of the target movement through the corresponding graph;
s108, restoring a target area frame of the current frame through the relative displacement value;
s109, the target area frame in the flow S108 is used as the target area frame S of the next frame0And the above steps S103 to S108 are repeated to track the target object.
First, a region in and around the target region frame is cyclically sampled to obtain shifted samples. And connecting the obtained shift samples into a cyclic matrix A, and performing operation updating on the cyclic matrix A through discrete Fourier transform. Compared with the prior art, the traditional method is lack of information acquisition of surrounding areas. The algorithm is also quite different from the present application. The speed is also ensured under the condition of ensuring higher precision, the background information is reasonably increased when a sample is taken, the robustness of the model is enhanced, and the problems of rapid movement and shielding in the tracking process are effectively solved.
In an optional embodiment of the present application, the video stream acquisition target area frame S is detected0Passing through the target area frame S0Constructing a regression model and carrying out operation transformation, and setting a target area frame S in the next frame of the video stream0Obtaining multiple extended search areas around, and framing the target area0And extracting feature vectors from a plurality of the extended search regions and performing operation transformation, wherein the method further comprises the following steps:
obtaining a target area frame S after detecting a first frame of a video0;
According to the target area frame S0Size of the first, constructing the bandwidth and S0A size-proportional Gaussian shape regression model y, and performing discrete Fourier transform operation on the y to obtain
With S0Searching as a basis for a next frameAn area which is positioned around the basic search area and obtains 4 extended search areas S with the same rectangular size1,S2,S3,S4;
According to the basic search area S0And the extended search area S1,S2,S3,S4Respectively extracting the directional gradient histogram features to obtain the shift sample x of the corresponding feature vector0,x1,x2,x3,x4。
Secondly, a target area frame S is obtained after the first frame of the video is detected0According to the target area frame S0Size of the first, constructing the bandwidth and S0A size-proportional Gaussian shape regression model y, and performing discrete Fourier transform operation on the y to obtainWith S0As the basic search area of the next frame, obtaining 4 extended search areas S with the same rectangular size around the basic search area1,S2,S3,S4Searching for the region S based on the basis0And the extended search area S1,S2,S3,S4Respectively extracting the directional gradient histogram features to obtain the shift sample x of the corresponding feature vector0,x1,x2,x3,x4. It is described in detail how to acquire modulus data values of a displaced sample and its samples. And provides a specific sample collection mode to adapt to a specific algorithm in the following process.
In an optional embodiment of the present application, the method further comprises:
from the shifted samples x that obtain the corresponding feature vector0,x1,x2,x3,x4For each of said shifted samples, connecting by cyclic shift a circulant matrix A into a respective region0,A1,A2,A3,A4. (feature direction obtained after sampling of target region image)The quantity x, after cyclic shift, constitutes a circulant matrix a).
Next, based on the shifted sample x for obtaining the corresponding feature vector0,x1,x2,x3,x4For each of said shifted samples, connecting by cyclic shift a circulant matrix A into a respective region0,A1,A2,A3,A4. And the modulus acquired by the method is converted into a corresponding cyclic matrix, so that a basis is provided for the following operation.
