CN114119666A - Irregular frame target tracking system and method adopting polar coordinate transformation method - Google Patents
Irregular frame target tracking system and method adopting polar coordinate transformation method Download PDFInfo
- Publication number
- CN114119666A CN114119666A CN202111422525.0A CN202111422525A CN114119666A CN 114119666 A CN114119666 A CN 114119666A CN 202111422525 A CN202111422525 A CN 202111422525A CN 114119666 A CN114119666 A CN 114119666A
- Authority
- CN
- China
- Prior art keywords
- filter
- characteristic
- image
- matrix
- feature
- 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 51
- 230000001788 irregular Effects 0.000 title claims abstract description 24
- 238000011426 transformation method Methods 0.000 title claims abstract description 20
- 240000004282 Grewia occidentalis Species 0.000 claims abstract description 103
- 238000001914 filtration Methods 0.000 claims abstract description 77
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 34
- 230000004044 response Effects 0.000 claims abstract description 33
- 238000005070 sampling Methods 0.000 claims abstract description 26
- 238000012952 Resampling Methods 0.000 claims abstract description 25
- 238000000605 extraction Methods 0.000 claims abstract description 25
- 238000003199 nucleic acid amplification method Methods 0.000 claims abstract description 12
- 230000003321 amplification Effects 0.000 claims abstract description 11
- 239000011159 matrix material Substances 0.000 claims description 87
- 230000006870 function Effects 0.000 claims description 48
- 238000001228 spectrum Methods 0.000 claims description 43
- 230000003190 augmentative effect Effects 0.000 claims description 34
- 230000009466 transformation Effects 0.000 claims description 31
- 238000012545 processing Methods 0.000 claims description 12
- 238000007781 pre-processing Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 6
- 230000007547 defect Effects 0.000 abstract description 2
- 230000000694 effects Effects 0.000 abstract description 2
- 230000015556 catabolic process Effects 0.000 abstract 1
- 238000006731 degradation reaction Methods 0.000 abstract 1
- 238000004364 calculation method Methods 0.000 description 23
- 239000000126 substance Substances 0.000 description 14
- 238000005457 optimization Methods 0.000 description 12
- 238000013519 translation Methods 0.000 description 8
- 230000008859 change Effects 0.000 description 7
- 238000012549 training Methods 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000003909 pattern recognition Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 102000002274 Matrix Metalloproteinases Human genes 0.000 description 2
- 108010000684 Matrix Metalloproteinases Proteins 0.000 description 2
- 230000003416 augmentation Effects 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 239000004576 sand Substances 0.000 description 2
- 230000035939 shock Effects 0.000 description 2
- 238000003786 synthesis reaction Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004883 computer application Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000002558 medical inspection Methods 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- 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
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- Computing Systems (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an irregular frame target tracking system and method adopting a polar coordinate transformation method, which are combined with a related filtering method to realize generation of an irregular target frame estimation result. Cutting the image according to the tracking target, and performing feature extraction and cosine window windowing; then, filtering by using a continuous correlation filter, and calculating by a Newton method to obtain an estimated target center point coordinate; generating four-corner point coordinates of a tracking target, and sampling a quadrilateral area formed by the generated four-corner point coordinates; and resampling the four corner point coordinates after the amplification and generating a continuous correlation filter and a scale and angle filter. The method based on continuous correlation filtering improves the defects of model degradation, boundary effect, response distortion and the like caused by the assumption that pixels of different feature extraction results are uniformly distributed in other correlation filtering methods, and improves the accuracy and robustness of the target tracking algorithm.
Description
Technical Field
The invention relates to the design problem of a single target tracking algorithm under monocular vision, in particular to an irregular frame target tracking system and method adopting a polar coordinate transformation method.
Background
Visual Object Tracking (Visual Object Tracking) is a challenging problem in the field of computer vision. The purpose of visual target tracking is to identify a tracking target from a continuous video sequence and extract a target position, so as to obtain a motion track of the target. Since target tracking involves recognition and optimization problems, target identification requires the application of optimization methods and control techniques. The main problems involved in target tracking include: (1) establishing a model, wherein the model refers to an appearance model of a tracking target, and the part determines the tracking effect of the target to a great extent; (2) an image input, wherein the image comprises a group of element view inputs, the view provides a scene comprising an object and a constraint between input image characteristics, and the part mainly relates to the sampling of the image, the extraction of the characteristics and the definition of the frame range of the object; (3) and detecting the target, namely utilizing a filter model to search and input to track the position of the target in the image.
Through decades of research and development, computer-based target detection and tracking have been applied to video surveillance, autopilot, unmanned aerial vehicles, human-computer interaction, medical inspection, robotics, autonomous navigation, industrial inspection and assembly, and some occasions where accurate measurements are required. In addition, the target tracking technology is also applied to the fields of video-based motion analysis and synthesis, identification based on motion information, image retrieval, hydrological observation, port management, medical image analysis, remote measurement, part quality detection and the like.
In the actual tracking algorithm development process, many problems still exist, such as: 1) tracking drift problems. Due to the complex application scene of target tracking, noise with different ranges is always introduced into the target appearance model in the training process, and the tracking drift problem is caused by the accumulation of the noise; 2) aspect ratio problems. The traditional target tracking algorithm is based on the selection of a fixed aspect ratio and a scale pool aiming at the description of target scale change, so that the predicted target scale information influences the tracking precision and causes the loss of a tracked target. In view of the above, research into further development of object tracking technology is still of great value to scientific research in the fields of computer vision and pattern recognition.
In addition, the most similar comparison documents to the present invention are: jizhang, Ningwang, full convolution twin network target tracking algorithm with rotation and scale estimation [ J ] computer application 2021,41(9):7. The main defects of the prior art are as follows: (1) in the prior art, a twin convolutional neural network method is adopted to construct a target tracking algorithm framework, but due to the intervention of a deep learning method, the calculation amount of the algorithm is increased, a large amount of off-line training is needed to ensure the accuracy of the algorithm, and the calculation speed of on-line tracking is only 15 frames per second and is far lower than the real-time requirement of 30 frames per second; (2) the twin network adopted in the prior art is a typical deep learning based approach. Due to the characteristics of deep learning, the ductility of the method is low, and the method cannot be combined with some traditional methods which can obviously reduce the calculation amount and improve the prediction accuracy, so that the precision and the real-time performance of the algorithm are reduced, and a good improvement method is provided; (3) the scale and angle prediction method in the prior art is stopped at single-corner point prediction, and a quadrilateral frame with a fixed frame aspect ratio is adopted for rotation and scale prediction of a tracking target, so that frames generated by an algorithm are all rectangular or parallelogram, and in an actual tracking scene, irregular quadrilaterals are more, so that the accuracy and robustness of the algorithm are limited by the assumption of the fixed frame aspect ratio of the single-corner point prediction method, and the accuracy cannot be improved in a breakthrough manner.
Disclosure of Invention
The purpose of the invention is as follows: the invention provides an irregular frame target tracking system and method adopting a polar coordinate transformation method aiming at the aspect ratio problem, wherein an augmented four-corner joint prediction method is utilized to obtain a high-confidence irregular frame, a continuous filtering frame is combined, and an algorithm is improved aiming at the tracking drift problem by a method of training a layered filter, so that tracking target information returned by the algorithm has high confidence level, and the accuracy of a tracking result is improved.
