CN114743124A - Real-time target tracking method for missile-borne platform - Google Patents

Real-time target tracking method for missile-borne platform Download PDF

Info

Publication number
CN114743124A
CN114743124A CN202210099550.8A CN202210099550A CN114743124A CN 114743124 A CN114743124 A CN 114743124A CN 202210099550 A CN202210099550 A CN 202210099550A CN 114743124 A CN114743124 A CN 114743124A
Authority
CN
China
Prior art keywords
target
frame
tracking
scale
missile
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
Application number
CN202210099550.8A
Other languages
Chinese (zh)
Inventor
吕梅柏
何菊
余桐
魏海瑞
刘晓东
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.)
Northwestern Polytechnical University
Original Assignee
Northwestern Polytechnical 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 Northwestern Polytechnical University filed Critical Northwestern Polytechnical University
Priority to CN202210099550.8A priority Critical patent/CN114743124A/en
Publication of CN114743124A publication Critical patent/CN114743124A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Aiming, Guidance, Guns With A Light Source, Armor, Camouflage, And Targets (AREA)

Abstract

The invention provides a real-time target tracking method for a missile-borne platform, which belongs to the field of image processing and comprises the following steps: step 1, a target tracker extracts a search area of a current image frame according to a target position of a first frame; step 2, extracting the characteristics of the search area extracted from the first frame, training a classifier by using the sample and the characteristics, and generating an initial position filter; step 3, in the subsequent image frame, the target tracker performs area cycle sampling according to the target position of the previous frame to obtain positive and negative samples, and extracts the sample characteristics to update the filter parameters; and 4, performing correlation calculation on the filter and the newly input image frame, wherein the region with the maximum response value is the target position. The method adopts a method of taking tracking as a main part and taking detection as an auxiliary part to fuse an improved related filtering algorithm and the improved single-stage detection algorithm to obtain an excellent long-term tracking system, and the algorithm expressive force is greatly improved.

