CN113033356A - Scale-adaptive long-term correlation target tracking method - Google Patents

Scale-adaptive long-term correlation target tracking method Download PDF

Info

Publication number
CN113033356A
CN113033356A CN202110265773.2A CN202110265773A CN113033356A CN 113033356 A CN113033356 A CN 113033356A CN 202110265773 A CN202110265773 A CN 202110265773A CN 113033356 A CN113033356 A CN 113033356A
Authority
CN
China
Prior art keywords
scale
frame
target
correlation
adaptive
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110265773.2A
Other languages
Chinese (zh)
Other versions
CN113033356B (en
Inventor
索继东
王思鹏
张伟红
柳晓鸣
陈晓楠
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian Maritime University
Original Assignee
Dalian Maritime University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Dalian Maritime University filed Critical Dalian Maritime University
Priority to CN202110265773.2A priority Critical patent/CN113033356B/en
Publication of CN113033356A publication Critical patent/CN113033356A/en
Application granted granted Critical
Publication of CN113033356B publication Critical patent/CN113033356B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a scale-adaptive long-term correlation target tracking method. Firstly, a first frame image is preprocessed to obtain a time context regression model RcTarget appearance regression model RtAnd a detector Drf. For the tracking of the subsequent frame, a search area is created according to the target position of the previous frame, and HOG characteristics are extracted to train a related filter template; carrying out translation estimation to estimate the target position of the current frame; then, a scale pool is constructed, the optimal scale of the predicted target is estimated in a self-adaptive mode, and the target state of the current frame is obtained; if the maximum response ysLess than threshold τrUsing DrfExecuting redetection, and updating the position of the target; next, R is updatedc(ii) a If the maximum response ysGreater than a threshold τaMore, moreNovel Rt(ii) a Then, D is updatedrf(ii) a Finally, the predicted target state R of the current frame is obtainedc、RtAnd Drf. The above steps are repeated until the video image sequence is finished. Compared with algorithms such as long-term correlation filtering (LCT) and the like, the method improves the performance of target tracking and has better robustness in various complex environments.

