CN117635665A - Anti-occlusion target tracking method based on correlation filtering - Google Patents
Anti-occlusion target tracking method based on correlation filtering Download PDFInfo
- Publication number
- CN117635665A CN117635665A CN202410103937.5A CN202410103937A CN117635665A CN 117635665 A CN117635665 A CN 117635665A CN 202410103937 A CN202410103937 A CN 202410103937A CN 117635665 A CN117635665 A CN 117635665A
- Authority
- CN
- China
- Prior art keywords
- target
- filter
- scale
- information
- detector
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 90
- 238000001914 filtration Methods 0.000 title claims abstract description 34
- 230000004927 fusion Effects 0.000 claims description 10
- 238000012706 support-vector machine Methods 0.000 claims description 6
- 230000000694 effects Effects 0.000 abstract description 7
- 230000004044 response Effects 0.000 description 23
- 238000004364 calculation method Methods 0.000 description 14
- 239000011159 matrix material Substances 0.000 description 11
- 238000010586 diagram Methods 0.000 description 8
- 238000001514 detection method Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000010276 construction Methods 0.000 description 3
- 238000006073 displacement reaction Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000003213 activating effect Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005315 distribution function Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of computer vision, in particular to an anti-occlusion target tracking method based on correlation filtering. The method comprises the following steps: positioning the position of the target in the target frame set based on the position filter to obtain a first position; if the first position is unreliable, obtaining a second position based on the position re-detector; if the confidence coefficient of the second position is not smaller than the first preset threshold value, determining the scale information of the target according to the second position and the scale filter, and taking the scale information of the second position and the target as target tracking information. After the first position is determined, whether the second position is required to be determined is determined according to the reliability of the first position, and then the target scale information is determined according to the second position and the scale filter, so that the target tracking information is determined, the follow-up tracking is guided by calculating the confidence coefficient of the result, the anti-interference capability and the anti-shielding capability are good under a complex scene, and the effect and the practicability of the target tracking are improved.
Description
Technical Field
The invention relates to the technical field of computer vision, in particular to an anti-occlusion target tracking method based on correlation filtering.
Background
Along with the development of computer vision technology, target tracking is a hot research problem in the field, and has great application prospects in the fields of man-machine interaction, autopilot, aerospace and the like. To date, the technology is mainly divided into two main systems: a generation formula and a discriminant formula.
The core of the object tracking algorithm based on the generation formula is to model the appearance of the object, find the area most similar to the appearance of the object in the video or image sequence frame by frame, and take the area as the tracking result. The tracker relies on the construction of an online model, and the algorithm is generally guaranteed to be real-time (more than or equal to 24 FPS) due to simple operation, but is sensitive to noise, background information is not fully utilized, and tracking drift is easy to occur. The core of the discriminant-based target tracking algorithm is to consider the target tracking problem as a classification problem, and simultaneously utilize target information and background information to realize the separation of the target and the background by extracting more efficient and higher-dimension features and training a classifier on line. The more complex the extracted features, the more robust the algorithm, but the real-time will be greatly reduced.
The foregoing is provided merely for the purpose of facilitating understanding of the technical solutions of the present invention and is not intended to represent an admission that the foregoing is prior art.
Disclosure of Invention
The invention mainly aims to provide an anti-occlusion target tracking method based on correlation filtering, and aims to solve the technical problems of poor effect and low practicability in target tracking in the prior art.
In order to achieve the above object, the present invention provides an anti-occlusion target tracking method based on correlation filtering, the method comprising the steps of:
positioning the position of the target in the target frame set based on the position filter to obtain a first position;
if the first position is unreliable, repositioning the position of the target in the target frame set based on a position re-detector to obtain a second position;
if the confidence coefficient of the second position is not smaller than a first preset threshold value, determining the scale information of the target according to the second position and the scale filter, and taking the scale information of the second position and the target as target tracking information.
Optionally, the positioning, based on the position filter, the position of the target in the target frame set to obtain the first position includes:
based on the position filter, extracting FHOG features and CN features at the corresponding position of the frame which is the frame where the target is located;
fusing the FHOG features and the CN features to obtain fusion features;
and determining the first position according to the fusion characteristic.
Optionally, after locating the position of the target in the target frame set based on the position filter to obtain the first position, the method further includes:
judging whether the first position is reliable or not;
and if so, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information.
