CN110189361A - A kind of multi-channel feature and the target following preferentially updated parallel - Google Patents

A kind of multi-channel feature and the target following preferentially updated parallel Download PDF

Info

Publication number
CN110189361A
CN110189361A CN201910451153.0A CN201910451153A CN110189361A CN 110189361 A CN110189361 A CN 110189361A CN 201910451153 A CN201910451153 A CN 201910451153A CN 110189361 A CN110189361 A CN 110189361A
Authority
CN
China
Prior art keywords
branch
feature
target
lower branch
frame
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
CN201910451153.0A
Other languages
Chinese (zh)
Inventor
胡昭华
李高飞
陈胡欣
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.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
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 Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN201910451153.0A priority Critical patent/CN110189361A/en
Publication of CN110189361A publication Critical patent/CN110189361A/en
Pending legal-status Critical Current

Links

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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • 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
    • 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
    • G06T2207/20084—Artificial neural networks [ANN]

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

The present invention is on the basis of traditional core correlation filtering, a kind of provide multi-channel feature and preferentially update parallel target following；The network structure for being suitable for target following is obtained using pre-training mode first, it is proposed new feature extraction mode later: upper branch extracts depth characteristic using trained network, lower branch then extracts traditional HOG and CN feature and constitutes manual feature set, and realizing more robust characteristic look compared to traditional single features indicates.The present invention proposes novel update mode simultaneously: preferentially parallel to update, the parallel tracking frame of two-way has obtained two different responses in present frame, optimal target position is found according to biggish response, update the Filtering Template and feature templates of two-way filter simultaneously using the corresponding parameter in optimum position in next frame, always use this strategy is preferentially parallel to update in subsequent tracking process, until tracking terminates.