The technical scheme is as follows: in order to achieve the purpose, the invention specifically adopts the following technical scheme:
the invention provides an irregular frame target tracking system adopting a polar coordinate transformation method, which comprises the following steps: an initial image sampling module for initializing parameters in the continuous correlation filtering algorithm and tracking the given position of the target, namely the target center point coordinate (x) in the first framec,yc) Cutting and sampling the image under the Cartesian coordinate system to obtain a sample image block x1(ii) a A cosine window windowing module for sampling image block x1Extracting features, and then performing series connection and cosine window windowing on the extracted features to obtain an image sample z'; the continuous correlation filtering module is used for obtaining a characteristic spectrum Z ' by carrying out fast Fourier transform on the characteristic Z ' of the image sample, then carrying out filtering processing on the characteristic spectrum Z ' by using a continuous correlation filter h to obtain a continuous filtering response spectrum r, and then calculating the continuous filtering response spectrum r by using a Newton method to obtain an estimated target central point coordinateAn irregular border estimation module to estimate coordinates of a target center point by using the estimated coordinatesGenerating four-corner coordinates of a tracking target, and sampling and preprocessing a quadrilateral region formed by the generated four-corner coordinates to obtain the coordinate values of the augmented four-corner pointsA continuous correlation filter generation module for generating the four-corner coordinate values after being amplifiedResampling for continuous correlation filter generation to obtain resampled image samplesAnd using the resampled image samplesGenerating a filter f to obtain a generated filterA scale and angle filter generation module to generate a scale and angle filter by using the resampled image samplesTo filterGenerating to obtain the generated scale and angle filterA finishing module for processing the augmented four corner point coordinate valuesGenerated filterAnd a generated scale and angle filterReserving and amplifying the four-corner coordinate valueApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd the video sequence is carried into a cosine window windowing module for operation, and the operation from the cosine window windowing module to an ending module is executed in a circulating way on each next frame until the video sequence is ended.
Further, in the irregular border estimation module, preprocessing is performedThe reason is as follows: carrying out feature extraction and windowing on the sampled quadrilateral area to obtain a single angular point feature map Z' for scale prediction; carrying out four-corner point rotation on the single corner point feature map Z' to obtain an augmented four-corner point feature spaceFor the feature space of the four corner points after the enlargementContinuous correlation filtering is carried out, and the coordinate value of the four corner points after being amplified is obtained through the coordinate value S' of the amplification
Further, in the continuous correlation filter generation module, the coordinate values of the four corner points after the amplification are usedResampling for successive correlation filter generation to obtain a correlation signal having w ═ wg+wh+wnWindowed resampled image samples for layer featuresWherein wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is shown, and w is the number of synthetic characteristic layers;
the generation expression of the filter f obtained according to the continuous correlation filtering theory is
Wherein, WHIs a conjugate transpose of the weight matrix W of the filter f, AHIs a conjugate inversion matrix of A matrix, the expression of A matrix isGamma is opposite angleA line weight matrix of the expressionGaussian type mark matrixIs expressed as
Resampling image samplesSubstituting the formula to generate a filter f to obtain a generated filter
Furthermore, in the scale and angle filter generation module, a single angle point filter is obtained according to the continuous correlation filtering theoryIs expressed as
Therein, <' >-1Is a matrix dot product operation, which is the inverse operation of the matrix dot product operation, k represents the k-th feature channel,is a linear combination of the frames preceding the current frame,is composed ofA result matrix after Fourier transformation, andis composed ofThe conjugate matrix of (a) is determined,is the result of x after polar transformation, andis composed ofThe result matrix after feature extraction, | · | is that each element in the matrix is respectively solved for its modulus;
four corner points are enlarged to obtain
Wherein k represents the kth characteristic channel, and upsilon is enlarged through four corner points
Resampling image samplesSubstituting the above formula for the filterGenerating to obtain the generated scale and angle filter
Furthermore, in the cosine window windowing module, the sample image block x is processed1Performing a feature extraction operation, the extracted features sequentially comprising: gray, HOG and CN; obtaining characteristics according to the characteristics GrayAtlas zgObtaining a characteristic map z according to the characteristic HOGhObtaining a characteristic map z according to the characteristic CNnWhereinwi(i∈[g,h,n]) Is a characteristic spectrum ziThe number of feature channels of (a); connecting the features Gray, the HOG and the CN in series according to the dimension of the number of the feature channels to obtain an image sampleWherein w ═ wg+wh+wn,wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is shown, and w is the number of synthetic characteristic layers; point-by-point two-dimensional cosine window cos for image sample zwindowThe window operation is expressed as z ═ z-windowThe as is a matrix dot product operation,and z' is a feature spectrum, cos, windowed with a cosine windowwindowIs obtained by a function form of the cosine window, and the expression is
coswindow=h(m)h(n)′
Wherein m and n are respectively the length and width of the frame of the tracked target, and h (m) and h (n) are both one-dimensional cosine window generating functions h (N), and the expression is
The period of the cosine window function is 2 pi/N, with N-m or N-N.
In addition, the invention provides an irregular frame target tracking method adopting a polar coordinate transformation method, which comprises the following steps: step 1, initializing parameters in a continuous correlation filtering algorithm, and tracking a given position of a target, namely a target central point coordinate (x) according to a first framec,yc) Cutting and sampling the images under the Cartesian coordinate system to obtain samplesThe image block x1(ii) a Step 2, sample image block x1Extracting features, and then performing series connection and cosine window windowing on the extracted features to obtain an image sample z'; step 3, the characteristic Z ' of the image sample is subjected to fast Fourier transform to obtain a characteristic spectrum Z ', the characteristic spectrum Z ' is subjected to filtering processing by using a continuous correlation filter h to obtain a continuous filtering response spectrum r, and the continuous filtering response spectrum r is calculated by a Newton method to obtain an estimated target central point coordinateStep 4, by using the estimated coordinates of the target center pointGenerating four-corner point coordinates of a tracking target, and sampling and preprocessing a quadrilateral region formed by the generated four-corner point coordinates to obtain the augmented four-corner point coordinatesStep 5, the four-corner coordinate value after being augmented is processedResampling for continuous correlation filter generation to obtain resampled image samplesAnd using the resampled image samplesGenerating a filter f to obtain a generated filterStep 6, using the resampled image sampleTo filterGenerating to obtain the generated scale and angle filterStep 7, the coordinate values of the four corner points after being augmented are processedGenerated filterAnd a generated scale and angle filterReserving and amplifying the coordinate values of the four corner pointsApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd carrying out operation in the step 2, and circularly performing the operation from the step 2 to the step 7 for the next frames until the video sequence is finished.