Description

Scale-adaptive long-term correlation target tracking method
Technical Field
The invention belongs to the field of visual target tracking, and particularly relates to a scale-adaptive long-term correlation target tracking method.
Background
The target tracking belongs to the content of video analysis, namely, the video image sequence is processed. The task of target tracking is to determine the position and size of a target in each subsequent frame by analyzing the group of video image sequences after information such as the position and size of the target in a first frame is given, and to accurately frame the target. The target tracking technology integrates knowledge of mathematics, physics, image processing and the like, and has wide application and development prospects in aspects of military, national defense, intelligent transportation and the like. For example, in the military field, for missile defense, guidance systems, air traffic control, etc.; the method is used for real-time monitoring of traffic flow, traffic accident detection, pedestrian counting and the like in the field of intelligent traffic.
The correlation filtering based tracking algorithm considers the tracking process as a process of template matching and ridge regression. The Kernel Correlation Filtering (KCF) is a correlation filtering target tracking algorithm added with a kernel function, the algorithm utilizes a multichannel directional gradient Histogram (HOG) to extract features, and then positive and negative samples are constructed by performing cyclic shift, but the KCF cannot deal with the problem of target scale change.
In order to solve the problem, a Discriminant Scale Space (DSST) target tracking algorithm adopts a three-dimensional scale space correlation filter joint tracking mode, firstly, a two-dimensional discriminant position filter is used for determining the position information of a target in a video sequence, and then, a one-dimensional scale filter is used for detecting a tracking target output by the position filter, so that the optimal scale of the current target is output. A long-term correlation filtering (LCT) algorithm is added with a correlation filter responsible for detecting confidence coefficient on the basis of a DSST algorithm position filtering and scale filtering framework, the added detection mechanism enables the algorithm to have good tracking performance in videos with the properties of shielding, exceeding visual fields and the like, but the tracking performance of the LCT algorithm needs to be improved in the environment with the properties of overlarge target scale change, low resolution, quick motion, illumination change and the like.
Disclosure of Invention
In view of this, the present invention provides a scale-adaptive long-term correlation (LCSA) target tracking algorithm to improve the accuracy of the existing target tracking algorithm, so that the target tracking algorithm can overcome the interference of various environmental factors.
Therefore, the invention provides the following technical scheme:
a scale-adaptive long-term correlation target tracking algorithm comprises the following steps:
(1) initializing a target detection frame; extracting the characteristics of the target according to the detection frame in the first frame, and initializing a time context regression model RcTarget appearance regression model RtAnd a detector Drf； wherein RcIs responsible for translation estimation, RtResponsible for the scale estimation, DrfIs responsible for the heavy detection;
(2) for the t-th frame, the target position (x) according to the t-1 th framet-1，yt-1) Cutting a search window in the t frame, extracting HOG characteristics, and training a related filter template;
(3) performing translation estimation using RcAnd calculating a response based on the correlation filter fraction obtained from the correlation filter template
Figure BDA0002971725640000021
And estimating the current frame position
Figure BDA0002971725640000022
(4) Constructing a scale pool using a multi-scale search strategy, mapping y by correlations and RtSelf-adaptive estimation of optimal rulerDegree of rotation
Figure BDA0002971725640000023
Obtaining the initial predicted target state of the t-th frame
Figure BDA0002971725640000024
(5) If it is not
Figure BDA0002971725640000025
Use of DrfPerforming double detection, finding candidate state set X, and for each state X 'in X'iCalculate confidence score y'iIf max (y'i)＞τtThen, then
Figure BDA0002971725640000026
wherein ,τrIs a first threshold value, τtIs the second threshold value, and is,
Figure BDA0002971725640000027
a correlation map representing the prediction; obtaining the final predicted target state of the t frame
Figure BDA0002971725640000028
(6) Updating Rc；
(7) If it is not
Figure BDA0002971725640000029
Updating Rt； wherein ,τaIs a third threshold;
(8) update Drf；
(9) And repeating (2) to (8) for the t +1 th frame until the video sequence is ended.