In an optional embodiment of the present application, the method further includes calculating a model parameter of the filter according to the feature vector, calculating a corresponding map according to the model parameter and the feature vector, obtaining a relative displacement value of the target moving through the corresponding map, and restoring a target region of the current frame according to the relative displacement value, where:
for cyclic matrix A0,A1,A2,A3,A4Performing discrete Fourier transform to obtain the frequency domain
According toCalculate out the corresponding S0,S1,S2,S3,S4Model parameters α ═ α for region dependent filters0,α1,α2,α3,α4};
Wherein i respectively corresponds to S0,S1,S2,S3,S4Target area of λ1The learning rate value is 0.015 for the fixed parameter according to the aboveSubjecting it to complex conjugate transformation to obtain corresponding
wherein Z represents that the current frame is based on the previous frame target area frame S0Obtained shifted samples, λ2The penalty factor value is 25 for the fixed parameter;
according toObtaining the maximum point of the corresponding graph r so as to obtain the relative displacement value of the target movement;
reducing the target area frame S of the current frame according to the relative shift value0Search region S as a basis for the next frame0。
Finally, firstly detecting and acquiring a target area frame S of a first frame of the video0From the target area frame S0Size of (d), constructing the bandwidth and S0A Gaussian shape regression model y with proportional size is obtained by performing DFT on the yThe S is0As a base search area for the next frame, at S0Obtaining 4 extended search areas S with same rectangular size around1,S2,S3,S4. Respectively pair S in the current frame of the video0,S1,S2,S3,S4HOG features (Histogram of oriented gradient) features are extracted. Respectively obtaining the feature vectors x from 5 regions0,x1,x2,x3,x4Representing the respective sample. Respectively circularly shifting the eigenvectors to obtain a corresponding circular matrix A0,A1,A2,A3,A4. Due to the nature of the circulant matrix, vector x can be efficiently aligned0,x1,x2,x3,x4A DFT transform (discrete fourier transform) is performed. Obtaining the frequency domain after DFT transformationAccording toModel parameter α ═ α for calculating correlation filter0,α1,α2,α3,α4}. Taking i as 0, 1, 2, 3 and 4, representing the corresponding marks of the target area and the surrounding area, the fixed parameter learning rate is a constant, and lambda10.015, eigenvector x0,x1,x2,x3,x4And a regression Gaussian model y obtained from the target region box S0 according to the regression Gaussian model ySubjecting it to complex conjugate transformation to obtain correspondingObtained by discrete Fourier transform on a regression Gaussian model y
According toObtaining a corresponding map r, a size and S of the current frame0And (5) the consistency is achieved. Where z denotes that the current frame is based on the previous frame target area box S0 (also called the basic search area S of the current frame)0) And obtaining the feature vector. Fixed parameter penalty factor lambda2=25,x0,x1,x2,x3,x4Representing S in the previous frame0,S1,S2,S3,S4Region-derived feature vector α0,α1,α2,α3,α4Obtained from the previous step and represents the training in the previous frameModel parameters of the correlation filter are refined. Finding the maximum point in the corresponding graph r, obtaining the relative shift value of the target movement, and further recovering the target area frame S0 of the current frame. The target area box S0 of the frame is then used as the base search area S of the next frame0. The steps S101 to S108 are repeated in this loop to perform the loop update so as to realize the entire tracking process.
The specific embodiment is as follows:
as shown in fig. 3 and fig. 5 to 7, the target area frame S for the first frame of the video0And (6) obtaining. The specific means usually adopts a detection module and frames S according to the target areaOSize of (d) to construct the bandwidth and S0A size-proportional Gaussian shape regression model y, and performing DFT (discrete Fourier transform) on y to obtain
Step a: with SOAs the base search area for the next frame, i.e. the second frame, at SORespectively obtain 4 extended search regions S with the same rectangular size1,S2,S3,S4. The tracking target is a target vehicle on one side of the left hand.
Step b: in the second frame of the video, respectively, the S0,S1,S2,S3,S4HOG features (Histogram of Oriented Gradient features) are extracted. Obtaining the feature vectors x in the above 5 regions respectively0,x1,x2,x3,x4Representing the respective sample. Converting the analog quantity of a specific video picture into a digital quantity which can be calculated, wherein the feature vector x0,x1,x2,x3,x4Obtaining corresponding cyclic matrix A after self-cyclic shift0,A1,A2,A3,A4. Due to the nature of the circulant matrix, vector x can be efficiently aligned0,x1,x2,x3,x4A DFT transform (discrete fourier transform) is performed. Obtaining the frequency domain after DFT transformation
And C: according toAnd calculating to obtain model parameters. Then according toObtaining a corresponding graph r of the second frame, the corresponding graph size and S0And (5) the consistency is achieved. Where z denotes a target area box S0 (also called the base search area S of the current frame) of the second frame based on the first frame0) And obtaining the feature vector.
And finding out the relative displacement value of the target movement according to the maximum point of the corresponding graph r. Restoring the second frame to obtain a target area frame S0 of the current second frame, and taking the area as a basic search area S of a third frame0And repeating the steps a, b and c for the third frame to realize the whole tracking process.