Optionally, if the first position is unreliable, repositioning the target in the target frame set based on a position re-detector to obtain a second position, including:
if the first position is unreliable, judging whether the confidence coefficient of the position re-detector is smaller than a second preset threshold value;
if the target tracking information is smaller than the target tracking information, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information;
and if the target frame is not smaller than the first frame, repositioning the target in the target frame set based on the position re-detector to obtain the second position.
Optionally, before determining the scale information of the target according to the second position and the scale filter if the confidence coefficient of the second position is not less than the first preset threshold, and taking the scale information of the second position and the target as target tracking information, the method further includes:
judging whether the confidence coefficient of the second position is smaller than the first preset threshold value or not;
if the target tracking information is smaller than the target tracking information, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information.
Optionally, the method further comprises: and determining the confidence of the second position and the position re-detector according to the appearance filter.
Optionally, the method further comprises: and determining the confidence of the second position and the position re-detector according to the correlation filtering and peak sidelobe ratio.
Optionally, before the positioning the position of the target in the target frame set based on the position filter, the method further includes:
initializing the position filter, the scale filter, the position re-detector and the appearance filter.
Optionally, the position filter is constructed based on a kernel-space ridge regression method, and the position re-detector is constructed based on an online support vector machine.
Optionally, the method further comprises: the scale filter adopts a PCA method to determine the scale information of the target.
The method is based on a position filter, positions of targets in a target frame set are positioned, and a first position is obtained; if the first position is unreliable, repositioning the position of the target in the target frame set based on the position re-detector to obtain a second position; if the confidence coefficient of the second position is not smaller than the first preset threshold value, determining the scale information of the target according to the second position and the scale filter, and taking the scale information of the second position and the target as target tracking information. According to the method, the first position is determined through the position filter, if the first position is unreliable, the position re-detector is adopted to determine the second position, if the confidence coefficient of the second position is not smaller than the first preset threshold value, the target scale information is determined according to the second position and the scale filter, further the target tracking information is obtained, the subsequent tracking is guided through the confidence coefficient of the calculation result, the method has good anti-interference capability and anti-shielding capability in complex scenes, and the effect and the practicability of target tracking are improved.
Drawings
FIG. 1 is a schematic flow chart of a first embodiment of an anti-occlusion target tracking method based on correlation filtering;
FIG. 2 is a schematic flow chart of a second embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention;
FIG. 3 is a schematic view of a sub-process in a third embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention;
FIG. 4 is a schematic diagram of another sub-process in a third embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention;
FIG. 5 is a schematic diagram of another sub-process in a third embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention.
The achievement of the objects, functional features and advantages of the present invention will be further described with reference to the accompanying drawings, in conjunction with the embodiments.
Detailed Description
It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
The embodiment of the invention provides an anti-occlusion target tracking method based on correlation filtering, and referring to fig. 1, fig. 1 is a flow diagram of a first embodiment of the anti-occlusion target tracking method based on correlation filtering.
In this embodiment, the anti-occlusion target tracking method based on correlation filtering includes the following steps:
step S1: positioning the position of the target in the target frame set based on the position filter to obtain a first position;
it should be noted that, the execution body of the method of the present embodiment may be a target tracking device having functions of position detection, scale detection, data processing, and program running. The following embodiments will illustrate the anti-occlusion target tracking method based on correlation filtering according to the present invention using a target tracking device as an execution subject.
Specifically, the target frame set is an input video or image sequence frame, the position filter is a tracker modified based on KCF (kernel correlation filter) and is used for determining the position of the selected target in the input video or image sequence frame; the first location is the location of the object in the set of object frames as determined by the location filter.
Further, in this embodiment, before step S10, the size of the input video or image sequence frame is determined, and if the size is greater than 360×360, the input frame is scaled to half of the original size, so that the algorithm computation is reduced, and the accuracy is not lost.
Step S2: if the first position is unreliable, repositioning the position of the target in the target frame set based on a position re-detector to obtain a second position;
specifically, judging the confidence coefficient of the first position, and if the confidence coefficient of the first position is higher than or equal to a preset first position threshold value, judging that the first position is reliable; if the confidence coefficient of the first position is lower than a preset first position threshold value, judging that the first position is unreliable, and repositioning the position of the target in the target frame set by adopting a position re-detector to obtain a second position. It should be noted that the first position threshold may be set according to actual situations.