Description

A kind of multi-channel feature and the target following preferentially updated parallel
Technical field
The invention belongs to image procossings and computer vision field, extract target signature by much channel communication, are updating Stage realizes that the target under various scenes is accurately tracked using the strategy preferentially updated parallel.It can be applied to unmanned, intelligence It can the fields such as security protection and video monitoring.
Background technique
Computer vision has numerous branches, and target following is one of the research direction on basis.Target following is given After the target position of first frame and the size of bounding box, specific objective is tracked in subsequent video sequence, is widely used in The fields such as intelligent security guard, unmanned plane detecting, missile guidance positioning.Target tracking algorism had obtained substantive quick hair in recent years Exhibition, mainly there is the tracking based on correlation filtering and the track algorithm based on deep learning.
Track algorithm based on correlation filtering class since high speed and higher precision are widely used, Middle Henriques et al. (Henriques J F, Caseiro R, Martins P, et al.High-speed tracking with kernelized correlation filters[J].IEEE Transactions on Pattern Analysis And Machine Intelligence, 2015,37 (3): 583-596.) propose core correlation filter high speed track algorithm (KCF) there is good performance in terms of speed and precision.KCF solves tracking phase sample deficiency by cyclic shift method The problem of, pass through the method training objective template of ridge regression later, the inseparable situation of linear is reflected using kernel function It is mapped to non-linear space and solves dual problem.Meanwhile KCF algorithm use direction histogram of gradients (HOG) feature replacement is traditional Gray feature achieves certain tracking effect and is promoted.But KCF algorithm can not cope with difference due to using single feature Tracking scene, and the update mode of KCF algorithm is to obtain position basis corresponding to each frame response maximum value enterprising Row updates, once target drift, dislocation, the mode of such update can be accumulated infinitely until target is lost completely.
With the progress of various vision contests, various network structures are introduced in tracking field.Nam et al. (Nam H, Han B.Learning multi-domain convolutional neural networks for visual tracking [C]//Computer Vision and Pattern Recognition(CVPR),2016IEEE Conference On.IEEE, 2016:4293-4302.) MDNet track algorithm is proposed on this basis.MDNet algorithm is appointed for image classification Existing difference between business and tracking proposes that one is directly obtained with tracking video preprocessor training network is suitable for tracking field Network structure.MDNet is directed to the difference of training sequence, proposes multiple domain training, and to each Sequence composition, its two classification connect entirely Layer is connect, for distinguishing the foreground and background of current sequence, it should be noted that the all-network layer before full articulamentum is all altogether It enjoys.Tracking phase with first frame video its bounding box regression model of training and extracts positive negative sample and updates net Network weight.The highest candidate samples of confidence level are chosen, bounding box recurrence is carried out later and acquires final result.MDNet with Track precision is high, but there are some problems, speed is too slow, and online updating also takes long time.
Summary of the invention
The target tracking algorism that the present invention proposes a kind of multi-channel feature and preferentially updates parallel, promotes the same of tracking accuracy When also reached ideal tracking performance.
The specific technical solution of the present invention is as follows:
A kind of multi-channel feature and the method for tracking target preferentially updated parallel, comprising the following steps:
Step 1: pre-training network
Initialization using the weight parameter of VGG-M network as training network, by using VOT video sequence to the mind Pre-training is carried out through network, determines parameter；The parameter includes learning rate 0.0001, the learning rate 0.001 of full articulamentum, Momentum coefficient 0.9 and weight attenuation coefficient 0.002；
Step 2: it constructs branch up and down and generates target signature collection
Upper and lower branch includes upper branch depth characteristic tracking branch and lower branch craft signature tracking branch；Upper and lower branch root Basic image pattern is obtained into sample set X by cyclic shift according to target position known to first frame；For upper branch by sample This collection X is input in the good neural network of above-mentioned pre-training, and the CNNs feature for extracting the last one convolutional layer constitutes depth characteristic Sample set X is similarly input to lower branch and extracts the manual feature set of CN and HOG feature composition respectively by collection；
Step 3: training stage
After obtaining the feature of step 2, upper and lower branch pass through respectively classifier f (X)=Xw find respectively it is the smallest Weight coefficient w, so that the sample of each branch and its recurrence the smallest square error of label, the minimum cost of training regression function Function are as follows:
Wherein w is also referred to as target template, and λ is punishment term coefficient, and y is recurrence label corresponding with X；When training data is line When property, target template expression formula are as follows:
Wherein, x is the feature templates of upper and lower branch, and F is discrete fourier calculating, ()*Indicate that conjugate matrices, ☉ indicate Dot product calculates, and most of sample characteristics collection is linearly inseparable, can solve feature set linearly inseparable using geo-nuclear tracin4 Problem, inversion operation are influenced algorithm speed, are obtained using circular matrix diagonalization and Fourier transformation:
Wherein, what K was indicated is the calculation of assessing between two functions, and inversion operation influences algorithm speed KxXIt is nuclear matrix K institute There is the first row of element, so far trains tracker to go to the optimal α of searching from optimal w is found, acquire training on this basis The target template w in stage；