Description

Missile-borne platform real-time target tracking method
Technical Field
The invention belongs to the field of image processing, and particularly relates to a real-time target tracking method for a missile-borne platform.
Background
The traditional tracking algorithm is widely applied to practical scenes in the industry, but the algorithm has a short board in complex background interference and long-term target tracking. An intelligent target tracking algorithm based on a twin network is good in performance under complex background and similar interference, but the algorithm is poor in real-time performance, and cannot be applied to an embedded mobile platform at present.
The classic target tracking algorithm is divided into two categories, namely a generating formula and a discriminating formula according to whether the algorithm has a target background classifying function, wherein the generating formula is established in a structure of a target feature subspace, and the discriminating formula is established on the basis of a classifying or regression mode to discriminate a target/background. Classical generative tracking algorithms include kalman filtering, particle filtering, mean-shift, and the like. The classical discriminant tracking algorithm is an MOSSE algorithm, Henriques et al in 2012 propose a CSK method, and a circulant matrix is introduced based on the MOSSE algorithm and is calculated and solved in a Fourier transform domain, so that the algorithm instantaneity is greatly improved. In 2014, the DSST algorithm proposed by MartinDanelljan combines the translation filtering with the scale filtering algorithm for the first time, and then a series of accelerated versions fdst appear. The KCF algorithm was proposed by Henriques et al in 2015.
The tracking algorithm based on deep learning can be said to be a classical target tracking algorithm which is rolled in a large number of ways in terms of precision. In 2015, the BohyungHan team designed a neural network (MDNet) with a multi-domain structure that performed well, but at a rate of only 1 frame/s. A twin network-based mountain-opening operation-SINT algorithm is proposed in 2016, and a Simm-FC algorithm is proposed in the same year. The Siamse-RPN network applies an RPN module in target detection to tracking, and an RPN sub-network is divided into two sub-modules of a classification target and a regression target. The DaSiamarPN method is used for further optimizing and improving the Siamese-RPN, a target segmentation task is introduced into target tracking by the SiamaMask algorithm, the optimal performance is achieved on the video tracking task, and the current fastest speed is achieved on the video target segmentation. Although the method achieves good performance on a public database, large-scale data is required to ensure the robustness of a trained tracking model, and the strong characterization capability of a deep network is obtained by sacrificing the instantaneity by relying on huge calculation amount, so that a long way is needed in the actual engineering application stage.
A deep convolutional network used by a deep learning-based single-stage detection algorithm has strong target characteristic representation power, an intelligent algorithm with a high recognition rate is applied to military technology instead of the traditional algorithm in the future, but huge counter propagation parameter calculation and an intricate network model are adopted, and resources such as data storage, power consumption, computational power and the like of a missile-borne platform are limited, so that the missile-borne platform is difficult to land on and achieve a real-time robust detection effect.
The traditional kernel correlation filtering tracking algorithm is high in speed but poor in long-term tracking effect, and the detection algorithm based on deep learning is high in precision but long in time consumption.
The related filtering tracking algorithm based on the CPU end in the early stage has strong real-time performance and wide engineering application, but has low tracking precision under the conditions of complex background interference, similar shielding and the like; a deep convolutional network used by a deep learning-based single-stage detection algorithm has strong target characteristic representation force, but the embedded realization power consumption is large, the real-time performance is low, and the good performance of intelligent detection is not exerted in engineering application such as actual military operation and the like. In order to guarantee the detection speed and simultaneously consider the detection precision, the development and research of a real-time target accurate detection tracking technology which can be used for a missile-borne GPU platform are strived to.
Therefore, the application provides a missile-borne platform real-time target tracking method.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a real-time target tracking method for a missile-borne platform.
In order to achieve the above purpose, the invention provides the following technical scheme:
a real-time target tracking method for a missile-borne platform comprises the following steps:
step 1, a target tracker extracts a search area of a current image frame according to a target position of a first frame;
step 2, extracting the characteristics of the search area extracted from the first frame, training a classifier by using the sample and the characteristics, and generating an initial position filter;
step 3, in the subsequent image frame, the target tracker performs area cycle sampling according to the target position of the previous frame to obtain positive and negative samples, and extracts the sample characteristics to update the filter parameters;