Further, the feature used by the position re-detector is FHOG, which can more effectively characterize the target information. FHOG is a method that improves upon the traditional hog (gradient statistics histogram) feature by incorporating Gaussian features into the relevant filtering framework. The second position refers to the position of the object in the set of object frames as determined by the position re-detector.
Step S3: if the confidence coefficient of the second position is not smaller than a first preset threshold value, determining scale information of the target according to the second position and the scale filter, and taking the scale information of the second position and the target as target tracking information;
specifically, if the confidence coefficient of the second position is greater than or equal to the first preset threshold value, the confidence coefficient of the second position is determined to be reliable, and then the scale information of the target can be determined according to the second position and the scale filter, and the obtained second position and the scale information are used as target tracking information.
Further, the scale filter is a tracker based on KCF (kernel correlation filter) improvement, and is used for estimating the scale of the target. The target tracking information comprises the position information and the scale information of the target, and the target tracking information is obtained, so that the target is tracked.
The method is based on a position filter, positions of targets in a target frame set are positioned, and a first position is obtained; if the first position is unreliable, repositioning the position of the target in the target frame set based on the position re-detector to obtain a second position; if the confidence coefficient of the second position is not smaller than the first preset threshold value, determining the scale information of the target according to the second position and the scale filter, and taking the scale information of the second position and the target as target tracking information. According to the method, the first position is determined through the position filter, if the first position is unreliable, the position re-detector is adopted to determine the second position, if the confidence coefficient of the second position is not smaller than the first preset threshold value, the target scale information is determined according to the second position and the scale filter, further the target tracking information is obtained, the subsequent tracking is guided through the confidence coefficient of the calculation result, the method has good anti-interference capability and anti-shielding capability in complex scenes, and the effect and the practicability of target tracking are improved.
Referring to fig. 2, fig. 2 is a flowchart of a second embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention;
based on the above-described first embodiment, in the present embodiment, step S1 includes:
s11: based on the position filter, extracting FHOG features and CN features at the corresponding position of the frame which is the frame where the target is located;
s12: fusing the FHOG features and the CN features to obtain fusion features;
s13: determining the first position according to the fusion characteristic;
specifically, in this embodiment, the position filter fuses FHOG features and CN features, and the calculation efficiency is improved by adopting a parallel calculation mode, that is, two independent correlation filters are adopted for the two features to calculate response and update parameters respectively, and the fusion formula is as follows:，/>、/>calculated responses for FHOG feature, CN feature, respectively,>the fusion coefficient is dynamically adjusted according to the response confidence coefficient, and the calculation formula is as follows: />In the followingRepresenting a minimum value, preventing the denominator from being zero.
Further, after the fusion feature is obtained, the first position is determined according to the fusion feature condition in the target frame set. In the embodiment, the first position is determined by performing feature matching on the combined feature formed after the FHOG feature and the CN feature are fused, so that the tracking precision is improved.
Referring to fig. 3, fig. 3 is a schematic flow chart of a third embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention;
based on the above-described second embodiment, in the present embodiment, after step S1, further includes:
s1a: judging whether the first position is reliable or not;
s1b: if so, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information;
specifically, after the first position is obtained, whether the first position is reliable or not is judged by calculating the confidence coefficient of the first position. When the first position is reliable, the first position is used as target position information, and target tracking information is determined according to the first position and the scale information.
It should be appreciated that if the first location is determined to be reliable at this time, this indicates that the target is not occluded or otherwise complicated in the target frame set. At this time, the target tracking information can be determined through the position filter and the scale filter.
In this embodiment, by determining whether the first position is reliable or not, different manners may be adopted to determine the target tracking information, thereby improving the tracking effect and efficiency.