Step 4: positioning stage
Since the 2nd frame, when image to be detected of input is m, according to obtaining candidate sample after detection range cyclic shift This collection matrix M；Candidate samples collection matrix M is sent to upper and lower two branches respectively and carries out feature templates extraction by step 2 method, Under the premise of finding out target template by step 3, the sample set to be detected of input is calculated, upper and lower branch is respectively obtained and waits for The response matrix of detection image are as follows:
F=F-1((F(KXM)☉F(α)) (4)
Taking position corresponding to f maximum value is the target position tracked；
Step 5: more new stage
After the input of present frame picture, the feature templates x of upper and lower branch present frame is acquired according to first four stepdeepWith xhand, and upper and lower branch Filtering Template wdeep, whand, after the input of next frame picture, acquire the best sound of upper and lower branch Answer fdeepAnd fhand,
Take the corresponding coordinate of the maximum response of the two as present frame target optimum position, in more two branches of new stage More new capital of next frame is updated using following optimal branch parameters, formula are as follows:
wdt=(1- ξ) wt-1+ξwdeept (5)
xdt=(1- ξ) xt-1+ξxdeept (6)
wht=(1- ξ) wt-1+ξwhandt (7)
xht=(1- ξ) xt-1+ξxhandt (8)
ξ is the model learning factor, w in formulat-1It is the previous frame optimum position target that a branch relatively obtains later above and below Template, wdeept, whandtRespectively indicate the target template of the upper and lower branch of present frame acquisition, wdt, whtIt respectively indicates for detecting The target template of upper and lower branch subsequent frame, xt-1For upper and lower branch previous frame relatively after obtained best features template, xdeept, xhandtIt is then the feature templates for the upper and lower branch that present frame obtains, xdt, xhtIt indicates for detecting branch subsequent frame up and down Feature templates；Step 2~step 5 is repeated, each frame Filtering Template and feature templates are updated, until terminating.
Wherein, in step 5, ξ value takes 0.02；
Wherein, in step 5, the corresponding depth characteristic collection extracted of feature templates in upper branch；Lower branch feature templates are corresponding The manual feature set of extraction.
The present invention has the following beneficial effects: compared with prior art
The present invention on the basis of traditional core correlation filtering first using pre-training mode obtain one be suitable for target with The network structure of track proposes new feature extraction mode later: upper branch extracts depth characteristic, lower branch using trained network Road then extracts traditional HOG and CN feature and constitutes manual feature set, realizes more robust spy compared to traditional single features Levying appearance indicates.The present invention proposes novel update mode simultaneously: preferentially parallel to update, the parallel tracking frame of two-way is current Frame has obtained two different responses, finds optimal target position according to biggish response, in next frame using most preferably The corresponding parameter in position updates the Filtering Template and feature templates of two-way filter simultaneously, always uses in subsequent tracking process This strategy is preferentially parallel to be updated, until tracking terminates.
The network that present invention training obtains has good effect to target following；The extraction of multi-channel feature can be coped with not Same challenge factor；The strategy preferentially updated parallel makes full use of the advantage of multiple branch circuit, has obtained the tracking effect of robust.It takes not Maximum response with branch present frame is as optimum target position, and the template renewal of two branches uses former frame in next frame The parameter of optimal location is updated simultaneously, and until tracking terminates, preferentially parallel update of multiple branch circuit compensates for single branch more New deficiency.
Track algorithm of the invention rotates in target, target deformation, blocks, quickly movement, homologue interference, low resolution When can still carry out the tracking of robust.
Detailed description of the invention
Fig. 1 is multi-channel feature proposed by the present invention and the method for tracking target general frame preferentially updated parallel；
Fig. 2 is convolutional neural networks structural schematic diagram proposed by the present invention；
Fig. 3 is performance comparison visualization block diagram of the present invention with upper and lower branch；
Fig. 4 is performance comparison OPE curve graph of the present invention with upper and lower branch；
Fig. 5 is that the present invention visualizes block diagram to 6 not homotactic tracking effects；
Fig. 6 is present invention figure compared with the synthesis tracking performance of 5 kinds of trackers under OPE assessment mode；
Specific embodiment
In order to more visually explain the purpose of the present invention, process and advantage, explained in conjunction with the drawings and specific embodiments The bright present invention.
Embodiment one:
The general frame of multi-channel feature of the present invention and the method for tracking target preferentially updated parallel is as shown in Figure 1, specific Including following operating procedure:
(1) step 1: pre-training network.
The network specific structure that the present invention uses is as shown in Fig. 2, mainly using the weight parameter of VGG-M network as training The initialization of network carries out pre-training to the neural network by using VOT video sequence, and the network is mainly by five convolution sums Two full articulamentum compositions, for the specification of network, the convolution kernel size of convolutional layer selected by this paper is unified for 3*3, and pond layer is adopted With maximum pond, the size of convolution kernel is unified for 2*2, and network carries out nonlinear processing using Relu function.The key of network Hyper parameter is ultimately determined to by experiment: learning rate 0.0001, the learning rate 0.001 of full articulamentum, momentum coefficient 0.9, And weight attenuation coefficient 0.002.The purpose of training network is to construct the convolutional neural networks for being suitable for tracking, therefore FC7 layers use Softmax cross entropy loss function to obtain output as the probability value of target or background, and the training stage carries out convolutional layer With the update of full articulamentum network weight, the weight parameter for saving convolutional layer after iteration each time is used for next iteration, and complete Connect network layer weight parameter in following iteration random initializtion followed by train, after network convergence, convolutional Neural net Network is saved for tracking phase and generates the depth characteristic for being suitble to tracking, and subsequent full articulamentum is given up after training terminates It goes, only with the last one convolutional layer, that is, conv5 output feature for tracking.