In fig. 3, in the detailed explanation of the technical solution of the present invention, relevant parameters and formulas are given, and the present invention is applied to the actual camera monitoring. As shown in fig. 5-7 below, the tracked video object is a white car, followed by frames from the video during the tracking process, and the target frame represents the tracking frame. The method is used for single target tracking, and can also be applied to a multi-target tracking system in combination with a detection module, which is not described herein again.
The application also provides an intelligent device, which comprises:
a detection unit for detecting a video stream acquisition target region frame S0;
A configuration arithmetic unit for passing through the target area frame S0Constructing a regression model and carrying out operation transformation;
an acquisition unit for framing S in the target area in the next frame of the video stream0Obtaining a plurality of expanded search areas around;
an extraction unit for extracting a target region frame S0And a plurality of said extended search area extractionTaking the feature vector and carrying out operation transformation;
the calculation unit is used for calculating the model parameters of the filter according to the characteristic vectors and calculating the corresponding graph according to the model parameters and the characteristic vectors;
the acquisition unit is further used for acquiring a relative displacement value of the target movement through the corresponding graph;
the restoring unit is used for restoring a target area of the current frame through the relative displacement value;
and the repeated circulation unit is used for realizing the tracking of the target object by repeated circulation.
Furthermore, the detection unit is further configured to obtain a target area frame S after detecting the first frame of the video0;
The construction unit is also used for constructing a frame S according to the target area0Size of (d), constructing bandwidth and S0A size-proportional gaussian shape regression model y;
the computing unit is also used for carrying out discrete Fourier transform operation on the y to obtain
The acquisition unit is also used for acquiring the data by S0As the basic search area of the next frame, obtaining 4 extended search areas S with the same rectangular size around the basic search area1,S2,S3,S4;
The extraction unit is used for searching the area S according to the basis0And the extended search area S1,S2,S3,S4Respectively extracting the directional gradient histogram features to obtain the shift sample x of the corresponding feature vector0,x1,x2,x3,x4。
Furthermore, the connection unit is further configured to obtain the shifted samples x of the corresponding feature vector according to the obtained shifted samples x0,x1,x2,x3,x4For each said shifted sample, connecting by cyclic shiftCirculant matrix A for respective region0,A1,A2,A3,A4。
Furthermore, the computing unit is also used for aligning the circulant matrix A0,A1,A2,A3,A4Performing discrete Fourier transform to obtain the frequency domain
And is also used according toCalculate out the corresponding S0Model parameters α ═ α of the filter associated with the regions S1, S2, S3, S40,α1,α2,α3,α4Wherein i respectively represents S0Target region, λ, of S1, S2, S3, S41The learning rate value is 0.015 for the fixed parameter according to the above Subjecting it to complex conjugate transformation to obtain corresponding
And also for passingAccording toA corresponding map r of the current frame is obtained, where Z denotes a shifted sample, λ, of the current frame obtained on the basis of the previous frame target area block S02The penalty factor value is 25 for the fixed parameter;
the acquisition unit is also used for acquiringObtaining the maximum point of the corresponding graph r so as to obtain the relative displacement value of the target movement;
the reduction unit is used for reducing the target area frame S of the current frame according to the relative shift value0Search region S as a basis for the next frame0。
Although the embodiments of the present invention have been described above, the above description is only for the convenience of understanding the present invention, and is not intended to limit the present invention. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (8)
1. A method for video object tracking based on correlation filtering, the method comprising:
s101, detecting a video stream to acquire a target area frame S0;
S102, passing the target area frame S0Constructing a regression model and carrying out operation transformation;
s103, in the next frame of the video stream, a target area frame S0Obtaining a plurality of expanded search areas around;
s104, regarding the target area frame S0Extracting characteristic vectors from a plurality of the extended search areas and performing operation transformation;
s105, calculating model parameters of the filter according to the characteristic vectors;
s106, calculating a corresponding graph according to the model parameters and the characteristic vectors;
s107, acquiring a relative displacement value of the target movement through the corresponding graph;
s108, restoring a target area frame of the current frame through the relative displacement value;
s109, the target area frame in the flow S108 is used as the target area frame S of the next frame0And repeating the above steps S103 to S108 to achieve the pair of targetsTracking of the object.