Referring to fig. 4, fig. 4 is a schematic diagram of another sub-flowchart of a third embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention;
based on the above-described second embodiment, in the present embodiment, step S2 includes:
s21: if the first position is unreliable, judging whether the confidence coefficient of the position re-detector is smaller than a second preset threshold value;
s22: if the target tracking information is smaller than the target tracking information, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information;
s23: if not, repositioning the position of the target in the target frame set based on the position re-detector to obtain the second position;
specifically, if the first position is unreliable, that is, the confidence level of the first position is lower than the preset position threshold. Further judging the confidence coefficient of the position re-detector, and if the confidence coefficient of the position re-detector is smaller than a second preset threshold value (namely, the result of the position re-detector is unreliable), determining the scale information of the target according to the first position and the scale filter.
Further, if the confidence level of the position re-detector is greater than the second preset threshold (i.e. the result indicating that the position re-detector is reliable), the target can be repositioned according to the position re-detector to obtain the second position.
In the present embodiment, when the first position is unreliable, the judgment target may be lost. When the first position is unreliable, if the confidence coefficient of the position re-detector is smaller than a second preset threshold value, the target is judged not to be lost, and if the confidence coefficient of the position re-detector is not smaller than the second preset threshold value, the target is judged to be lost (blocked), so that the position re-detector can be repositioned according to the position, and the anti-interference capability of tracking is improved.
Referring to fig. 5, fig. 5 is a schematic diagram of another sub-flowchart of a third embodiment of an anti-occlusion object tracking method based on correlation filtering according to the present invention;
based on the above-described second embodiment, in the present embodiment, before step S3, further includes:
s3a: judging whether the confidence coefficient of the second position is smaller than the first preset threshold value or not;
s3b: if the target tracking information is smaller than the target tracking information, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information;
specifically, after the second position is obtained, whether the second position is reliable or not is firstly judged, namely, when the confidence coefficient of the second position is larger than or equal to a first preset threshold value, the obtained second position is reliable, and when the confidence coefficient of the second position is smaller than the first preset threshold value, the second position is unreliable. When the second position is unreliable, the scale information is determined according to the first position obtained by the position filter and the scale filter, and then the first position and the scale information are used as target tracking information.
In this embodiment, when the target is not lost (not blocked), the target tracking information is determined through the first position and the scale filter, and when the target is lost (blocked), the target tracking information is determined through the second position and the scale filter, so that the accuracy of target tracking is improved.
Based on the above second embodiment, in this embodiment, the confidence level of the second position and the position re-detector is determined according to the appearance filter, and the confidence level is measured by using the appearance filter, so that the reliability of the prediction accuracy and the feature matching can be conveniently determined by the confidence level.
In this embodiment, the confidence level of the second position and the position re-detector is determined according to the correlation filtering and peak sidelobe ratio, so that the prediction accuracy of the confidence level is improved.
In this embodiment, before step S1, the method further includes:
initializing the position filter, the scale filter, the position re-detector and the appearance filter, wherein the initialization can improve the calculation accuracy of the position filter, the scale filter, the position re-detector and the appearance filter.
Optionally, the position filter is built based on a kernel space ridge regression method, the position re-detector is built based on an online support vector machine, the position re-detector is built based on the online support vector machine, the speed and the precision are considered, the position filter is built based on the kernel space ridge regression method, and the tracking precision is improved.
Optionally, the scale filter uses PCA to determine the scale information of the target. The PCA method is adopted to reduce the data dimension, so that the calculated amount is reduced, the calculated time is shortened, and the tracking speed is improved.
In this embodiment, in the first frame of the input video or image sequence, the tracking target of interest is selected, and the position filter, appearance filter, scale filter, and position re-detector of the target tracking device are initialized. Inputting the video or image sequence into the target tracking device frame by frame to detect the target: the position filter uses a correlation kernel filter method to estimate the position of the target; calculating the confidence coefficient of the current result by using an appearance filter, and judging whether the current detection result is unreliable by using a double-threshold method; if the determination is unreliable and the confidence of the position re-detector is high, activating the position re-detector to relocate the target, and if the confidence of the relocated result exceeds a threshold, accepting the re-detection result.
In the scale filter, the PCA method is used for carrying out data dimension reduction, so that the operand is reduced; according to the confidence level of the result, the learning rates of the position filter, the appearance filter and the scale filter are adjusted in real time, and parameters are updated more effectively aiming at the selected target; and judging whether the current result is reliable or not according to the confidence level of the result, and determining whether to use the current sample to update parameters, so that the interference caused by unreliable samples is reduced, and the accuracy, the robustness and the instantaneity of target tracking are further improved.