Further, in step 4, the pretreatment is: carrying out feature extraction and windowing on the sampled quadrilateral area to obtain a single angular point feature map Z' for scale prediction; performing four-corner point rotation on the single corner point feature map Z' to obtain an augmented four-corner point feature spaceFor the feature space of the four corner points after the enlargementCarrying out continuous correlation filtering, and obtaining the coordinate values of the four corner points after being amplified through the coordinate value S' after being amplified
Further, in the above-mentioned case,in step 5, the augmented four corner point coordinate values are usedResampling for successive correlation filter generation to obtain a correlation signal having w ═ wg+wh+wnWindowed resampled image samples for layer featuresWherein wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is shown, and w is the number of synthetic characteristic layers;
the generation expression of the filter f obtained according to the continuous correlation filtering theory is
Wherein, WHIs a conjugate transpose of the weight matrix W of the filter f, AHIs a conjugate inversion matrix of A matrix, the expression of A matrix isGamma is a diagonal weight matrix with the expression ofGaussian type mark matrixIs expressed as
Resampling image samplesSubstituting the formula to generate a filter f to obtain a generated filter
Further, in step 6, a single corner filter is obtained according to the continuous correlation filtering theoryIs expressed as
Therein, <' >-1Is a matrix dot product operation, which is the inverse operation of the matrix dot product operation, k represents the k-th feature channel,is a linear combination of the frames preceding the current frame,is composed ofA result matrix after Fourier transformation, andis composed ofThe conjugate matrix of (a) is determined,is the result of x after polar transformation, andis composed ofThe result matrix after feature extraction, | · | is that each element in the matrix is respectively solved for its modulus;
four corner points are enlarged to obtain
Wherein k represents the kth characteristic channel, and upsilon is enlarged through four corner points
Resampling image samplesSubstituting the above formula for the filterGenerating to obtain the generated scale and angle filter
Further, in step 2, the sample image block x1Performing a feature extraction operation, the extracted features sequentially comprising: gray, HOG and CN; obtaining a characteristic map z according to the characteristic GraygObtaining a characteristic map z according to the characteristic HOGhObtaining a characteristic map z according to the characteristic CNnWherein wi(i∈[g,h,n]) Is a characteristic spectrum ziThe number of feature channels of (a); connecting the features Gray, the HOG and the CN in series according to the dimension of the number of the feature channels to obtain an image sampleWherein w ═ wg+wh+wn,wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnIs the number of CN characteristic layers, w is the synthesis characteristicThe number of layers; point-by-point two-dimensional cosine window cos for image sample zwindowThe window operation is expressed as z ═ z-windowThe as is a matrix dot product operation,and z' is a feature spectrum, cos, windowed with a cosine windowwindowIs obtained from the function form of cosine window, and the expression is
coswindow=h(m)h(n)′
Wherein m and n are respectively the length and width of the frame of the tracked target, and h (m) and h (n) are both one-dimensional cosine window generating functions h (N), and the expression is
The period of the cosine window function is 2 pi/N, with N-m or N-N.
Has the advantages that:
(1) the single-corner-point scale and rotation prediction model used by the invention adopts logarithmic polar coordinate transformation, and solves the joint prediction problem of target rotation and scale transformation. Because the continuous correlation filtering framework can apply the Meisen theorem in the signalology, the calculation process can be simplified on the basis of ensuring the calculation precision;
(2) according to the method, a high-confidence irregular frame four-corner-point prediction model with a non-fixed aspect ratio is added to a single-corner-point-scale and rotation-based prediction model, so that the problem of the fixed aspect ratio of a related filtering algorithm is solved, and the overall prediction precision of the algorithm is improved;
(3) the invention adopts a target tracking algorithm framework based on continuous correlation filtering. Due to the application of the cyclic matrix and the frequency domain calculation, the frame reduces the calculation amount and can ensure the real-time performance of tracking.
The algorithm has robustness, accuracy and real-time performance, can withstand the challenges of practical application scenarios, and has practical applicability. The invention provides a new solution for the aspect ratio relation of the frame, can form a unique irregular quadrilateral prediction frame, can meet the frame precision requirement in practical application, has low requirement on hardware by an algorithm, and has economic application value.
Drawings
FIG. 1 is a flowchart of an irregular frame target tracking method using polar coordinate transformation according to the present invention.
Detailed Description
The following description and the accompanying drawings are included to explain embodiments of the present invention in detail.
Example one
The embodiment discloses an irregular frame target tracking system adopting a polar coordinate transformation method, which is shown in fig. 1. Adopt irregular frame target tracking system of polar coordinate transform method, include: an initial image sampling module for initializing parameters in the continuous correlation filtering algorithm and tracking the given position of the target, namely the target center point coordinate (x) in the first framec,yc) Cutting and sampling the image under the Cartesian coordinate system to obtain a sample image block x1(ii) a A cosine window windowing module for sampling image block x1Extracting features, and then performing series connection and cosine window windowing on the extracted features to obtain an image sample z'; the continuous correlation filtering module is used for obtaining a characteristic spectrum Z ' by fast Fourier transform of the image sample characteristic Z ', then filtering the characteristic spectrum Z ' by using a continuous correlation filter h to obtain a continuous filtering response spectrum r, and then calculating the continuous filtering response spectrum r by a Newton method to obtain an estimated target central point coordinateAn irregular border estimation module for estimating the coordinates of the center point of the object by using the estimated coordinatesGenerating four-corner point coordinates of a tracking target, and sampling and preprocessing a quadrilateral area formed by the generated four-corner point coordinates to obtain the coordinate values of the four-corner points after the four-corner points are enlargedA continuous correlation filter generation module for generating the four corner point coordinate values after the amplificationResampling for continuous correlation filter generation to obtain resampled image samplesAnd using the resampled image samplesGenerating a filter f to obtain a generated filterA scale and angle filter generation module to generate a scale and angle filter by using the resampled image samplesTo filterGenerating to obtain the generated scale and angle filterA finishing module for processing the augmented four corner point coordinate valuesGenerated filterAnd a generated scale and angle filterReserving and amplifying the coordinate values of the four corner pointsApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd the video sequence is carried into a cosine window windowing module for operation, and the operation from the cosine window windowing module to an ending module is executed in a circulating way on each next frame until the video sequence is ended.
In the initial image sampling module, the parameters of each part involved in the relevant filtering algorithm are initialized. From a given position of the tracked object in the first frame (object center point coordinates (x)c,yc) Length and width m, n) of object frame), cut out the sampling to the image under the cartesian coordinate system, the scope is four angular points: upper left corner point s1=(xc+m/2,yc+ n/2), upper right corner point s2=(xc-m/2,yc+ n/2), lower right corner point s3=(xc-m/2,yc-n/2), lower left corner point s4=(xc+m/2,yc-n/2) defining a quadrangular area. The set of coordinate points may be represented as: s ═ S1 s2 s3 s4]T. Cropped sample image block x1Will be used to train the correlation filter after superimposing the cosine window.
In the cosine window windowing module, the sample image block x1Performing a feature extraction operation, the extracted features comprising in order: gray, HOG and CN; obtaining a characteristic map z according to the characteristic GraygObtaining a characteristic map z according to the characteristic HOGhObtaining a characteristic map z according to the characteristic CNnWherein wi(i∈[g,h,n]) Is a characteristic spectrum ziThe number of feature channels of (a); connecting the features Gray, the HOG and the CN in series according to the dimension of the number of the feature channels to obtain an image sampleWherein w ═ wg+wh+wn,wgFor the number of Gray feature layers, whFor the HOG feature layerNumber, wnThe number of CN characteristic layers is, and w is the number of synthetic characteristic layers; point-by-point two-dimensional cosine window cos for image sample zwindowThe window operation is expressed as z ═ z-windowThe as is a matrix dot product operation,and z' is a feature spectrum, cos, windowed with a cosine windowwindowIs obtained from the function form of cosine window, and the expression is
coswindow=h(m)h(n)′
Wherein m and n are respectively the length and width of the frame of the tracked target, and h (m) and h (n) are both one-dimensional cosine window generating functions h (N), and the expression is
The period of the cosine window function is 2 pi/N, with N-m or N-N. The function distribution characteristic of the one-dimensional cosine window is that h (N) is 0 when i is 0, h (N) is 1 when i is N/4, h (N) is 2 when i is N/2, h (N) is 1 when i is 3N/2, and h (N) is 0 when i is N, i shows a tendency that the cosine equation increases monotonically first and then decreases monotonically. And a two-dimensional cosine window cos obtained by the outer product of two one-dimensional cosine windowswindowThe method has the characteristic that the closer to the center, the larger the pixel value is, and the pixel values of a circle around the boundary are all 0.