Further, the correlation filter template trained in the step (2) is specifically:
Figure BDA0002971725640000031
where, w denotes the correlation filter,xm,nrepresenting image blocks x having m x n pixels, y (m, n) being denoted xm,nAs a gaussian sample label generated for the training sample,
Figure BDA0002971725640000032
represents the mapping to kernel space and λ represents the regularization parameter.
Further, the correlation filter template trained in the step (2) is specifically:
Figure BDA0002971725640000033
wherein the coefficient a is defined by the formula:
Figure BDA0002971725640000034
f denotes the discrete Fourier operator, xm,nThe image block x is represented by m × n pixels, and x and y represent pixel coordinates.
Further, the tracking task is to calculate a correlation map by means of the image blocks z of a new one of the image frames of size m x n, said step (3) being responsive
Figure BDA0002971725640000035
Determined according to the following formula:
Figure BDA0002971725640000036
wherein f denotes a learned target appearance model,. indicates a Hadamard product, by finding
Figure BDA0002971725640000039
Finds the predicted position of the target.
Further, the scale search policy of step (4) is:
the dimension of the template is fixed as ST＝(sx,sy) The scale pool is set to S ═ t1,t2,…tkFor the current frame at tist|tiSample k sizes in e S to find a sumThe appropriate target scale is then interpolated bilinearly so that the samples at each scale become STThe size of the samples is consistent;
the final target scale response value is calculated as follows:
Figure BDA0002971725640000037
Figure BDA0002971725640000038
is the ith scale sample in the scale pool, with size tiSt。
Further, R is updatedcUpdate RtThe method comprises the following steps:
for Rc and RtThe coefficients f and a in the model are updated frame by frame at the learning rate α as:
Figure BDA0002971725640000041
Figure BDA0002971725640000042
further, DrfA support vector machine detector;
update DrfThe method comprises the following steps:
in each frame, a training set is given { (v)i,ci) 1, 2., N } and N samples, where viIs the feature vector generated by the ith sample, ciE { +1, -1} is a sample label, and the objective function for solving the support vector machine detector hyperplane h is:
Figure BDA0002971725640000043
wherein ,
Figure BDA0002971725640000044
<h,v>denotes h andthe inner product of v; λ represents a regularization parameter;
updating hyperplane parameters using a passive algorithm:
Figure BDA0002971725640000045
wherein ,
Figure BDA0002971725640000046
is the gradient of the loss function with respect to h, and τ e (0, ∞) is a hyper-parameter that controls the update rate of h.
Further, τrIs 0.15, τtIs 0.5, τaIs 0.38.
The invention has the following beneficial effects:
according to the scale-adaptive long-term correlation target tracking method provided by the invention, a scale adaptive strategy and an LCT target tracking frame are effectively fused, and firstly, a scale pool is introduced, so that an algorithm can adaptively select the optimal scale for finding the position of a tracking target. The multi-scale search can be combined with the position estimation filter more stably, and the situation that the scale estimation is over-biased is not easy to occur.
The scale self-adaptive long-term correlation target tracking method provided by the invention has the advantages that the tracking precision is obviously improved compared with that of an LCT algorithm in classical target tracking scenes such as scale change, aspect ratio, low resolution, rapid movement, complete shielding, partial shielding, beyond visual field, illumination change, visual field point conversion, camera movement, similar objects and the like on an Unmanned Aerial vehicle 123(UAV 123).
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a flowchart of a scale-adaptive long-term relevance target tracking method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a scale-adaptive long-term correlation target tracking method according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an adaptive scale model of a scale-adaptive long-term correlation target tracking method according to an embodiment of the present invention;
FIG. 4 is a graph of tracking accuracy of the present invention and other algorithms in a UAV123 data set over 123 video sequences;
FIG. 5 is a graph of the success rate of the present invention and other algorithms in a UAV123 data set at 123 video sequences;
FIG. 6 is a graph of tracking accuracy and success rate in the context of Scale Variation in the UAV123 data set by the present invention and other algorithms;
FIG. 7 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of Aspect Ratio Change in the UAV123 data set;
FIG. 8 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of Low Resolution in the UAV123 data set;
FIG. 9 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of Fast Motion in the UAV123 data set;