Further, in this embodiment, the reliability of the tracking result is determined by using a three-threshold method, after the latest reliable position of the target is obtained, the scale of the target is estimated by using a scale filter, after the reliable estimation of the position and the scale of the target is obtained, the current sample is determined to be reliable by using the three-threshold method, and then the position filter, the appearance filter, the scale filter and the position re-detector are updated, and the learning rate is dynamically adjusted.
In this embodiment, the property of fast fourier transform is used to transform the convolution operation in the time domain to the dot product operation in the frequency domain, reducing the algorithm complexity. The appearance filter is similar to the position filter in construction, but only takes the target size of the target positionExtracting FHOG characteristics and performing related operations; and to reduce the computational effort, the solution of the scale filter is decomposed into molecules +.>Denominator->Respectively updating; the position estimation and the scale estimation of the target are separated, and the latter takes the result of the former as prior information; in the scale filter, performing data dimension reduction on the features by using a PCA method; the position re-detector is built based on an on-line support vector machine.
Further, after the FHOG feature and the CN feature are adaptively fused into a feature with higher dimension, a detection sample is created by using a circular convolution matrix, and convolved to obtain a response diagram. The response diagram reflects the correlation degree of different areas, the larger the numerical value is, the higher the correlation degree is, the most conforming position is found in the response diagram by using the Newton iteration method, and the new position is taken as a target.
In this embodiment, the confidence is obtained by combining APCE (Average Peak-to Correlation Energy) with PSR (Peak to Sidelobe Ratio), and the calculation formula is:wherein->Is a proportionality coefficient, preferably 0.8. Wherein, the calculation formula of APCE is: />，、/>、/>Respectively represent a response maximum value, a response minimum value and a response +.>A response value for the location; the calculation formula of PSR is:，/>response value representing peak value +.>Mean value of sidelobes>Representing the standard deviation of the sidelobes.
In this embodiment, in the present invention, the reliability of the tracking result is determined by adopting a three-threshold method: ratio of position filter response maximum to historical response meanRatio of position filter response confidence to historical confidence +.>Ratio of appearance filter response maximum to history response maximum +.>The reliability judgment of the tracking result is more accurate. When all three are smaller than the threshold value, the result at this time is determined to be unreliable. It should be noted that, in the present embodiment, the threshold values may be 0.45, 0.4, and 0.55, respectively;
when the target is blocked, the characteristics of the target change, and the tracking result is interfered. In order to solve the problem, the reliability of the tracking result is judged by adopting the three-threshold method. When all the three indexes are lower than the threshold value, the target is shielded at the moment, a position re-detector is activated, when the result of the position re-detector is reliable, the target is repositioned in the image, and when the confidence of the re-detection result exceeds the threshold value, the re-detection result is accepted as the final position of the target;
after the latest reliable position of the target is obtained, a scale filter is used for carrying out scale on the targetEstimating: dividing an image block where a target is located into S scales:,/>represents a scale change factor, the target size is +.>The method comprises the steps of carrying out a first treatment on the surface of the Then extracting FHOG features from S scale images, scaling to the original size, reducing the dimension of the feature data by a PCA method, performing correlation filtering calculation on the feature data by a scale filter to obtain S scale response graphs, and selecting the scale corresponding to the maximum value as a target scale estimation result, namely #>The method comprises the steps of carrying out a first treatment on the surface of the After obtaining reliable estimation of the position and the scale of the target, after judging that the current sample is reliable by using a three-threshold method, updating a position filter, an appearance filter, a scale filter and a target re-detector, and in order to improve the learning efficiency, reduce the interference, and adjust the learning rate in real time according to the confidence coefficient:；
the size of the correlation filter is fixed, and the convolution operation in the time domain is converted into dot product operation in the frequency domain by using the property of fast Fourier change, so that the algorithm complexity is reduced;
the position filter is constructed based on a kernel space ridge regression method: manually selecting the size to beAfter the goal of (2), a learning sample +.>The method comprises the steps of carrying out a first treatment on the surface of the The kernel space-based ridge regression method enables the sum of square errors to be minimized to obtain a calculation formula of coefficients: />In the formula->Representing a norm->For regularization coefficient, ++>Representing a mapping function mapped to kernel space, +.>Representing a label corresponding to the sample; constructing a corresponding sample tag ++using a Gaussian distribution function according to the magnitude of the displacement>The smaller the displacement amount, the closer the tag value is to 1, whereas the closer the tag value is to 0; to mitigate the boundary effects of loops, the constructed samples are windowed, the discontinuities at the edges of the samples are suppressed, and the object can be highlighted: />Wherein->；