2. The video object tracking method according to claim 1, wherein the video stream acquisition target area box S is detected0Passing through the target area frame S0Constructing a regression model and carrying out operation transformation, and setting a target area frame S in the next frame of the video stream0Obtaining multiple extended search areas around, and framing the target area0And extracting feature vectors from a plurality of the extended search regions and performing operation transformation, wherein the method further comprises the following steps:
obtaining a target area frame S after detecting a first frame of a video0;
According to the target area frame S0Size of the first, constructing the bandwidth and S0A size-proportional Gaussian shape regression model y, and performing discrete Fourier transform operation on the y to obtain
With S0As the basic search area of the next frame, obtaining 4 extended search areas S with the same rectangular size around the basic search area1,S2,S3,S4;
According to the basic search area S0And the extended search area S1,S2,S3,S4Respectively extracting the directional gradient histogram features to obtain the shift sample x of the corresponding feature vector0,x1,x2,x3,x4。
3. The video object tracking method of claim 2, further comprising:
from the shifted samples x that obtain the corresponding feature vector0,x1,x2,x3,x4For each of said shifted samples, connecting by cyclic shift a circulant matrix A into a respective region0,A1,A2,A3,A4。
4. The method of claim 3, wherein the model parameters of the filter are calculated according to the eigenvectors, the corresponding map is calculated according to the model parameters and the eigenvectors, the relative displacement value of the target movement is obtained from the corresponding map, and the target region of the current frame is restored according to the relative displacement value, the method further comprising:
for cyclic matrix A0,A1,A2,A3,A4Performing discrete Fourier transform to obtain the frequency domain
According toCalculate out the corresponding S0,S1,S2,S3,S4Model parameters α ═ α for region dependent filters0,α1,α2,α3,α4};
Wherein i respectively corresponds to S0,S1,S2,S3,S4Target area of λ1The learning rate value is 0.015 for the fixed parameter according to the aboveSubjecting it to complex conjugate transformation to obtain corresponding
wherein Z represents that the current frame is based on the previous frame target area frame S0Obtained shifted samples, λ2The penalty factor value is 25 for the fixed parameter;
according toObtaining the maximum point of the corresponding graph r so as to obtain the relative displacement value of the target movement;
reducing the target area frame S of the current frame according to the relative shift value0Search region S as a basis for the next frame0。
5. An intelligent device, comprising:
a detection unit for detecting a video stream acquisition target region frame S0;
A configuration arithmetic unit for passing through the target area frame S0Constructing a regression model and carrying out operation transformation;
an acquisition unit for framing S in the target area in the next frame of the video stream0Obtaining a plurality of expanded search areas around;
an extraction unit for extracting a target region frame S0Extracting characteristic vectors from a plurality of the extended search areas and performing operation transformation;
the calculation unit is used for calculating the model parameters of the filter according to the characteristic vectors and calculating the corresponding graph according to the model parameters and the characteristic vectors;
the acquisition unit is further used for acquiring a relative displacement value of the target movement through the corresponding graph;
the restoring unit is used for restoring a target area of the current frame through the relative displacement value;
and the repeated circulation unit is used for realizing the tracking of the target object by repeated circulation.