In the continuous correlation filtering module, the image sample characteristics Z ' are subjected to fast Fourier transform to obtain a characteristic map spectrum Z ', and then the characteristic map Z ' is subjected to filtering processing by using a continuous correlation filter h to obtain a continuous filtering response map spectrum r, wherein the formula is
Wherein the content of the first and second substances,representing an analog Fourier transform (iFFT), < is a matrix dot product operation, and
then, calculating the continuous filtering response spectrum r by a Newton method to obtain the estimated coordinates of the target central point
In the irregular frame estimation module, the coordinates of the center point of the target to be estimated are obtainedThen, generating four-corner point coordinates, wherein the generated four-corner point coordinates are as follows: point on the upper left cornerUpper right corner pointLower right corner pointLower left corner pointThe set of coordinate points may be represented as: s ═ S'1 s′2 s′3 s′4]T. And then sampling a quadrilateral area formed by the generated four-corner point coordinates, and performing feature extraction and windowing operation to obtain a single corner point feature map Z' for scale prediction. Performing four-corner point rotation on Z' to obtain an augmented four-corner point feature spaceWhereinIs a spatial orthogonal superposition operator, and hasFor a single corner point correlation filter gsAlso carrying out amplification to obtain an amplified scale and angle estimation filterFormula pair of augmented four-corner point feature space by applying four-corner point joint prediction modelCarrying out continuous correlation filtering, and obtaining the coordinate values of the four corner points after being amplified by applying the formula of the amplified coordinate values S
The formula of the four-corner point joint prediction model and the formula of the augmented coordinate value S' are obtained in the following mode:
let the image I be in N × N size in Cartesian coordinate system with its upper coordinateHas a pixel value ofThe origin of the coordinate system is the upper left corner of the image, the positive direction of the x axis is the positive right direction, and the positive direction of the y axis is the positive down direction. Where i is 1,2, …, N. According to the definition, the coordinates of the pixel points at the upper left corner of the image I are (1,1), and the coordinates of the pixel points at the lower right corner of the image I are (N, N).
The polar coordinate representation of the image I can be regarded as the coordinate in the Cartesian coordinate systemTo the polar coordinate system, coordinateNon-linear, non-uniform transformation of (a). The polar coordinate is transformed into SfThen there is
In order to keep the integrity of the angle change of the polar coordinate, namely, the positive and negative values are uniformly distributed, and the image can be uniformly divided, namely, the image is diffused outwards from the datum point, and the values of all parts are basically uniform. The central point (x) of the image I under the Cartesian coordinate systemc,yc) As a reference point for polar coordinate transformation.
Under a polar coordinate system, an arbitrary coordinate point (rho)i,θi) Middle mode length rhoiIs expressed as
Phase thetaiIs expressed as
Wherein r is a reference standard direction vector, and the positive direction of the polar coordinate system is the same as the positive direction of the Cartesian coordinate system.
Image I (x) after polar coordinate changei,yi) Is an image I' (ρ)i,θi). In a Cartesian coordinate system, the phase angle θiHas a more concise expression form as shown in the following
In addition, an image I in a Cartesian coordinate system is divided into0Proceeding with die lengthAnd phaseAnd generating a changed image I1When, the following correspondence relationship is given
Correspondingly, the antipodal coordinate mark is
As shown in the two formulas, the change of the image scale and the rotation in the Cartesian coordinate system can be converted into the translational motion along the axis in the polar coordinate system, and based on the characteristic, the target scale and the rotation transformation in the target tracking can be more simply calculated and estimated.
Through polar coordinate system transformation, a prediction model of single-angle-point scale and angle transformation of the trapezoid tracking target frame under polar coordinates can be constructed. The response calculation to the correlation filtering in polar coordinates has
Wherein the content of the first and second substances,according to the s ═ p, theta pair IiSampling polar coordinates, wherein s' is the coordinate of s in a polar coordinate system, and the conversion formula is
Wherein, H and W are respectively the image I sampled according to the tracking target frameiLength and width of (d); rho' is the module length of s after the polar coordinate system transformation; and theta' is the phase of s after the polar coordinate system transformation.
The Fourier domain calculation used by the classic DCF target tracking framework can also be used for calculation under polar coordinate transformation, and the specific response calculation formula is
Wherein the content of the first and second substances,as a function of polar transformation, gsA model weight matrix is estimated for the scale and angle. According to the formulaThe scale and angle estimation under polar coordinates does not need to be converted and sampled for calculationAccording to the formulaOnly the weight matrix g of the scale and angle estimation model needs to be determinedsThe filtering operation can be done directly.
Converting the image into a polar coordinate system, solving the problem of scale and angle transformation of the tracking target under a Cartesian coordinate system, and converting the problem into the problem of translation and rotation of the tracking target under an estimated polar coordinate; or to be understood as: and solving the coordinate change problem of the original image and the image after translation and rotation in the polar coordinate system, wherein the problem can be solved through a phase correlation algorithm in the field of pattern recognition.
According to the theory of the phase correlation algorithm, image I1The result after feature extraction is phi (I)1) Image I1The result after feature extraction is phi (I)2) Image I2Is an image I1The impulse function matrix G between the translated and rotated images can be calculated as
Wherein k represents the kth characteristic channel,representing the frequency domain form of the variable, for the shock function matrix G after inverse Fourier transform(k)The values of elements designated by coordinates corresponding to translation and rotation of the function matrix are not 0, and the values of the other elements are 0. Obtaining translation and rotation values corresponding to the coordinates of the maximum value element by superposing the impact function matrixes G of all D characteristic channels, and further enabling the impact function matrixes G to be used for processingConverting into Cartesian coordinate system to complete the scale and angle values of the tracked target, which can be expressed as
Wherein the content of the first and second substances,for the inverse Fourier transform, max (H) is a function of the maximum value, and max (H) returns are the maximum values in the matrix H. find (H; ζ) is a search function defined as: and returning the position information of the element zeta in the matrix H, namely the row value and column value information of zeta in H. Here, the search function functions have the role of: at filter response matrixIn search of response maximumThe position information s', H and ζ in the response map are temporary variables that are set for convenient function interpretation, and have no other role in the algorithm.
Corresponding calculation formula according to related filtering algorithmAnd formulaCan obtain the product
Wherein f issThe argument of which is the image sample x, is a function of the response value. And for single corner point correlation filterHas the following calculation formula
Wherein the content of the first and second substances,is a linear combination of frames preceding the current frame. And | is to calculate the modulus of each element in the matrix.
Wherein λ iswIs a linear combinationThe learning rate parameter of (2) is,is transformed into x by polar coordinatesAnd performing a result matrix after feature extraction.
Wherein, betajIs composed ofThe previous actual coefficients, and j is 1,2, …, i.i. is equal to or greater than 1.
For the joint prediction model of four corner points, orderWherein the content of the first and second substances,is a spatial positive overlap adding operator, and hasIf with x1Is a standard angle, xi+1Then is equal to xiThe rotation angle therebetween is 90 deg., and the direction is counterclockwise. The expression of the four corner point joint prediction model is
Due to x2,x3,x4Are all x1Is obtained by rotation transformation, then x1,x2,x3,x4All the dimensions of (A) are subject to a template S ═ W, H]Wherein W and H are each x1Length and width. The image size limitation adopts an image compression method, namely, the image is reduced without directly deleting image information.