The cyclic matrix is composed of displacement matrixTo realize: for an n-dimensional column vector sample:，/>can be combined by +.>The method comprises the following steps: />;
Using the dual spatial theory, the coefficients of the position filter satisfy:,/>representing inverse fourier transform ++>Representing a fast discrete fourier transform (Discrete Fast Fourier Transform, DFFT); the Gaussian kernel is adopted: />，/>Number of channels representing a feature>Representing complex conjugate>Representing a dot product operation. Motion filter->The size of the target position is +.>Is->Performing correlation calculation, and obtaining output responses of all candidate samples by utilizing the property of Fourier transformation, wherein the output responses are as follows: />，/>Representing the learned target appearance characteristics, and finding an optimal position in the output response by using Newton iteration as an estimated target position;
in-place filterThere are two parameters that need to be updated: target appearance characteristics->And classifier parameters->Iteration is performed using linear interpolation: />，/>The method comprises the steps of carrying out a first treatment on the surface of the Learning rate->Dynamically adjusting by using the method;
the appearance filter is similar to the position filter in construction, but only takes the target size of the target positionExtracting FHOG characteristics and performing related operations;
scale filter to reduce the algorithm complexity, a one-dimensional filter is constructed in the frequency domain:，/>is the scale filter parameter in the frequency domain, +.>Is obtained by performing DFFT on each row of the feature matrix extracted under S scales>Matrix of->Is a desired response generated by a Gaussian function of magnitude +.>The method comprises the steps of carrying out a first treatment on the surface of the Solving to obtain: />；
To reduce the computational effort, the scale filter is decomposed into moleculesDenominator->Update respectively:，/>，/>the dimension of (2) is +.>，/>The dimension of (2) is +.>The method comprises the steps of carrying out a first treatment on the surface of the Similarly, the learning rate is dynamically adjusted;
scale filter scale feature map to be detectedSolving to obtain a response: />In->Selecting a scale corresponding to the maximum response as an estimation result of the target scale;
the position estimation of the object is separated from the scale estimation, which takes the result of the former as a priori information.
In the scale filter, the PCA method is used to perform data dimension reduction on the features.Representation->The appearance characteristics of the dimensions are that,representation->Characteristic covariance matrix of frame,>is the learning rate of the covariance matrix, +.>Representing a low dimensional space, +.>Representation->The mapping matrix of the dimension is: when->There is->When->There is->. Wherein->Is->Front->Matrix of columns>；/>、/>Intermediate variable->Eigenvalue decompositionCome from (I)>Is ordered from small to large +.>Diagonal matrix of eigenvalues of +.>Is a matrix formed by unit eigenvectors of corresponding eigenvalues; intermediate variable->The calculation formula of (2) is as follows: when->There is->When->Sometimes haveWherein->，/>By->Is centered to:；
the position re-detector is constructed based on an online support vector machine, and selects image blocks with target sizes in a region near the target position, and creates positive samples and negative samples by using Intersection over Union (IoU): ioU0.75 is positive sample, ioU->0.28 is a negative sample, discarding the middle blurred sample.
Experiments prove that samples do not need to be stored from the head to train the position re-detector, and the maximum accommodating image frame of the samples is set to 10 and the maximum iteration number is set to 20 under the balance speed and precision.
The method is based on a position filter, positions of targets in a target frame set are positioned, and a first position is obtained; if the first position is unreliable, repositioning the position of the target in the target frame set based on the position re-detector to obtain a second position; if the confidence coefficient of the second position is not smaller than the first preset threshold value, determining the scale information of the target according to the second position and the scale filter, and taking the scale information of the second position and the target as target tracking information. According to the method, the first position is determined through the position filter, if the first position is unreliable, the position re-detector is adopted to determine the second position, if the confidence coefficient of the second position is not smaller than the first preset threshold value, the target scale information is determined according to the second position and the scale filter, further the target tracking information is obtained, the subsequent tracking is guided through the confidence coefficient of the calculation result, the method has good anti-interference capability and anti-shielding capability in complex scenes, and the effect and the practicability of target tracking are improved.