6. The intelligent device according to claim 5, further comprising:
the detection unit is also used for acquiring a target area frame S after detecting the first frame of the video0;
The construction operation unit is also used for constructing the frame S according to the target area0Size of (d), constructing bandwidth and S0A size-proportional gaussian shape regression model y;
the computing unit is also used for carrying out discrete Fourier transform operation on the y to obtain
The acquisition unit is also used for acquiring the data by S0As the basic search area of the next frame, obtaining 4 extended search areas S with the same rectangular size around the basic search area1,S2,S3,S4;
The extraction unit is used for searching the area S according to the basis0And the extended search area S1,S2,S3,S4Respectively extracting the directional gradient histogram features to obtain the shift sample x of the corresponding feature vector0,x1,x2,x3,x4。
7. The intelligent device according to claim 6, further comprising:
a connection unit for obtaining the shift sample x of the corresponding feature vector0,x1,x2,x3,x4For each of said shifted samples, connecting by cyclic shift a circulant matrix A into a respective region0,A1,A2,A3,A4。
8. The intelligent device according to claim 7, characterized in that it comprises:
the meterA computing unit for computing the cyclic matrix A0,A1,A2,A3,A4Performing discrete Fourier transform to obtain the frequency domain
And is also used according toCalculate out the corresponding S0Model parameters α ═ α of the filter associated with the regions S1, S2, S3, S40,α1,α2,α3,α4Wherein i respectively represents S0Target region, λ, of S1, S2, S3, S41The learning rate value is 0.015 for the fixed parameter according to the aboveSubjecting it to complex conjugate transformation to obtain corresponding
And also for passingAccording toA corresponding map r of the current frame is obtained, where Z denotes a shifted sample, λ, of the current frame obtained on the basis of the previous frame target area block S02The penalty factor value is 25 for the fixed parameter;
the acquisition unit is also used for acquiringObtaining the maximum point of the corresponding graph r so as to obtain the relative displacement value of the target movement;
the reduction unit is used for reducing the phase difference value according to the relative shift valueTarget region frame S of previous frame0Search region S as a basis for the next frame0。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710742685.0A CN107590820B (en) | 2017-08-25 | 2017-08-25 | Video object tracking method based on correlation filtering and intelligent device thereof |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710742685.0A CN107590820B (en) | 2017-08-25 | 2017-08-25 | Video object tracking method based on correlation filtering and intelligent device thereof |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107590820A CN107590820A (en) | 2018-01-16 |
CN107590820B true CN107590820B (en) | 2020-06-02 |
Family
ID=61042885
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710742685.0A Active CN107590820B (en) | 2017-08-25 | 2017-08-25 | Video object tracking method based on correlation filtering and intelligent device thereof |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107590820B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108961311B (en) * | 2018-06-20 | 2021-06-22 | 吉林大学 | Dual-mode rotor craft target tracking method |
CN109064493B (en) * | 2018-08-01 | 2021-03-09 | 苏州飞搜科技有限公司 | Target tracking method and device based on meta-learning |
CN111311645A (en) * | 2020-02-25 | 2020-06-19 | 四川新视创伟超高清科技有限公司 | Ultrahigh-definition video cut target tracking and identifying method |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015163830A1 (en) * | 2014-04-22 | 2015-10-29 | Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi | Target localization and size estimation via multiple model learning in visual tracking |
CN105447887A (en) * | 2015-11-06 | 2016-03-30 | 掌赢信息科技(上海)有限公司 | Historical-route-based target tracking method and electronic equipment |
CN106204639A (en) * | 2016-06-27 | 2016-12-07 | 开易(北京)科技有限公司 | Based on frequency domain regression model target tracking method, system and senior drive assist system |
CN106296723A (en) * | 2015-05-28 | 2017-01-04 | 展讯通信(天津)有限公司 | Target location method for tracing and device |
CN106570893A (en) * | 2016-11-02 | 2017-04-19 | 中国人民解放军国防科学技术大学 | Rapid stable visual tracking method based on correlation filtering |
CN106570486A (en) * | 2016-11-09 | 2017-04-19 | 华南理工大学 | Kernel correlation filtering target tracking method based on feature fusion and Bayesian classification |
CN106815859A (en) * | 2017-01-13 | 2017-06-09 | 大连理工大学 | Target tracking algorism based on dimension self-adaption correlation filtering and Feature Points Matching |