and 4, performing correlation calculation on the filter and the newly input image frame, wherein the region with the maximum response value is the target position.
Preferably, the step 1 specifically includes:
step 11, carrying out automatic target detection on a video frame shot by a missile-borne platform through a multi-scale and single-stage target detector, and taking an output result of the detector as initialization of a tracker;
and step 12, extracting a search area of the current image frame by the target tracker according to the target position of the first frame.
11. Preferably, the step 2 specifically includes:
step 21, after the hog-lbp-Fourier descriptor feature fusion is carried out on the search area extracted from the first frame, the fusion feature extraction is carried out;
step 22, training a classifier by using the samples and the characteristics, and generating an initial position filter;
step 23, in the subsequent image frame, the target tracker performs area cycle sampling according to the target position of the previous frame; meanwhile, with the target center position as a target center, zooming the candidate sample to be tested to obtain a scale pool space;
and 24, performing hog-lbp-Fourier descriptor feature fusion on the extracted area sample, and then performing fusion feature extraction.
Preferably, the step 4 specifically includes:
step 41, training a position filter according to the positive and negative samples, wherein the area with the maximum response of the position filter is the current target position; meanwhile, extracting target characteristics in a scale space by an algorithm to perform scale correlation filtering detection to obtain a response matrix, wherein the corresponding maximum peak value is the current target scale;
step 42, solving the current confidence score by using a secondary tracking confidence coefficient judging mechanism in the tracking process, and judging whether the tracking state of each frame of image is stable;
step 43: the confidence score is transmitted to the detector to determine whether to activate the detector, and the tracker is transmitted to determine whether to update the target template.
Preferably, the acquired image needs to be preprocessed before the target detection is performed.
Preferably, the target sample size selection principle:
Figure BDA0003491935420000041
wherein, W and H represent the width and height of the current frame target respectively, λ is the scale factor, and S is the total scale number.
Preferably, the kernel correlation filtering algorithm specifically includes:
the kernel correlation filtering algorithm maps ridge regression of a linear space to a nonlinear space through a kernel function, a dual problem and common constraints are solved in the nonlinear space, and meanwhile, a multichannel Histogram of Oriented Gradients (HOG) is adopted in the algorithm to better represent target characteristics;
let the training sample set be (x)i,yi) An error function of
min||Xw-y||2+λ||w||2
Wherein X represents an input sample cyclic matrix, w is a coefficient matrix to be solved, and lambda is used for controlling the structural complexity of the system;
to minimize the error function, the above equation is derived and the derivative is taken to be 0, and the following is found:
w＝(XHX+λI)-1XHy
let w be expressed as a linear combination of samples:
Figure BDA0003491935420000042
wherein
Figure BDA0003491935420000043
For a nonlinear mapping function, the solution of the ridge regression in kernel space is:
Figure BDA0003491935420000044
wherein
Figure BDA0003491935420000045
The expression x is self-correlated in the Fourier domain, and the output response equation obtained in the fast detection process is as follows:
Figure BDA0003491935420000046
it can be derived from the above formula that only the template training parameter α and the training sample set x need to be updated in the process of updating the tracker, and meanwhile, the template update factor η is set by using a linear interpolation method, and the specific update formula is as follows:
xt＝ηxt-1+(1-η)xt
αt＝ηαt-1+(1-η)αt。
preferably, the feature fusion specifically comprises:
for a single visual feature fi(i-1, 2,3) establishing a target single body with better universalityFeature submodel MfiConstructing a target model M based on multi-feature fusion by using a self-adaptive weighted fusion strategy I;
Figure BDA0003491935420000051
preferably, the scale filter performing the scale-dependent filtering detection includes:
firstly, determining a new position of a target in a new frame of image by using a position correlation filter;
then taking the current target center position as a target center, and taking the candidate sample X to be detected as a target centeriScaling to obtain a scale pool space { S }1,S2,S3,S4,S5}；
Extracting features to carry out scale correlation filtering detection to obtain a response matrix miAnd the corresponding maximum peak value is the current target scale.
Preferably, the principle of determining whether the tracking state of each frame of image is stable is as follows:
the first-stage discrimination is carried out according to the maximum peak value of the filtering, and the second-stage discrimination is carried out by introducing an average peak value correlation energy index to measure the tracking response effect; fmaxpAPCE being the maximum peak of the P frame picturepThe average peak correlation energy of the image of the P-th frame is obtained if the following condition is satisfied:
Fmaxp＜|u1±γσ1|2