FIG. 10 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of Full Occlusion in the UAV123 dataset;
FIG. 11 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the Partial Occupusion context in the UAV123 dataset;
FIG. 12 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the UAV123 data set Out-of-View (Out-of-View) context;
FIG. 13 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of a Background around the UAV123 dataset;
FIG. 14 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of the drilling Variation in the UAV123 data set;
figure 15 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of a Viewpoint Change in the UAV123 dataset;
FIG. 16 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of Camera Motion in the UAV123 data set;
FIG. 17 is a graph of tracking accuracy and success rate for the present invention and other algorithms in the context of Similar objects in the UAV123 data set.
Detailed Description
In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
Referring to fig. 1 and fig. 2, two flowcharts of a scale-adaptive long-term correlation target tracking method in an embodiment of the present invention are respectively shown, where the method includes the following steps:
pretreatment: and reading the target position of the first frame to obtain a large background detection frame, a small background detection frame and a Gaussian regression label corresponding to each detection frame.
Processing of the video of the first frame: extracting target direction gradient Histogram (HOG) features according to the detection frame, and performing Fourier transform to a frequency domain to obtain xfAdding a Gaussian kernel function to obtain a Gaussian response k in the frequency domainfWherein the Gaussian kernel function is
Figure BDA0002971725640000071
Both regression models are mapped to kernel space, defined as k (x, x ') ═ Φ (x) · Φ (x'). Using xf and kfCalculating classifier parameters to obtain a time context regression model RcTarget appearance regression model RtAnd a detector Drf, wherein RcIs responsible for translation estimation, RtResponsible for the scale estimation, DrfIs responsible for the retesting.
Processing of the current frame (tth frame):
(1) target position (x) according to t-1 th framet-1，yt-1) Cutting a search window in the t frame, extracting HOG characteristics, and training a related filter template;
wherein, the trained relevant filter template specifically comprises:
Figure BDA0002971725640000072
wherein w denotes a correlation filter, xm,nRepresenting image block x having M x N pixels, y (M, N) being represented by xm,nAs a gaussian sample label generated for the training sample,
Figure BDA0002971725640000073
representing to kernel spaceλ represents a regularization parameter.
The above formula can also be written as w ═ Σm,na(m,n)φ(xm,n)。
Wherein the coefficient a is defined by the formula:
Figure BDA0002971725640000081
f denotes the discrete Fourier operator, xm,nThe image block x is represented by m × n pixels, and x and y represent pixel coordinates.
(2) Performing translation estimation using a context regression model RcAnd the associated filter fraction computing response ytAnd estimating the position (x) of the t-th framet，yt)；
The tracking task may calculate the correlation map by means of the image blocks z of a new one of the image frames of size m x n. The response of step (2) may be determined according to the following equation:
Figure BDA0002971725640000082
wherein f denotes a learned target appearance model,. indicates a Hadamard product, by finding
Figure BDA0002971725640000083
Finds the predicted position of the target.
(3) As shown in FIG. 3, a scale pool is constructed using a multi-scale search strategy, with a relevant mapping ysAnd a target appearance regression model RtSelf-adaptive estimation of optimal scale
Figure BDA0002971725640000084
Obtaining the initial predicted target state of the t-th frame
Figure BDA0002971725640000085
Wherein, the scale search strategy is as follows: the dimension of the template is fixed as ST＝(sx,sy) The scale pool is set to S ═ t1,t2,…tkFor the current frame at tist|tiE to S), k sizes are sampled to find a proper target scale, and then bilinear interpolation is adopted to enable the samples of all scales to become S and STThe size of the samples is consistent.
The final target scale response value is calculated as follows:
Figure BDA0002971725640000086
Figure BDA0002971725640000087
is the ith scale sample in the scale pool, with size tiSt. Through the process of bilinear interpolation,
Figure BDA0002971725640000088
will be adjusted to ST。
(4) Setting a first threshold τrSecond threshold τt，
Figure BDA0002971725640000089
A correlation map representing the prediction; if it is not
Figure BDA00029717256400000810
Use of DrfPerforming a redetection to find a set of candidate states X, for each state X in Xi'calculating confidence score y'iIf max (y'i)＞τtThen, then