CN106934338A (en) * | 2017-01-09 | 2017-07-07 | 浙江汉凡软件科技有限公司 | A kind of long-term pedestrian tracting method based on correlation filter |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101767564B1 (en) * | 2015-11-12 | 2017-08-11 | 성균관대학교산학협력단 | A method of analysing images of rod-like particles |
-
2017
- 2017-08-25 CN CN201710742685.0A patent/CN107590820B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015163830A1 (en) * | 2014-04-22 | 2015-10-29 | Aselsan Elektronik Sanayi Ve Ticaret Anonim Sirketi | Target localization and size estimation via multiple model learning in visual tracking |
CN106296723A (en) * | 2015-05-28 | 2017-01-04 | 展讯通信(天津)有限公司 | Target location method for tracing and device |
CN105447887A (en) * | 2015-11-06 | 2016-03-30 | 掌赢信息科技(上海)有限公司 | Historical-route-based target tracking method and electronic equipment |
CN106204639A (en) * | 2016-06-27 | 2016-12-07 | 开易(北京)科技有限公司 | Based on frequency domain regression model target tracking method, system and senior drive assist system |
CN106570893A (en) * | 2016-11-02 | 2017-04-19 | 中国人民解放军国防科学技术大学 | Rapid stable visual tracking method based on correlation filtering |
CN106570486A (en) * | 2016-11-09 | 2017-04-19 | 华南理工大学 | Kernel correlation filtering target tracking method based on feature fusion and Bayesian classification |
CN106934338A (en) * | 2017-01-09 | 2017-07-07 | 浙江汉凡软件科技有限公司 | A kind of long-term pedestrian tracting method based on correlation filter |
CN106815859A (en) * | 2017-01-13 | 2017-06-09 | 大连理工大学 | Target tracking algorism based on dimension self-adaption correlation filtering and Feature Points Matching |
Also Published As
Publication number | Publication date |
---|---|
CN107590820A (en) | 2018-01-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113269237B (en) | Assembly change detection method, device and medium based on attention mechanism | |
CN110991272B (en) | Multi-target vehicle track recognition method based on video tracking | |
CN108734723B (en) | Relevant filtering target tracking method based on adaptive weight joint learning | |
CN107358623B (en) | Relevant filtering tracking method based on significance detection and robustness scale estimation | |
CN107369166B (en) | Target tracking method and system based on multi-resolution neural network | |
CN111311647B (en) | Global-local and Kalman filtering-based target tracking method and device | |
CN107452015B (en) | Target tracking system with re-detection mechanism | |
CN107169994B (en) | Correlation filtering tracking method based on multi-feature fusion | |
CN107590820B (en) | Video object tracking method based on correlation filtering and intelligent device thereof | |
CN110175649B (en) | Rapid multi-scale estimation target tracking method for re-detection | |
CN109087333B (en) | Target scale estimation method and device based on correlation filtering tracking algorithm | |
Feng et al. | A novel saliency detection method for wild animal monitoring images with WMSN | |
CN111028263A (en) | Moving object segmentation method and system based on optical flow color clustering | |
CN108898619B (en) | Target tracking method based on PVANET neural network | |
CN109410246B (en) | Visual tracking method and device based on correlation filtering | |
CN112233143B (en) | Target tracking method, device and computer readable storage medium | |
CN113139904B (en) | Image blind super-resolution method and system | |
CN113989556A (en) | Small sample medical image classification method and system | |
CN110689559B (en) | Visual target tracking method based on dense convolutional network characteristics | |
CN112767450A (en) | Multi-loss learning-based related filtering target tracking method and system | |
CN109241932B (en) | Thermal infrared human body action identification method based on motion variance map phase characteristics | |
CN114708307B (en) | Target tracking method, system, storage medium and device based on correlation filter | |
CN113706580B (en) | Target tracking method, system, equipment and medium based on relevant filtering tracker | |
CN113763415B (en) | Target tracking method, device, electronic equipment and storage medium | |
CN113033356A (en) | Scale-adaptive long-term correlation target tracking method |
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 | ||
TA01 | Transfer of patent application right | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20200427 Address after: 730000 a3-013, zone a, 704 Jiatan Road, Chengguan District, Lanzhou City, Gansu Province (station) Applicant after: Lanzhou feisou Information Technology Co., Ltd Address before: 100082 Beijing city Haidian District Xitucheng Road 10, High-Tech Mansion BUPT, 1209 Applicant before: BEIJING FEISOU TECHNOLOGY Co.,Ltd. |
|
GR01 | Patent grant | ||
GR01 | Patent grant |