APCEp＜|u2±γσ2|2
the tracking is stable; otherwise, the tracking fails.
The real-time target tracking method for the missile-borne platform provided by the invention has the following beneficial effects:
the invention establishes a target tracking-detection integrated long-term target tracker based on an embedded platform, and simultaneously ensures the detection speed of the algorithm and the detection precision.
According to the method, a target tracking confidence coefficient discriminator and a single-stage target detector are added based on a KCF kernel correlation filter, the detector is started to perform target relocation when the target is judged to be lost, a good long-term target tracker which mainly tracks and assists in detection is obtained, and the algorithm robustness is guaranteed while the detection frame rate is taken into account.
The invention aims to adopt a method which takes tracking as a main part and detection as an auxiliary part to fuse an improved related filtering algorithm and the improved single-stage detection algorithm to obtain an excellent long-term tracking system, thereby greatly improving the algorithm expressive force.
Drawings
In order to more clearly illustrate the embodiments of the present invention and the design thereof, the drawings required for the embodiments will be briefly described below. The drawings in the following description are only some embodiments of the invention and it will be obvious to a person skilled in the art that other drawings can be obtained from these drawings without inventive effort.
Fig. 1 is a flowchart of a missile-borne platform real-time target tracking method according to embodiment 1 of the present invention;
FIG. 2 is a scale filter;
FIG. 3 is a maximum response score discrimination chart;
FIG. 4 is an APCE decision diagram;
FIG. 5 is a graph of target scale change tracking results;
FIG. 6 shows the tracking result of the fast motion of the target;
FIG. 7 is a target deformation tracking result;
FIG. 8 shows the result of target occlusion tracking;
FIG. 9 shows the results of scale change tracking;
FIG. 10 shows the results of ray change tracking;
FIG. 11 is a network structure.
Detailed Description
In order that those skilled in the art will better understand the technical solutions of the present invention and can practice the same, the present invention will be described in detail with reference to the accompanying drawings and specific examples. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
Example 1
The invention provides a missile-borne platform tracking algorithm combining correlation filtering and machine learning, and a one-dimensional scale filter and a feature fusion method are added on the basis of a kernel correlation filtering type fast tracking algorithm. And establishing a secondary tracking confidence degree judging mechanism, introducing a target re-detection function, stopping template updating and starting a single-stage target detector when judging that the target tracking is lost. And a method which takes tracking as a main part and detection as an auxiliary part is adopted to fuse the improved correlation filtering algorithm and the detection algorithm to obtain an excellent long-term tracking system. The long-term tracking system consists of the following three parts: a single-stage target detector; an improved kernel correlation filter algorithm; a confidence arbiter is tracked.
The embodiment provides a method for tracking a real-time target of a missile-borne platform, which is specifically shown in fig. 1 to 11, and comprises the following steps:
step 1: a high-precision multi-scale and single-stage target detector is designed by self, the target detector performs automatic target detection on a video frame in an image acquired by the missile-borne platform through an intelligent detection algorithm, and a detection result is used as an initialization value of a target tracker;
step 2: the target tracker extracts a search area of a current image frame according to the target position of a first frame image in the detection result, and expands to generate a search frame;
and step 3: carrying out hog-lbp-Fourier descriptor feature fusion on a search area where the search box is located, and then carrying out fusion feature extraction;
and 4, step 4: training a classifier by using the fused features and generating an initial position filter;
and 5: in the subsequent image frame, the target tracker performs regional cycle sampling according to the target position of the previous frame; meanwhile, with the target center position as a target center, zooming the candidate sample to be tested to obtain a scale pool space;
step 6: performing hog-lbp-Fourier descriptor feature fusion on the extracted area sample again, and then performing fusion feature extraction;
and 7: training a position filter according to the positive and negative samples, wherein the area with the maximum position filtering response is the current target position; meanwhile, extracting target characteristics in a scale space by using a kernel correlation filtering algorithm, and performing scale correlation filtering detection through a scale filter to obtain a response matrix, wherein the corresponding maximum peak value is the current target scale;
target scale selection principle:
Figure BDA0003491935420000071
wherein, W and H represent the width and height of the current frame target respectively, λ is the scale factor, and S is the total scale number;