In the expression of the four-corner point joint prediction model,is gsAndcorresponding to the form of augmentation. After the maximum value optimization is carried out on the final response obtained by the calculation of the four-corner joint prediction model, the obtained augmented coordinate value is
Using the augmented four corner point coordinate values in the continuous correlation filter generation moduleResampling for filter generation to obtain w ═ wg+wh+wnWindowed resampled image samples for layer featuresAccording to the theory of continuous filtering, the overall optimization function E (f) of continuous correlation filtering is
Wherein alpha isjPenalty term corresponding to j-th correlation filterPenalty parameter of yjIs a Gaussian response map corresponding to the j-th correlation filter,being a continuous filter, fdIs a filter for the D-th layer feature, D is the total number of layers of image features,is a filter fdA corresponding weight matrix.
Discretizing each characteristic layer by continuous correlation filtering, and then the training formula with correlation filtering is as follows
Wherein, [ k ]]For the index variable after the discretization,discrete compensation coefficients corresponding to the d-th layer features,is the d-th layer characteristic.
The training formula of the correlation filtering is substituted into the overall optimization function of the continuous correlation filtering, and the pasival formula is applied to obtain the continuous correlation filtering
Wherein the content of the first and second substances,is composed ofIn the form of the enlargement of (a) or (b),is composed ofIn the form of an extension of | · |)2Is thatStandard euclidean norm below.From 2K +1 rows and eachAll have d characteristic channels and 2Kd+1 columns, each element value is respectively
Wherein A isHIs a conjugate transpose matrix of A matrix, and the expression of the matrix A isGamma is a diagonal weight matrix with the expression ofGaussian type mark matrixIs expressed as
The filter is generated by the expression
Resampling image samplesSubstituting the formula to generate a filter f to obtain a generated filter
In the scale and angle filter generation module, the formula is matchedThe four corner points are enlarged to obtain
Resampling image samplesSubstituting the above formula for the filterGenerating to obtain the generated scale and angle filter
In the ending module, the coordinate values of the four corner points after being augmented are processedGenerated filterAnd a generated scale and angle filterReserving and amplifying the coordinate values of the four corner pointsApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd the video sequence is carried into a cosine window windowing module for operation, and the operation from the cosine window windowing module to an ending module is executed in a circulating way on each next frame until the video sequence is ended.
Example two
The embodiment provides an irregular frame target tracking method adopting a polar coordinate transformation method, as shown in fig. 1. The irregular frame target tracking method adopting the polar coordinate transformation method comprises the following steps: step 1, initializing parameters in a continuous correlation filtering algorithm, and tracking a given position of a target, namely a target central point coordinate (x) according to a first framec,yc) Cutting and sampling the image in Cartesian coordinate system to obtain sample image block x1(ii) a Step 2, sample image block x1Extracting features, and then performing series connection and cosine window windowing on the extracted features to obtain an image sample z'; step 3, obtaining a characteristic spectrum Z 'by fast Fourier transform of the image sample characteristic Z', and then using continuousThe related filter h carries out filtering processing on the characteristic spectrum Z' to obtain a continuous filtering response spectrum r, and then the continuous filtering response spectrum r is calculated through a Newton method to obtain the estimated coordinates of the target center pointStep 4, by using the estimated coordinates of the target center pointGenerating four-corner coordinates of a tracking target, and sampling and preprocessing a quadrilateral region formed by the generated four-corner coordinates to obtain the coordinate values of the augmented four-corner pointsStep 5, the coordinate values of the four corner points after being augmented are processedResampling for continuous correlation filter generation to obtain resampled image samplesAnd using the resampled image samplesGenerating a filter f to obtain a generated filterStep 6, using the resampled image sampleTo filterGenerating to obtain the generated scale and angle filterStep 7, the four-corner coordinate value after being augmented is processedGenerated filterAnd a generated scale and angle filterReserving and amplifying the coordinate values of the four corner pointsApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd carrying out operation in the step 2, and circularly executing the operation from the step 2 to the step 7 for the next frames until the video sequence is ended.
In step 1, the parameters of the respective parts involved in the correlation filtering algorithm are initialized. From a given position of the tracked object in the first frame (object center point coordinates (x)c,yc) Length and width m, n) of object frame), cut out the sampling to the image under the cartesian coordinate system, the scope is four angular points: upper left corner point s1=(xc+m/2,yc+ n/2), upper right corner point s2=(xc-m/2,yc+ n/2), lower right corner point s3=(xc-m/2,yc-n/2), lower left corner point s4=(xc+m/2,yc-n/2) defining a quadrangular area. The set of coordinate points may be represented as: s ═ S1 s2 s3 s4]T. Cropped sample image block x1Will be used to train the correlation filter after superimposing the cosine window.
In step 2, the sample image block x is processed1Performing a feature extraction operation, the extracted features sequentially comprising: gray, HOG and CN; obtaining a characteristic map z according to the characteristic GraygObtaining a characteristic map z according to the characteristic HOGhObtaining a characteristic map z according to the characteristic CNnWherein wi(i∈[g,h,n]) Is a characteristic spectrum ziThe number of feature channels of (a); connecting the features Gray, the HOG and the CN in series according to the dimension of the number of the feature channels to obtain an image sampleWherein w ═ wg+wh+wn,wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is, and w is the number of synthetic characteristic layers; point-by-point two-dimensional cosine window cos for image sample zwindowThe window operation is expressed as z ═ z-windowThe as is a matrix dot product operation,and z' is a feature spectrum, cos, windowed with a cosine windowwindowIs obtained from the function form of cosine window, and the expression is
coswindow=h(m)h(n)′
Wherein m and n are respectively the length and width of the frame of the tracked target, and h (m) and h (n) are both one-dimensional cosine window generating functions h (N), and the expression is
The period of the cosine window function is 2 pi/N, with N-m or N-N. The function distribution characteristic of the one-dimensional cosine window is that h (N) is 0 when i is 0, h (N) is 1 when i is N/4, h (N) is 2 when i is N/2, h (N) is 1 when i is 3N/2, and h (N) is 0 when i is N, i shows a tendency that the cosine equation increases monotonically first and then decreases monotonically. And a two-dimensional cosine window cos obtained by the outer product of two one-dimensional cosine windowswindowThe method has the characteristic that the closer to the center, the larger the pixel value is, and the pixel values of a circle around the boundary are all 0.
In step 3, the image sample characteristics Z ' are subjected to fast Fourier transform to obtain a characteristic spectrum Z ', and then the characteristic spectrum Z ' is subjected to filtering processing by using a continuous correlation filter h to obtain a continuous filtering response spectrum r, wherein the formula is
Wherein the content of the first and second substances,representing an analog Fourier transform (iFFT), < is a matrix dot product operation, and
then, calculating the continuous filtering response spectrum r by a Newton method to obtain the estimated coordinates of the target central point
In step 4, the estimated coordinates of the center point of the target are obtainedThen, generating four-corner point coordinates, wherein the generated four-corner point coordinates are as follows: point on the upper left cornerUpper right corner pointLower right corner pointLower left corner pointThe set of coordinate points may be represented as: s ═ S'1 s′2 s′3 s′4]T. Then the generated four-corner point coordinate structure is alignedAnd sampling the formed quadrilateral area, and performing feature extraction and windowing operation to obtain a single corner feature map Z' for scale prediction. Rotating the four corner points of Z' to obtain the feature space of the four corner points after being enlargedWhereinAre orthogonal superposition operators of space, and haveFor a single corner point correlation filter gsAlso carrying out amplification to obtain an amplified scale and angle estimation filterFormula pair of augmented four-corner point feature space by applying four-corner point joint prediction modelCarrying out continuous correlation filtering, and obtaining the coordinate values of the four corner points after being amplified by applying the formula of the amplified coordinate values S
The formula of the four-corner point joint prediction model and the formula of the augmented coordinate value S' are obtained in the following mode:
let the image I be in N × N size in Cartesian coordinate system with its upper coordinateHas a pixel value ofThe origin of the coordinate system is the upper left corner of the image, the positive direction of the x axis is the positive right direction, and the positive direction of the y axis is the positive down direction. Where i is 1,2, …, N. According to the definition, the coordinates of the pixel points at the upper left corner of the image IAnd the coordinate of the pixel point at the lower right corner of the image is (1,1) and is (N, N).