Figure BDA00029717256400000811
Obtaining the final predicted target state of the t frame
Figure BDA00029717256400000812
(5) Updating model Rc；
(6) Setting a third threshold τaIf, if
Figure BDA00029717256400000813
Updating model Rt；
The steps of (A), (B), (C5) In (6), for Rc、RtThe model, the coefficients f and A in the model are updated frame by frame at a learning rate alpha as follows:
Figure BDA0002971725640000091
Figure BDA0002971725640000092
(7) update detector Drf；
In the examples of the present invention, DrfIs a Support Vector Machine (SVM) detector. For SVM, in each frame, a training set is given { (v)i,ci) 1, 2., N } and N samples, where viIs the feature vector generated by the ith sample, ciE { +1, -1} is a sample label, and the objective function for solving the hyperplane h of the SVM detector is:
Figure BDA0002971725640000093
wherein ,
Figure BDA0002971725640000094
<h,v>represents the inner product of h and v.
Hyperplane parameters are efficiently updated using a passive algorithm:
Figure BDA0002971725640000095
wherein ,
Figure BDA0002971725640000096
is the gradient of the loss function with respect to h, and τ e (0, ∞) is a hyper-parameter that controls the update rate of h.
Obtaining the final predicted target state of the t frame from the steps (2) to (4)
Figure BDA0002971725640000097
Obtaining R of the current frame from the steps (5) - (7)c、Rt and Drf；
(8) Repeating (1) to (7) for the t +1 th frame until the end of the video sequence.
According to the scale-adaptive long-term correlation target tracking method provided by the embodiment of the invention, a scale adaptive strategy and an LCT target tracking frame are effectively fused, and firstly, a scale pool is introduced, so that an algorithm can adaptively select the optimal scale for finding the position of a tracking target. The multi-scale search can be combined with the position estimation filter more stably, and the situation that the scale estimation is over-biased is not easy to occur.
Based on the above embodiments, the present embodiment provides a simulation experiment.
Simulation conditions are as follows: the emulation provided in this embodiment is performed in the Intel (R) core (TM) i3-4170CPU @3.70GHz 3.70GHz, hardware environment with 4.00GB memory, and software environment with MATLAB R2016 a. The experimental parameters were set as follows: regularization parameter λ 10-4The gaussian kernel σ is 0.1, the learning rate α is 0.01, and the threshold τ is set tor＝0.15，τa＝0.38，τtScale pool set to [1, 0.99, 1.01, ] 0.5]. The algorithm presented herein is then compared to LCT and other existing classical target tracking algorithms.
Simulation content: the method provided by the evaluation is carried out on a large reference data set UAV-123 containing 123 videos, and the evaluation mode selects one-pass success rate (OPE), namely, the target position given by the first frame starts to track, and the target position cannot be reinitialized after the tracking fails.
Fig. 4 to 17 are graphs of experimental results of the experiment, where the LCSA represents the scale-adaptive long-term correlation filtering tracking method proposed by the present invention, and the LCT, KCF _ GaussHog, CSK, IVT, and DFT represent other excellent target tracking algorithms, respectively. The LCSA algorithm provided by the invention has the score of 0.40 on the success rate curve comparison graph in FIG. 5 and the score of 0.58 on the accuracy curve comparison graph in FIG. 4. By taking a long-term correlation tracking algorithm (LCT) as a benchmark algorithm, the LCSA is improved by 2.56% in AUC success rate and is improved by 5.26% in accuracy compared with the LCT according to experimental data on the UAV-123. Although the accuracy is slightly lower than that of the LCT under the background speckle condition on the UAV123, the performance is improved compared with classical target tracking algorithms such as the LCT under classical target tracking scenes such as scale change, aspect ratio, low resolution, rapid motion, complete shielding, partial shielding, beyond visual field, illumination change, visual field point transition, camera motion and similar objects.
Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A scale-adaptive long-term correlation tracking algorithm, comprising the steps of:
(1) initializing a target detection frame; extracting the characteristics of the target according to the detection frame in the first frame, and initializing a time context regression model RcTarget appearance regression model RtAnd a detector Drf； wherein RcIs responsible for translation estimation, RtResponsible for the scale estimation, DrfIs responsible for the heavy detection;
(2) for the t-th frame, the target position (x) according to the t-1 th framet-1，yt-1) Cutting a search window in the t frame, extracting HOG characteristics, and training a related filter template;