and 8: in the tracking process, a secondary tracking confidence coefficient judging mechanism is used for solving the current confidence degree score and judging whether the tracking state of each frame of image is stable or not; the first-stage discrimination discriminates according to the maximum peak value of filtering, and the second-stage discrimination introduces the correlation energy index of the average peak value to measure the tracking response effect; fmaxpAPCE being the maximum peak of the P frame picturepThe average peak correlation energy of the P-th frame image is obtained if the following conditions are satisfied:
Fmaxp＜|u1±γσ1|2 (2)
APCEp＜|u2±γσ2|2 (3)
the tracking is stable; otherwise, the tracking fails;
and step 9: transmitting the confidence score into a detector to judge whether the detector is started or not, and simultaneously transmitting the confidence score into a tracker to judge whether the target template is updated or not; if the target is lost, target relocation is carried out, and the step 1 is switched to; otherwise, go to step 5.
Specifically, the kernel correlation filtering algorithm in this embodiment specifically includes:
the kernel correlation filtering algorithm maps ridge regression of a linear space to a nonlinear space through a kernel function, solves dual problems and common constraints in the nonlinear space, and adopts a multichannel directional gradient Histogram (HOG) to better represent target characteristics;
let the training sample set be (x)i,yi) An error function of
min||Xw-y||2+λ||w||2 (4)
Wherein X represents an input sample cyclic matrix, w is a coefficient matrix to be solved, and lambda is used for controlling the structural complexity of the system;
to minimize the error function, the above equation is derived and the derivative is taken to be 0, and the following is found:
w＝(XHX+λI)-1XHy (5)
let w be expressed as a linear combination of samples:
Figure BDA0003491935420000081
wherein
Figure BDA0003491935420000082
For a nonlinear mapping function, the solution of the ridge regression in kernel space is:
Figure BDA0003491935420000083
wherein
Figure BDA0003491935420000084
The expression x is self-correlated in the Fourier domain, and the output response equation obtained in the fast detection process is as follows:
Figure BDA0003491935420000091
it can be derived from the above formula that only the template training parameter α and the training sample set x need to be updated in the process of updating the tracker, and meanwhile, the template update factor η is set by using a linear interpolation method, and the specific update formula is as follows:
xt＝ηxt-1+(1-η)xt (9)
αt＝ηαt-1+(1-η)αt (10)。
specifically, the feature fusion in this embodiment specifically includes:
for a single visual feature fi(i is 1,2,3), establishing a target single-feature sub-model M with better universalityfiConstructing a target model M based on multi-feature fusion by using a self-adaptive weighted fusion strategy I;
Figure BDA0003491935420000092
specifically, the performing of the scale-dependent filtering detection by the scale filter in this embodiment includes:
firstly, determining a new position of a target in a new frame of image by using a position correlation filter;
then taking the current target center position as a target center, and taking the candidate sample X to be detected as a target centeriScaling to obtain a scale pool space { S }1,S2,S3,S4,S5}；
Extracting features to carry out scale correlation filtering detection to obtain a response matrix miThe corresponding maximum peak value is the current target scale;
target scale selection principle:
Figure BDA0003491935420000093
wherein, W and H represent the width and height of the current frame target respectively, λ is the scale factor, and S is the total scale number.
A Biker video sequence in the OTB100 is adopted to perform tracking test, a target is lost when the 51 st frame of a video is lost, so that the tracking fails, and the experimental result is shown in FIG. 3 and FIG. 4:
FIG. 3 is a graph of the maximum response score discrimination result, with the abscissa representing the number of frames and the ordinate representing the maximum response score calculated for each frame; fig. 4 is a graph of the mean peak correlation energy results, with the abscissa representing frame number and the ordinate representing the mean peak correlation energy value. The analysis of fig. 3 and 4 is combined to conclude the following:
(1) when the target tracking is stable, the maximum response fraction and the average peak correlation energy value are large, for example: in the 34 th frame image, Fmax is 0.35, and APCE is 20.5;
(2) when the target tracking condition is poor, the maximum response fraction and the average peak correlation energy value are correspondingly reduced, for example: in the 48 th image, Fmax is 0.28, and APCE is 18.6;
(3) when the target is lost, the maximum peak value and the APCE value are sharply decreased, the values fluctuate sharply, and the target is lost at frame 51, Fmax is 0.21, and APCE is 15.9.
(4) After the target is lost, the correlation filtering tracker introduces background error information, and continues tracking with the background as the target, at this time, the maximum peak value gradually increases, and 63-frame images Fmax is 0.267, and APCE is 12.5.
From the above experiments, it can be known that the maximum response score and the average peak correlation energy value can be used as the criterion for the target tracking condition, FmaxpFor maximum response score of P frame image, APCEpThe average peak correlation energy of the image of the P-th frame is obtained if the following condition is satisfied:
Fmaxp＜|u1±γσ1|2(13)
APCEp＜|u2±γσ2|2 (14)
the tracking is stable; if a large mutation occurs, tracking fails.