The polar coordinate representation of the image I can be regarded as the coordinate in the Cartesian coordinate systemTo the polar coordinate system, coordinateNon-linear, non-uniform transformation of (a). The polar coordinate is transformed into SfThen there is
In order to keep the integrity of the angle change of the polar coordinate, namely, the positive and negative values are uniformly distributed, and the image can be uniformly divided, namely, the image is diffused outwards from the datum point, and the values of all parts are basically uniform. The central point (x) of the image I under the Cartesian coordinate systemc,yc) As a reference point for polar coordinate transformation.
Under a polar coordinate system, an arbitrary coordinate point (rho)i,θi) Middle mode length rhoiIs expressed as
Phase thetaiIs expressed as
Wherein r is a reference standard direction vector, and the positive direction of the polar coordinate system is the same as the positive direction of the Cartesian coordinate system.
Image I (x) after polar coordinate changei,yi) Is an image I' (ρ)i,θi). In a Cartesian coordinate system, the phase angle θiHas a more concise expression form as shown in the following
In addition, an image I in a Cartesian coordinate system is divided into0Proceeding with die lengthAnd phaseAnd generating a changed image I1When, the following correspondence relationship is given
Correspondingly, the antipodal coordinate mark is
As shown in the two formulas, the change of the image scale and the rotation in the Cartesian coordinate system can be converted into the translational motion along the axis in the polar coordinate system, and based on the characteristic, the target scale and the rotation transformation in the target tracking can be more simply calculated and estimated.
Through polar coordinate system transformation, a prediction model of single-angle-point scale and angle transformation of the trapezoid tracking target frame under polar coordinates can be constructed. The response calculation to the correlation filtering in polar coordinates has
Wherein the content of the first and second substances,according to the s ═ p, theta pair IiSampling polar coordinates, where s' is the coordinate of s in polar coordinate system, and convertingIs given by the formula
Wherein, H and W are respectively the image I sampled according to the tracking target frameiLength and width of (d); rho' is the module length of s after the polar coordinate system transformation; and theta' is the phase of s after the polar coordinate system transformation.
The Fourier domain calculation used by the classic DCF target tracking framework can also be used for calculation under polar coordinate transformation, and the specific response calculation formula is
Wherein the content of the first and second substances,as a function of polar transformation, gsA model weight matrix is estimated for the scale and angle. According to the formulaThe scale and angle estimation under polar coordinates does not need to be converted and sampled for calculationAccording to the formulaOnly the weight matrix g of the scale and angle estimation model needs to be determinedsThe filtering operation can be done directly.
Converting the image into a polar coordinate system, solving the problem of scale and angle transformation of the tracking target under a Cartesian coordinate system, and converting the problem into the problem of translation and rotation of the tracking target under an estimated polar coordinate; or to be understood as: and solving the coordinate change problem of the original image and the image after translation and rotation in the polar coordinate system, wherein the problem can be solved through a phase correlation algorithm in the field of pattern recognition.
According to the theory of the phase correlation algorithm, image I1The result after feature extraction is phi (I)1) Image I1The result after feature extraction is phi (I)2) Image I2Is an image I1The impulse function matrix G between the translated and rotated images can be calculated as
Wherein k represents the kth characteristic channel,representing the frequency domain form of the variable, for the shock function matrix G after inverse Fourier transform(k)The values of elements designated by coordinates corresponding to translation and rotation of the function matrix are not 0, and the values of the other elements are 0. Obtaining translation and rotation values corresponding to the coordinates of the maximum value element by superposing the impact function matrixes G of all D characteristic channels, and further enabling the impact function matrixes G to be used for processingConverting into Cartesian coordinate system to complete the scale and angle values of the tracked target, which can be expressed as
Wherein the content of the first and second substances,for the inverse Fourier transform, max (H) is a function of the maximum value, and max (H) returns are the maximum values in the matrix H. find (H; ζ) is a search function defined as: and returning the position information of the element zeta in the matrix H, namely the row value and column value information of zeta in H. Here, the search function functions have the role of: at filter response matrixIn search of response maximumThe position information s', H and ζ in the response map are temporary variables that are set for convenient function interpretation, and have no other role in the algorithm.
Corresponding calculation formula according to related filtering algorithmAnd formulaCan obtain the product
Wherein f issThe argument of which is the image sample x, is a function of the response value. And for single corner point correlation filterHas the following calculation formula
Wherein the content of the first and second substances,is a linear combination of frames preceding the current frame. And | is to calculate the modulus of each element in the matrix.
Wherein λ iswIs a linear combinationThe learning rate parameter of (2) is,is transformed into x by polar coordinatesAnd performing a result matrix after feature extraction.
Wherein, betajIs composed ofThe previous actual coefficients, and j is 1,2, …, i.i. is equal to or greater than 1.
For the joint prediction model of four corner points, orderWherein the content of the first and second substances,is a spatial positive overlap adding operator, and hasIf with x1Is a standard angle, xi+1Then is equal to xiThe rotation angle therebetween is 90 deg., and the direction is counterclockwise. The expression of the four corner point joint prediction model is
Due to x2,x3,x4Are all x1Is obtained by rotation transformation, then x1,x2,x3,x4All the dimensions of (A) are subject to a template S ═ W, H]Wherein W and H are each x1Length and width. The image size limitation adopts an image compression method, namely, the image is reduced without directly deleting image information.
In the expression of the four-corner point joint prediction model,is gsAndcorresponding to the form of augmentation. After the maximum value optimization is carried out on the final response obtained by the calculation of the four-corner joint prediction model, the obtained augmented coordinate value is
In step 5, the augmented four corner point coordinate values are usedResampling for filter generation is performed to obtain a signal with w ═ wg+wh+wnWindowed resampled image samples for layer featuresAccording to the theory of continuous filtering, the overall optimization function E (f) of continuous correlation filtering is
Wherein alpha isjPenalty term corresponding to j-th correlation filterPenalty parameter of yjIs a Gaussian response map corresponding to the j-th correlation filter,being a continuous filter, fdIs a filter for the D-th layer feature, D is the total number of layers of image features,is a filter fdA corresponding weight matrix.
Discretizing each characteristic layer by continuous correlation filtering, and then the training formula with correlation filtering is as follows
Wherein, [ k ]]For the index variable after the discretization,discrete compensation coefficients corresponding to the d-th layer features,is the d-th layer characteristic.
The training formula of the correlation filtering is substituted into the overall optimization function of the continuous correlation filtering, and the pasival formula is applied to obtain the continuous correlation filtering
Wherein the content of the first and second substances,is composed ofIn the form of the enlargement of (a) or (b),is composed ofIn the form of an extension of | · |)2Is thatStandard euclidean norm below.From 2K +1 rows and eachAll have d characteristic channels and 2Kd+1 columns, each element value is respectively
Wherein A isHIs a conjugate transpose matrix of A matrix, and the expression of the matrix A isGamma is a diagonal weight matrix with the expression ofGaussian type mark matrixIs expressed as
The filter is generated by the expression
Resampling image samplesSubstituting the formula to generate a filter f to obtain a generated filter
Resampling image samplesSubstituting the above formula for the filterGenerating to obtain the generated scale and angle filter
In step 7, the augmented four corner point coordinate values are calculatedGenerated filterAnd generated scale and angle filterReserving and amplifying the coordinate values of the four corner pointsApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd carrying out operation in the step 2, and circularly executing the operation from the step 2 to the step 7 for the next frames until the video sequence is ended.