It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or system. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or system that comprises the element.
The foregoing embodiment numbers of the present invention are merely for the purpose of description, and do not represent the advantages or disadvantages of the embodiments.
From the above description of the embodiments, it will be clear to those skilled in the art that the above-described embodiment method may be implemented by means of software plus a necessary general hardware platform, but of course may also be implemented by means of hardware, but in many cases the former is a preferred embodiment. Based on such understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art in the form of a software product stored in a storage medium (e.g. read-only memory/random-access memory, magnetic disk, optical disk), comprising instructions for causing a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to perform the method according to the embodiments of the present invention.
The foregoing description is only of the preferred embodiments of the present invention, and is not intended to limit the scope of the invention, but rather is intended to cover any equivalents of the structures or equivalent processes disclosed herein or in the alternative, which may be employed directly or indirectly in other related arts.
Claims (10)
1. An anti-occlusion target tracking method based on correlation filtering, which is characterized by comprising the following steps:
positioning the position of the target in the target frame set based on the position filter to obtain a first position;
if the first position is unreliable, repositioning the position of the target in the target frame set based on a position re-detector to obtain a second position;
if the confidence coefficient of the second position is not smaller than a first preset threshold value, determining the scale information of the target according to the second position and the scale filter, and taking the scale information of the second position and the target as target tracking information.
2. The method for tracking an anti-occlusion target based on correlation filtering of claim 1, wherein the positioning the target in the target frame set based on the position filter to obtain the first position comprises:
based on the position filter, extracting FHOG features and CN features at the corresponding position of the frame which is the frame where the target is located;
fusing the FHOG features and the CN features to obtain fusion features;
and determining the first position according to the fusion characteristic.
3. The correlation-filtering-based anti-occlusion object tracking method of claim 1, wherein after locating the position of the object in the object frame set based on the position filter to obtain the first position, further comprising:
judging whether the first position is reliable or not;
and if so, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information.
4. The correlation filtering-based anti-occlusion object tracking method of claim 1, wherein repositioning the object in the set of object frames based on a position re-detector if the first position is unreliable, to obtain a second position, comprises:
if the first position is unreliable, judging whether the confidence coefficient of the position re-detector is smaller than a second preset threshold value;
if the target tracking information is smaller than the target tracking information, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information;
and if the target frame is not smaller than the first frame, repositioning the target in the target frame set based on the position re-detector to obtain the second position.
5. The method for anti-occlusion object tracking based on correlation filtering according to claim 1, wherein before determining scale information of the object according to the second position and the scale filter if the confidence level of the second position is not less than a first preset threshold, and taking the scale information of the second position and the object as object tracking information, further comprises:
judging whether the confidence coefficient of the second position is smaller than the first preset threshold value or not;
if the target tracking information is smaller than the target tracking information, determining the scale information of the target according to the first position and the scale filter, and taking the first position and the scale information as target tracking information.
6. The correlation filtering-based anti-occlusion target tracking method of claim 4, further comprising: and determining the confidence of the second position and the position re-detector according to the appearance filter.
7. The correlation-filtering-based anti-occlusion target tracking method of claim 6, further comprising: and determining the confidence of the second position and the position re-detector according to the correlation filtering and peak sidelobe ratio.
8. The correlation-filtering-based anti-occlusion object tracking method of claim 7, wherein before the positioning the object in the set of object frames based on the position filter, further comprising:
initializing the position filter, the scale filter, the position re-detector and the appearance filter.
9. The correlation filtering-based anti-occlusion object tracking method of claim 8, wherein said position filter is constructed based on a kernel-space ridge regression method, and said position re-detector is constructed based on an on-line support vector machine.