The design of the target weight detector in this embodiment is as follows:
after the target loss is judged, the single-stage target detector is started, and the network structure is shown in fig. 11 and is specifically designed as follows:
(1) multi-scale prediction
When the size of the input image is 416 × 416, the four feature map scales obtained after multi-scale prediction are respectively as follows: (a)13 × 13, the feature map output size at the deepest level of the network is high, and the local receptive field of the feature map is large in this case, and therefore, is suitable for predicting a large target. (b)26 multiplied by 26, the scale output by the feature map at the bottom layer is spliced with the last feature map with the size of 26 multiplied by 26 in the network after 2 times of upsampling, and the feature map has a medium-scale receptive field and is more suitable for predicting a medium-size target. (c)52 × 52, as with the method of the scale b, finally takes the feature map with the size of 52 × 52 as an output, and the receptive field is smaller, so that the feature map is most suitable for predicting smaller targets. (d)104 × 104 takes the same operation as above, and finally takes the feature map with the size of 104 × 104 as an output, and the receptive field is the smallest, so that the feature map is most suitable for predicting the smallest target.
(2) Darknet-53 infrastructure network
The network uses a new deep network containing more convolution layers, contains 53 convolution layers and uses the residual error structure of ResNet for reference, the structure can accelerate the training speed of the network and improve the training effect, and when the number of layers of the network model is deepened, the simple structure can well solve the problem of degradation of the detection rate, so that the network structure can develop in a deeper direction.
(3) Classifier class prediction
Since in most scenarios the target object may contain multiple tags, Softmax assigns only one class tag to each target. Multiple logistic classifiers can be substituted for Softmax without reducing accuracy. Finally, in the training process, binary cross-entropy loss (Binarycross-entropyLoss) is selected for class prediction.
DBL is the basic component of the network, and is composed of a convolutional layer, a BN layer and an activation function.
And (2) resn: n represents a number, res1, res2, …, res8, etc., indicating how many res _ units are contained in the res _ block. By taking the residual structure in the ResNet algorithm as a reference, the network structure can be made deeper by using the structure.
concat: and (5) tensor splicing. Stitching the upsampling of the middle layer of darknet and the next layer expands the dimensionality of the tensor.
And obtaining an optimal training model after network training and optimization, performing online target detection to output a current lost target position, initializing a related filtering tracker, updating a template, and continuously performing continuous target tracking.
In order to solve the problems of target shielding and embedded transplantation, the missile-borne platform real-time target tracking method provided by the embodiment adds a scale filter to solve the problem of target scale change, adds a two-stage target tracking discrimination mechanism, stops template updating and starts a single-stage target weight detector when the target is judged to be lost, and adopts a method taking tracking as a main part and detection as an auxiliary part to fuse an improved related filtering algorithm and a single-stage detection algorithm to obtain an excellent long-term tracking system, so that the accuracy and the real-time performance of the algorithm are maximized, and the expression of the algorithm is greatly improved.
In order to compare the anti-interference performance of the tracker more intuitively, a SRDCF, DSST and KCF algorithm and the method provided by the invention are selected for qualitative experiment comparison, video tracking tests are carried out on a maritime data set shot by an external field and a part of OTB100 data set, algorithm performance tests are respectively carried out when the target size changes, the target moves rapidly, the target deforms seriously, is shielded or light rays change strongly, and the comparison tracking results are as follows:
comparing fig. 5 to fig. 10, it can be found that the tracking of the algorithm provided by the present invention is stable under the conditions of target scale change, light change, target occlusion, etc. In fig. 6, due to the abrupt change of the target displacement, the target is lost by other algorithms in the 78 th image frame, and the method provided by the invention can still correctly lock the target position and track and stabilize in the subsequent image frame. In the marine vessel tracking video in fig. 10, the artificial target is shielded at the 1940 th frame, and the KCF algorithm tracking failure can be seen, but the method provided by the invention adopts a repositioning mechanism to reposition the target for continuous tracking, so that the method has stronger anti-interference performance. FIG. 11 shows that when the target scale of the ship changes, the method provided by the present invention has better tracking accuracy after adding the scale filter.
The above-mentioned embodiments are only preferred embodiments of the present invention, and the scope of the present invention is not limited thereto, and any simple modifications or equivalent substitutions of the technical solutions that can be obviously obtained by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.