(3) performing translation estimation using RcAnd calculating a response based on the correlation filter fraction obtained from the correlation filter template
Figure FDA0002971725630000011
And estimating the current frame position
Figure FDA0002971725630000012
(4) Constructing a scale pool using a multi-scale search strategy, mapping y by correlations and RtSelf-adaptive estimation of optimal scale
Figure FDA0002971725630000013
Obtaining the initial predicted target state of the t-th frame
Figure FDA0002971725630000014
(5) If it is not
Figure FDA0002971725630000015
Use of DrfPerforming double detection, finding candidate state set X, and for each state X 'in X'iCalculate confidence score y'iIf max (y'i)＞τtThen, then
Figure FDA0002971725630000016
i＝arg maxiy′i； wherein ,τrIs a first threshold value, τtIs the second threshold value, and is,
Figure FDA0002971725630000017
a correlation map representing the prediction; obtaining the final predicted target state of the t frame
Figure FDA0002971725630000018
(6) Updating Rc；
(7) If it is not
Figure FDA0002971725630000019
Updating Rt； wherein ,τaIs a third threshold;
(8) update Drf；
(9) And repeating (2) to (8) for the t +1 th frame until the video sequence is ended.
2. The scale-adaptive long-term correlation tracking algorithm according to claim 1, wherein the correlation filter template trained in step (2) is specifically:
Figure FDA0002971725630000021
wherein w denotes a correlation filter, xm,nRepresenting image blocks x having m x n pixels, y (m, n) being denoted xm,nAs a gaussian sample label generated for the training sample,
Figure FDA0002971725630000022
represents the mapping to kernel space and λ represents the regularization parameter.
3. The scale-adaptive long-term correlation tracking algorithm according to claim 1, wherein the correlation filter template trained in step (2) is specifically:
Figure FDA0002971725630000023
wherein the coefficient a is defined by the formula:
Figure FDA0002971725630000024
f denotes the discrete Fourier operator, xm,nThe image block x is represented by m × n pixels, and x and y represent pixel coordinates.
4. A scale-adaptive long-term correlation tracking algorithm according to claim 3, wherein the tracking task is to calculate a correlation map by means of image blocks z of a new frame in an image frame of size m x n, and said step (3) is responsive
Figure FDA0002971725630000025
Determined according to the following formula:
Figure FDA0002971725630000026
wherein f denotes a learned target appearance model,. indicates a Hadamard product, by finding
Figure FDA0002971725630000029
Finds the predicted position of the target.
5. The scale-adaptive long-term correlation tracking algorithm according to claim 1, wherein the scale search policy of step (4) is:
the dimension of the template is fixed as ST＝(sx,sy) The scale pool is set to S ═ t1,t2,…tkFor the current frame at tist|tiE S, sampling k sizes to find a proper target scale, and then adopting bilinear interpolation to enable the samples of all scales to become STThe size of the samples is consistent;
the final target scale response value is calculated as follows:
Figure FDA0002971725630000027
Figure FDA0002971725630000028
is the ith scale sample in the scale pool, with size tiSt。
6. The scale-adaptive long-term correlation tracking algorithm according to claim 3, wherein R is updatedcUpdate RtThe method comprises the following steps:
for Rc and RtThe coefficients f and a in the model are updated frame by frame at the learning rate α as:
Figure FDA0002971725630000031
Figure FDA0002971725630000032
7. the scale-adaptive long-term correlation tracking algorithm according to claim 1, wherein D isrfA support vector machine detector;
update DrfThe method comprises the following steps:
in each frame, a training set is given { (v)i,ci) 1, 2., N } and N samples, where viIs the feature vector generated by the ith sample, ciE { +1, -1} is a sample label, and the objective function for solving the support vector machine detector hyperplane h is:
Figure FDA0002971725630000033
wherein l (h, (v, c)) -max {0,1-c < h, v > }, < h, v > represents the inner product of h and v; λ represents a regularization parameter;
updating hyperplane parameters using a passive algorithm:
Figure FDA0002971725630000034
wherein ,
Figure FDA0002971725630000035
is the gradient of the loss function with respect to h, and τ e (0, ∞) is a hyper-parameter that controls the update rate of h.
8. The scale-adaptive long-term correlation tracking algorithm according to claim 1, wherein τ isrIs 0.15, τtIs 0.5, τaIs 0.38。
CN202110265773.2A 2021-03-11 2021-03-11 Scale-adaptive long-term correlation target tracking method Active CN113033356B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110265773.2A CN113033356B (en) 2021-03-11 2021-03-11 Scale-adaptive long-term correlation target tracking method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110265773.2A CN113033356B (en) 2021-03-11 2021-03-11 Scale-adaptive long-term correlation target tracking method