10. The correlation-filtering-based anti-occlusion target tracking method of claim 9, further comprising: the scale filter adopts a PCA method to determine the scale information of the target.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410103937.5A CN117635665A (en) | 2024-01-25 | 2024-01-25 | Anti-occlusion target tracking method based on correlation filtering |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410103937.5A CN117635665A (en) | 2024-01-25 | 2024-01-25 | Anti-occlusion target tracking method based on correlation filtering |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117635665A true CN117635665A (en) | 2024-03-01 |
Family
ID=90034240
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410103937.5A Pending CN117635665A (en) | 2024-01-25 | 2024-01-25 | Anti-occlusion target tracking method based on correlation filtering |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117635665A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034193A (en) * | 2018-06-20 | 2018-12-18 | 上海理工大学 | Multiple features fusion and dimension self-adaption nuclear phase close filter tracking method |
| CN111476819A (en) * | 2020-03-19 | 2020-07-31 | 重庆邮电大学 | Long-term target tracking method based on multi-correlation filtering model |
| CN113537241A (en) * | 2021-07-16 | 2021-10-22 | 重庆邮电大学 | Long-term correlation filtering target tracking method based on adaptive feature fusion |
| CN113989331A (en) * | 2021-11-12 | 2022-01-28 | 山西大学 | Long-term target tracking method based on context multi-clue information and adaptive response |
-
2024
- 2024-01-25 CN CN202410103937.5A patent/CN117635665A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109034193A (en) * | 2018-06-20 | 2018-12-18 | 上海理工大学 | Multiple features fusion and dimension self-adaption nuclear phase close filter tracking method |
| CN111476819A (en) * | 2020-03-19 | 2020-07-31 | 重庆邮电大学 | Long-term target tracking method based on multi-correlation filtering model |
| CN113537241A (en) * | 2021-07-16 | 2021-10-22 | 重庆邮电大学 | Long-term correlation filtering target tracking method based on adaptive feature fusion |
| CN113989331A (en) * | 2021-11-12 | 2022-01-28 | 山西大学 | Long-term target tracking method based on context multi-clue information and adaptive response |
Non-Patent Citations (2)
| Title |
|---|
| 王海军: "《复杂场景下单目标视觉跟踪方法》", 30 November 2021, 北京：北京航空航天大学出版社, pages: 6 - 13 * |
| 魏龙生等: "《计算机视觉中的相关滤波跟踪和图像质量评价》", 30 April 2021, 武汉：华中科技大学出版社, pages: 62 - 64 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111260689B (en) | Confidence enhancement-based correlation filtering visual tracking method | |
| Yuan et al. | Particle filter re-detection for visual tracking via correlation filters | |
| EP1818870B1 (en) | Computerized method for tracking object in sequence of frames | |
| EP2064652B1 (en) | Method of image processing | |
| CN110120065B (en) | Target tracking method and system based on hierarchical convolution characteristics and scale self-adaptive kernel correlation filtering | |
| Olson | Maximum-likelihood template matching | |
| CN111311647B (en) | Global-local and Kalman filtering-based target tracking method and device | |
| Zhang et al. | Robust image corner detection based on scale evolution difference of planar curves | |
| CN110111370B (en) | Visual object tracking method based on TLD and depth multi-scale space-time features | |
| CN111754548B (en) | Multi-scale correlation filtering target tracking method and device based on response discrimination | |
| CN109859242B (en) | Target tracking method for prediction adaptive learning | |
| CN109410246B (en) | Visual tracking method and device based on correlation filtering | |
| CN115239760A (en) | Target tracking method, system, equipment and storage medium | |
| Yan et al. | Dual-template adaptive correlation filter for real-time object tracking | |
| CN117635665A (en) | Anti-occlusion target tracking method based on correlation filtering | |
| CN113470074B (en) | Self-adaptive space-time regularization target tracking method based on block discrimination | |
| CN113033356B (en) | Scale-adaptive long-term correlation target tracking method | |
| Wang et al. | Algorithm for target detection in smart city combined with depth learning and feature extraction | |
| CN115345902A (en) | Infrared image dim target detection tracking method and system based on machine learning | |
| CN112613574B (en) | Training method of image classification model, image classification method and device | |
| US11821986B1 (en) | Target tracking method, system, device and storage medium | |
| Jing et al. | An improved target tracking method based on extraction of corner points | |
| CN111291745B (en) | Target position estimation method and device, storage medium and terminal | |
| CN109740470B (en) | Target tracking method, computer device, and computer-readable storage medium | |
| CN113608618B (en) | Hand region tracking method and system |
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 |