Claims (10)

1. A real-time target tracking method for a missile-borne platform is characterized by comprising the following steps:
step 1, a target tracker extracts a search area of a current image frame according to a target position of a first frame;
step 2, extracting the characteristics of the search area extracted from the first frame, training a classifier by using the sample and the characteristics, and generating an initial position filter;
step 3, in the subsequent image frame, the target tracker performs area cycle sampling according to the target position of the previous frame to obtain positive and negative samples, and extracts the sample characteristics to update the filter parameters;
and 4, performing correlation calculation on the filter and the newly input image frame, wherein the region with the maximum response value is the target position.
2. The missile-borne platform real-time target tracking method according to claim 1, wherein the step 1 specifically comprises:
step 11, carrying out automatic target detection on a video frame shot by a missile-borne platform through a multi-scale and single-stage target detector, and taking an output result of the detector as initialization of a tracker;
and step 12, the target tracker extracts a search area of the current image frame according to the target position of the first frame.
3. The missile-borne platform real-time target tracking method according to claim 2, wherein the step 2 specifically comprises:
step 21, performing hog-lbp-Fourier descriptor feature fusion on the search area extracted from the first frame, and then performing fusion feature extraction;
step 22, training a classifier by using the samples and the characteristics, and generating an initial position filter;
step 23, in the subsequent image frame, the target tracker performs area cycle sampling according to the target position of the previous frame; meanwhile, with the target center position as a target center, scaling the candidate sample to be tested to obtain a scale pool space;
and 24, performing hog-lbp-Fourier descriptor feature fusion on the extracted area sample, and then performing fusion feature extraction.
4. The missile-borne platform real-time target tracking method according to claim 3, wherein the step 4 specifically comprises:
step 41, training a position filter according to the positive and negative samples, wherein the area with the maximum response of the position filter is the current target position; simultaneously, extracting target characteristics in a scale space by an algorithm to perform scale correlation filtering detection to obtain a response matrix, wherein the corresponding maximum peak value is the current target scale;
step 42, solving the current confidence score by using a secondary tracking confidence discrimination mechanism in the tracking process, and judging whether the tracking state of each frame of image is stable;
step 43: the confidence score is transmitted to the detector to determine whether to activate the detector, and the tracker is transmitted to determine whether to update the target template.
5. The method of claim 4, wherein the acquired images need to be preprocessed before the target detection.
6. The missile-borne platform real-time target tracking method according to claim 5, wherein a target sample size selection principle is as follows:
Figure FDA0003491935410000021
wherein, W and H represent the width and height of the current frame target respectively, λ is the scale factor, and S is the total scale number.
7. The missile-borne platform real-time target tracking method according to claim 6, wherein the kernel-dependent filtering algorithm specifically comprises:
the kernel correlation filtering algorithm maps ridge regression of a linear space to a nonlinear space through a kernel function, a dual problem and common constraints are solved in the nonlinear space, and meanwhile, a multichannel Histogram of Oriented Gradients (HOG) is adopted in the algorithm to better represent target characteristics;
let the training sample set be (x)i,yi) An error function of
min||Xw-y||2+λ||w||2
Wherein X represents an input sample cyclic matrix, w is a coefficient matrix to be solved, and lambda is used for controlling the structural complexity of the system;
to minimize the error function, the above equation is derived and the derivative is taken to be 0, and the following is found:
w＝(XHX+λI)-1XHy
let w be expressed as a linear combination of samples:
Figure FDA0003491935410000022
wherein
Figure FDA0003491935410000034
For a nonlinear mapping function, the solution of the ridge regression in kernel space is:
Figure FDA0003491935410000031
wherein
Figure FDA0003491935410000035
The expression x is self-correlated in the Fourier domain, and the output response equation obtained in the fast detection process is as follows:
Figure FDA0003491935410000032
it can be derived from the above formula that only the template training parameter α and the training sample set x need to be updated in the process of updating the tracker, and meanwhile, the template update factor η is set by using a linear interpolation method, and the specific update formula is as follows:
xt＝ηxt-1+(1-η)xt
αt＝ηαt-1+(1-η)αt。
8. the missile-borne platform real-time target tracking method according to claim 7, wherein the feature fusion specifically comprises:
for a single visual feature fi(i is 1,2,3), and establishing a target single-feature submodel M with better universalityfiConstructing a target model M based on multi-feature fusion by using a self-adaptive weighted fusion strategy I;
Figure FDA0003491935410000033
9. the missile-borne platform real-time target tracking method according to claim 8, wherein the scale filter performing scale-dependent filter detection comprises:
firstly, determining a new position of a target in a new frame of image by using a position correlation filter;
then taking the current target center position as a target center, and taking the candidate sample X to be detected as a target centeriScaling to obtain a scale pool space { S }1,S2,S3,S4,S5}；
Extracting features to carry out scale correlation filtering detection to obtain a response matrix miAnd the corresponding maximum peak value is the current target scale.
10. The method for tracking the real-time target of the missile-borne platform according to claim 9, wherein the principle of judging whether the tracking state of each frame of image is stable is as follows:
the first-stage discrimination is carried out according to the maximum peak value of the filtering, and the second-stage discrimination introduces the related energy index of the average peak value to measure the tracking response effect; fmaxpAPCE being the maximum peak of the P frame picturepThe average peak correlation energy of the image of the P-th frame is obtained if the following condition is satisfied:
Fmaxp＜|u1±γσ1|2
APCEp＜|u2±γσ2|2
the tracking is stable; otherwise, the tracking fails.
CN202210099550.8A 2022-01-27 2022-01-27 Real-time target tracking method for missile-borne platform Pending CN114743124A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210099550.8A CN114743124A (en) 2022-01-27 2022-01-27 Real-time target tracking method for missile-borne platform