Publications (2)

Publication Number Publication Date
CN113033356A true CN113033356A (en) 2021-06-25
CN113033356B CN113033356B (en) 2023-10-17

Family

ID=76469899

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110265773.2A Active CN113033356B (en) 2021-03-11 2021-03-11 Scale-adaptive long-term correlation target tracking method

Country Status (1)

Country Link
CN (1) CN113033356B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113902773A (en) * 2021-09-24 2022-01-07 南京信息工程大学 Long-term target tracking method using double detectors
CN117292306A (en) * 2023-11-27 2023-12-26 四川迪晟新达类脑智能技术有限公司 Edge equipment-oriented vehicle target detection optimization method and device

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110569723A (en) * 2019-08-02 2019-12-13 西安工业大学 Target tracking method combining feature fusion and model updating
KR102175491B1 (en) * 2019-05-08 2020-11-06 군산대학교산학협력단 Method and apparatus for tracking object based on correlation filter

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102175491B1 (en) * 2019-05-08 2020-11-06 군산대학교산학협력단 Method and apparatus for tracking object based on correlation filter
CN110569723A (en) * 2019-08-02 2019-12-13 西安工业大学 Target tracking method combining feature fusion and model updating

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
张伟;温显斌;: "基于多特征和尺度估计的核相关滤波跟踪算法", 天津理工大学学报, no. 03 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113902773A (en) * 2021-09-24 2022-01-07 南京信息工程大学 Long-term target tracking method using double detectors
CN117292306A (en) * 2023-11-27 2023-12-26 四川迪晟新达类脑智能技术有限公司 Edge equipment-oriented vehicle target detection optimization method and device

Also Published As

Publication number Publication date
CN113033356B (en) 2023-10-17

Similar Documents

Publication Publication Date Title
CN108776975B (en) Visual tracking method based on semi-supervised feature and filter joint learning
CN111311647B (en) Global-local and Kalman filtering-based target tracking method and device
CN106981071B (en) Target tracking method based on unmanned ship application
CN110569723A (en) Target tracking method combining feature fusion and model updating
CN109859241B (en) Adaptive feature selection and time consistency robust correlation filtering visual tracking method
CN110120064B (en) Depth-related target tracking algorithm based on mutual reinforcement and multi-attention mechanism learning
CN111523447B (en) Vehicle tracking method, device, electronic equipment and storage medium
CN111582349B (en) Improved target tracking algorithm based on YOLOv3 and kernel correlation filtering
CN107368802B (en) Moving target tracking method based on KCF and human brain memory mechanism
CN112785622B (en) Method and device for tracking unmanned captain on water surface and storage medium
CN113033356B (en) Scale-adaptive long-term correlation target tracking method
CN110827327B (en) Fusion-based long-term target tracking method
CN110751670B (en) Target tracking method based on fusion
CN110689559B (en) Visual target tracking method based on dense convolutional network characteristics
CN110751671B (en) Target tracking method based on kernel correlation filtering and motion estimation
CN113706580B (en) Target tracking method, system, equipment and medium based on relevant filtering tracker
Zhang et al. Boosting the speed of real-time multi-object trackers
CN110751673B (en) Target tracking method based on ensemble learning
CN111612816B (en) Method, device, equipment and computer storage medium for tracking moving target
CN114842506A (en) Human body posture estimation method and system
CN110660079A (en) Single target tracking method based on space-time context
Huang et al. An anti-occlusion and scale adaptive kernel correlation filter for visual object tracking
Li et al. Adaptive object tracking algorithm based on eigenbasis space and compressive sampling
Sun et al. Correlation filter tracking algorithm based on multiple features and average peak correlation energy
US11821986B1 (en) Target tracking method, system, device and storage medium

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
GR01 Patent grant
GR01 Patent grant