The method and the way for implementing the technical scheme are many, and the above is only the preferred embodiment of the invention. It should be noted that, for those skilled in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be construed as the protection scope of the present invention. All the components not specified in the present embodiment can be realized by the prior art.
Claims (10)
1. An irregular-frame target tracking system using a polar coordinate transformation method, comprising:
an initial image sampling module for initializing parameters in the continuous correlation filtering algorithm and tracking the given position of the target, namely the target center point coordinate (x) in the first framec,yc) Cutting and sampling the image in Cartesian coordinate system to obtain sample image block x1;
A cosine window windowing module for windowing the sample image block x1Extracting features, and then performing series connection and cosine window windowing on the extracted features to obtain an image sample z';
the continuous correlation filtering module is used for performing fast Fourier transform on the image sample characteristics Z ' to obtain a characteristic spectrum Z ', then performing filtering processing on the characteristic spectrum Z ' by using a continuous correlation filter h to obtain a continuous filtering response spectrum r, and then calculating the continuous filtering response spectrum r by a Newton method to obtain an estimated target central point coordinate
An irregular border estimation module to estimate the coordinates of the target center point by using the estimated coordinatesGenerating four-corner point coordinates of the tracking target, and sampling and preprocessing a quadrilateral region formed by the generated four-corner point coordinates to obtain the augmented four-corner point coordinate values
A continuous correlation filter generation module for generating the augmented four corner point coordinate valuesResampling for continuous correlation filter generation to obtain resampled image samplesAnd using the resampled image samplesGenerating a filter f to obtain a generated filter
A scale and angle filter generation module to generate a scale and angle filter by using the resampled image samplesTo filterGenerating to obtain the generated scale and angle filter
A finishing module for processing the augmented four corner point coordinate valuesThe generated filterAnd the generated scale and angle filterReserving and amplifying the coordinate values of the four corner pointsApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd the video sequence is carried into a cosine window windowing module for operation, and the operation from the cosine window windowing module to an ending module is executed in a circulating way on each next frame until the video sequence is ended.
2. The irregular-frame target tracking system using the polar coordinate transformation method according to claim 1, characterized in that: in the irregular border estimation module, the preprocessing is: carrying out feature extraction and windowing on the sampled quadrilateral area to obtain a single angular point feature map Z' for scale prediction; carrying out four-corner point rotation on the single corner point feature map Z' to obtain an augmented four-corner point feature spaceFor the feature space of the four corner points after the enlargementCarrying out continuous correlation filtering, and obtaining the coordinate values of the four corner points after the amplification through the coordinate values S' after the amplification
3. The irregular-frame target tracking system using the polar coordinate transformation method according to claim 2, characterized in that: in the continuous correlation filter generation module, the correlation filter is generated,
using the augmented four corner point coordinate valuesResampling for successive correlation filter generation to obtain a correlation signal having w ═ wg+wh+wnWindowed resampled image samples for layer featuresWherein wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is shown, and w is the number of synthetic characteristic layers;
the generation expression of the filter f obtained according to the continuous correlation filtering theory is
Wherein, WHIs a conjugate transpose of the weight matrix W of the filter f, AHIs a conjugate transpose matrix of A matrix, and the expression of the matrix A isGamma is a diagonal weight matrix with the expression ofGaussian type mark matrixIs expressed as
4. The irregular-frame target tracking system using the polar coordinate transformation method according to claim 3, characterized in that: in the scale and angle filter generation module,
according to the continuous phaseSingle corner filter obtained by filter-related theoryIs expressed as
Therein, <' >-1Is a matrix dot product operation, which is the inverse operation of the matrix dot product operation, k represents the k-th feature channel,is a linear combination of the frames preceding the current frame,is composed ofA result matrix after Fourier transform, andis composed ofThe conjugate matrix of (a) is determined,is the result of x after polar transformation, andis composed ofThe result matrix after feature extraction, | · | is that each element in the matrix is respectively solved for its modulus;
four corner points are enlarged to obtain
5. The irregular-frame target tracking system using the polar coordinate transformation method according to claim 4, characterized in that: in the cosine window windowing module, the cosine window is set to zero,
for sample image block x1Performing a feature extraction operation, the extracted features sequentially comprising: gray, HOG and CN;
obtaining a characteristic map z according to the characteristic GraygObtaining a characteristic map z according to the characteristic HOGhObtaining a characteristic map z according to the characteristic CNnWherein z isg,zh,wi(i∈[g,h,n]) Is a characteristic spectrum ziThe number of feature channels of (a);
connecting the features Gray, the HOG and the CN in series according to the dimension of the number of the feature channels to obtain an image sampleWherein w ═ wg+wh+wn,wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is shown, and w is the number of synthetic characteristic layers;
point-by-point two-dimensional cosine window cos for image sample zwindowThe window operation is expressed as z ═ z-windowThe as is a matrix dot product operation,and z' is a feature spectrum, cos, windowed with a cosine windowwindowIs obtained from the function form of cosine window, and the expression is
coswindow=h(m)h(n)′
Wherein m and n are respectively the length and width of the frame of the tracked target, and h (m) and h (n) are both one-dimensional cosine window generating functions h (N), and the expression is
The period of the cosine window function is 2 pi/N, with N-m or N-N.
6. An irregular frame target tracking method adopting a polar coordinate transformation method comprises the following steps:
step 1, initializing parameters in a continuous correlation filtering algorithm, and tracking a given position of a target, namely a target central point coordinate (x) according to a first framec,yc) Cutting and sampling the image in Cartesian coordinate system to obtain sample image block x1;
Step 2, for the sample image block x1Extracting features, and then performing series connection and cosine window windowing on the extracted features to obtain an image sample z';
step 3, obtaining a characteristic spectrum Z 'by fast Fourier transform of the image sample characteristic Z', and then using continuous correlation filteringThe wave filter h carries out filtering processing on the characteristic spectrum Z' to obtain a continuous filtering response spectrum r, and then the continuous filtering response spectrum r is calculated through a Newton method to obtain an estimated target central point coordinate
Step 4, by using the estimated target center point coordinatesGenerating four-corner point coordinates of the tracking target, and sampling and preprocessing a quadrilateral region formed by the generated four-corner point coordinates to obtain the augmented four-corner point coordinate values
Step 5, the coordinate values of the four corner points after the amplification are processedResampling for continuous correlation filter generation to obtain resampled image samplesAnd using the resampled image samplesGenerating a filter f to obtain a generated filter
Step 6, using the resampled image sampleTo filterGenerating to obtain the generated scale and angle filter
Step 7, the coordinate values of the four corner points after the amplification are processedThe generated filterAnd the generated scale and angle filterReserving and amplifying the coordinate values of the four corner pointsApplied to the image of the next frame to obtain a new image block x of the cropped sampleiAnd carrying out operation in the step 2, and circularly executing the operation from the step 2 to the step 7 for the next frames until the video sequence is ended.