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210099550.8A CN114743124A (en) 2022-01-27 2022-01-27 Real-time target tracking method for missile-borne platform

Publications (1)

Publication Number Publication Date
CN114743124A true CN114743124A (en) 2022-07-12

Family

ID=82275319

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210099550.8A Pending CN114743124A (en) 2022-01-27 2022-01-27 Real-time target tracking method for missile-borne platform

Country Status (1)

Country Link
CN (1) CN114743124A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115359240A (en) * 2022-07-15 2022-11-18 北京中科思创云智能科技有限公司 Small target detection method, device and equipment based on multi-frame image motion characteristics
CN116702093A (en) * 2023-08-08 2023-09-05 海南智慧海事科技有限公司 Marine target positioning method based on big data fusion
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

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115359240A (en) * 2022-07-15 2022-11-18 北京中科思创云智能科技有限公司 Small target detection method, device and equipment based on multi-frame image motion characteristics
CN115359240B (en) * 2022-07-15 2024-03-15 北京中科思创云智能科技有限公司 Small target detection method, device and equipment based on multi-frame image motion characteristics
CN116702093A (en) * 2023-08-08 2023-09-05 海南智慧海事科技有限公司 Marine target positioning method based on big data fusion
CN116702093B (en) * 2023-08-08 2023-12-08 海南智慧海事科技有限公司 Marine target positioning method based on big data fusion
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

Similar Documents

Publication Publication Date Title
Wang et al. Detecting faces using region-based fully convolutional networks
Xu et al. Segregated temporal assembly recurrent networks for weakly supervised multiple action detection
Wang et al. Face r-cnn
CN114743124A (en) Real-time target tracking method for missile-borne platform
Liu et al. A robust boosting tracker with minimum error bound in a co-training framework
CN109255289B (en) Cross-aging face recognition method based on unified generation model
Li et al. Graph mode-based contextual kernels for robust SVM tracking
CN112836639A (en) Pedestrian multi-target tracking video identification method based on improved YOLOv3 model
Chen et al. Learning linear regression via single-convolutional layer for visual object tracking
CN109190544B (en) Human identity recognition method based on sequence depth image
US11908457B2 (en) Orthogonally constrained multi-head attention for speech tasks
CN103902960A (en) Real-time face recognition system and method thereof
Naghavi et al. Integrated real-time object detection for self-driving vehicles
Luo et al. SFA: small faces attention face detector
Xia et al. Face occlusion detection using deep convolutional neural networks
CN112733942A (en) Variable-scale target detection method based on multi-stage feature adaptive fusion
CN109697727A (en) Method for tracking target, system and storage medium based on correlation filtering and metric learning
KR102272921B1 (en) Hierarchical object detection method for extended categories
Dewi et al. Deep Learning for Advanced Similar Musical Instrument Detection and Recognition
Prasad et al. A fast and self-adaptive on-line learning detection system
Kaur et al. Hybrid classification method for the human activity detection
Shangguan et al. Improved region proposal network for enhanced few-shot object detection
Khurshid et al. Incremental multi-model dictionary learning for face tracking
Mozaffari et al. Facial expression recognition using deep neural network
Zhang et al. Rt-track: robust tricks for multi-pedestrian tracking

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