7. The irregular-frame target tracking method using the polar coordinate transformation method according to claim 6, characterized in that: in step 4, the pretreatment is as follows: carrying out feature extraction and windowing on the sampled quadrilateral area to obtain a single angular point feature map Z' for scale prediction; performing four-corner point rotation on the single corner point feature map Z' to obtain an augmented four-corner point feature space Z; for the feature space of the four corner points after the enlargementCarrying out continuous correlation filtering, and obtaining the coordinate values of the four corner points after the amplification through the coordinate values S' after the amplification
8. The irregular-frame target tracking method using the polar coordinate transformation method according to claim 7, characterized in that: in the step 5, the process is carried out,
using the augmented four corner point coordinate valuesResampling for successive correlation filter generation to obtain a correlation signal having w ═ wg+wh+wnWindowed resampled image samples for layer featuresWherein wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is shown, and w is the number of synthetic characteristic layers;
the generation expression of the filter f obtained according to the continuous correlation filtering theory is
Wherein, WHIs a conjugate transpose of the weight matrix W of the filter f, AHIs a conjugate transpose matrix of A matrix, and the expression of the matrix A isGamma is a diagonal weight matrix with the expression ofGaussian type mark matrixIs expressed as
9. The irregular-frame target tracking method using the polar coordinate transformation method according to claim 8, characterized in that: in a step 6, the process is carried out,
obtaining a single corner filter according to the continuous correlation filtering theoryIs expressed as
Therein, <' >-1Is a matrix dot product operation, which is the inverse operation of the matrix dot product operation, k represents the k-th feature channel,is a linear combination of the frames preceding the current frame,is composed ofA result matrix after Fourier transform, andis composed ofThe conjugate matrix of (a) is determined,is the result of x after polar transformation, andis composed ofThe result matrix after feature extraction, | · | is that each element in the matrix is respectively solved for its modulus;
four corner points are enlarged to obtain
10. The irregular-frame target tracking method using the polar coordinate transformation method according to claim 9, characterized in that: in the step 2, the process is carried out,
for sample image block x1Performing a feature extraction operation, the extracted features sequentially comprising: gray, HOG and CN;
obtaining a characteristic map z according to the characteristic GraygObtaining a characteristic map z according to the characteristic HOGhObtaining a characteristic map z according to the characteristic CNnWherein z isg,zh,wi(i∈[g,h,n]) Is a characteristic spectrum ziThe number of feature channels of (a);
connecting the features Gray, the HOG and the CN in series according to the dimension of the number of the feature channels to obtain an image sampleWherein w ═ wg+wh+wn,wgFor the number of Gray feature layers, whNumber of layers for HOG feature, wnThe number of CN characteristic layers is shown, and w is the number of synthetic characteristic layers;
point-by-point two-dimensional cosine window cos for image sample zwindowThe window operation is expressed as z ═ z-windowThe as is a matrix dot product operation,and z' is a feature spectrum, cos, windowed with a cosine windowwindowIs obtained from the function form of cosine window, and the expression is
coswindow=h(m)h(n)′
Wherein m and n are respectively the length and width of the frame of the tracked target, and h (m) and h (n) are both one-dimensional cosine window generating functions h (N), and the expression is
The period of the cosine window function is 2 pi/N, with N-m or N-N.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111422525.0A CN114119666A (en) | 2021-11-26 | 2021-11-26 | Irregular frame target tracking system and method adopting polar coordinate transformation method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111422525.0A CN114119666A (en) | 2021-11-26 | 2021-11-26 | Irregular frame target tracking system and method adopting polar coordinate transformation method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114119666A true CN114119666A (en) | 2022-03-01 |
Family
ID=80370204
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111422525.0A Pending CN114119666A (en) | 2021-11-26 | 2021-11-26 | Irregular frame target tracking system and method adopting polar coordinate transformation method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114119666A (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102982556A (en) * | 2012-11-01 | 2013-03-20 | 江苏科技大学 | Video target tracking method based on manifold particle filter algorithm |
CN111738045A (en) * | 2020-01-19 | 2020-10-02 | 中国科学院上海微系统与信息技术研究所 | Image detection method and device, electronic equipment and storage medium |
US20210227132A1 (en) * | 2018-05-30 | 2021-07-22 | Arashi Vision Inc. | Method for tracking target in panoramic video, and panoramic camera |
-
2021
- 2021-11-26 CN CN202111422525.0A patent/CN114119666A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102982556A (en) * | 2012-11-01 | 2013-03-20 | 江苏科技大学 | Video target tracking method based on manifold particle filter algorithm |
US20210227132A1 (en) * | 2018-05-30 | 2021-07-22 | Arashi Vision Inc. | Method for tracking target in panoramic video, and panoramic camera |
CN111738045A (en) * | 2020-01-19 | 2020-10-02 | 中国科学院上海微系统与信息技术研究所 | Image detection method and device, electronic equipment and storage medium |
Non-Patent Citations (2)
Title |
---|
XUEDONG WU, ET AL.: "Scale-Adaptive Kernel Correlation Filter with Maximum Posterior Probability Estimation and Combined Features for Visual Target Tracking", RESEARCHGATE, 30 June 2021 (2021-06-30) * |
伍雪冬等: "基于上下文感知和尺度自适应的实时目标跟踪", 红外技术, vol. 43, no. 05, 31 May 2021 (2021-05-31), pages 429 - 436 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111862126B (en) | Non-cooperative target relative pose estimation method combining deep learning and geometric algorithm | |
CN111028292B (en) | Sub-pixel level image matching navigation positioning method | |
CN107516322B (en) | Image object size and rotation estimation calculation method based on log polar space | |
CN106447601B (en) | Unmanned aerial vehicle remote sensing image splicing method based on projection-similarity transformation | |
Wu et al. | Autonomous seam recognition and feature extraction for multi-pass welding based on laser stripe edge guidance network | |
Henson et al. | Attitude-trajectory estimation for forward-looking multibeam sonar based on acoustic image registration | |
Crowley et al. | Continuity properties of the appearance manifold for mobile robot position estimation | |
CN112183675B (en) | Tracking method for low-resolution target based on twin network | |
CN113516693B (en) | Rapid and universal image registration method | |
Xian et al. | Super-resolved fine-scale sea ice motion tracking | |
CN101556694A (en) | Matching method of rotating images | |
CN116907509A (en) | AUV underwater auxiliary navigation method, system, equipment and medium based on image matching | |
Liu et al. | Phase correlation pixel‐to‐pixel image co‐registration based on optical flow and median shift propagation | |
Tian et al. | Real time augmented reality tracking registration based on motion blur template matching image construction model | |
CN113837924A (en) | Water bank line detection method based on unmanned ship sensing system | |
CN112184785B (en) | Multi-mode remote sensing image registration method based on MCD measurement and VTM | |
CN113763274A (en) | Multi-source image matching method combining local phase sharpness orientation description | |
CN114066795A (en) | DF-SAS high-low frequency sonar image fine registration fusion method | |
Jindal et al. | An ensemble mosaicing and ridgelet based fusion technique for underwater panoramic image reconstruction and its refinement | |
CN114119666A (en) | Irregular frame target tracking system and method adopting polar coordinate transformation method | |
CN109886988A (en) | A kind of measure, system, device and the medium of Microwave Imager position error | |
CN116109682A (en) | Image registration method based on image diffusion characteristics | |
CN111899284B (en) | Planar target tracking method based on parameterized ESM network | |
Wei et al. | Beam-domain image mosaic of forward-looking sonar using expression domain mapping model | |
Kostrykin et al. | Globally Optimal and Scalable Video Image Stitching for Robotic Visual Inspection of Electric Generators |
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 |