(2) it step 2: constructs branch up and down and generates target signature collection.
The composition of upper and lower branch as shown in Figure 1, be divided into branch depth characteristic tracking branch and lower branch craft feature with Track branch.Basic image pattern is obtained sample set X by cyclic shift by the target position according to known to first frame.By sample Collection X is input to the good convolutional neural networks of branch pre-training, and the CNNs feature of the last one convolutional layer is taken to constitute depth characteristic The step for collection, i.e. Fig. 1 carries out branch training stage extraction CNNs feature；Sample set X is similarly input to lower branch point CN and HOG feature is indescribably taken to constitute manual feature set.That is Fig. 1 carries out the step for lower branch training stage extraction craft feature.
(3) step 3: training stage.
After obtaining the feature of step 2, two branches find respective the smallest weight system by classifier f (X)=Xw Number w, so that sample and its recurrence label have the smallest square error, the least cost function of training regression function are as follows:
Wherein w is also referred to as target template, and λ is punishment term coefficient, and y is recurrence label corresponding with X.When training data is line When property, target template expression formula are as follows:
Wherein, x is the feature templates of upper and lower branch, and F is discrete fourier calculating, ()*Indicate that conjugate matrices, ☉ indicate Dot product calculates, and most of sample characteristics collection is linearly inseparable, can solve feature set linearly inseparable using geo-nuclear tracin4 Problem, inversion operation are influenced algorithm speed, are obtained using circular matrix diagonalization and Fourier transformation:
Wherein what K was indicated is the calculation of assessing between two functions, and inversion operation influences algorithm speed KxXIt is nuclear matrix K all The first row of element so far trains tracker to go to the optimal α of searching from optimal w is found, and acquires Fig. 1 training on this basis The target template in stage
(4) step 4: positioning stage.
If Fig. 1 is since the 2nd frame, when the picture to be detected of input is m, according to being waited after detection range cyclic shift Select sample set matrix M.Candidate samples collection matrix M is sent to two branches respectively and carries out feature extraction, it should be noted that two The feature extraction of branch positioning stage is consistent with step 2, therefore does not do repeated explanation herein, has found out optimal mesh in upper step Under the premise of marking template, the sample set to be detected of input is calculated, the response of upper and lower branch image to be detected is respectively obtained Matrix are as follows:
F=F-1((F(KXM)☉F(α)) (4)
Position corresponding to f maximum value is the target position of tracking.
(5) step 5: more new stage.After the input of present frame picture, it is current that each branch is acquired according to first four step The target signature template x of framedeepAnd xhandAnd Filtering Template wdeep, whand, after the input of next frame picture, acquire each The best response f on roaddeepAnd fhand,
Take the corresponding coordinate of the maximum response of the two as present frame target optimum position, in more two branches of new stage More new capital of next frame is updated using optimal branch parameters, formula are as follows:
wdt=(1- ξ) wt-1+ξwdeept (5)
xdt=(1- ξ) xt-1+ξxdeept (6)
wht=(1- ξ) wt-1+ξwhandt (7)
xht=(1- ξ) xt-1+ξxhandt (8)
ξ is the model learning factor in formula, and being worth is 0.02, wt-1It is the optimum bit that two branches of previous frame relatively obtain later Set target template, wdeept, whandtRespectively indicate the target template of the branch up and down of present frame acquisition, wdt, whtIt respectively indicates and is used for Detect the target template of branch subsequent frame up and down, xt-1For two branches of previous frame relatively after obtained best features template, xdeept, xhandtIt is then the feature templates for the branch up and down that present frame obtains, xdt, xhtIt indicates for detecting branch subsequent frame up and down Feature templates, repeat step 2 to step 5, later each frame target template and feature templates more line mode be same as above, directly To end.
Evaluation contents:
The present invention measures performance of the invention, OPE (One-pass Evaluation) tracking by OPE evaluation criteria Device disposably assesses video sequence, including accuracy figure and success rate figure, randomly chooses 70 in total from OTB2015 containing not Video sequence with the challenge factor carries out part and whole performance test.And with other trackers (FDSST, KCF, CSK, 5 kinds of trackers such as CNT, SCM) in different challenge factors, (plane external rotation, background clutter, illumination variation, motion blur are hidden Situations such as gear) under compare.
Fig. 3 is performance comparison visualization block diagram of the present invention with upper and lower branch, preferentially parallel as can be seen from these figures The validity of more new strategy, even if good tracking effect still may be implemented in branch tracking failure, the present invention up and down.
Further Fig. 4 is performance comparison OPE curve graph of the present invention with upper and lower branch, can be with from the curve values in Fig. 4 Find out that effect of the invention compares algorithm still better than it.
Fig. 5 is the present invention and 5 kinds of trackers in (a) Bolt 2, (b) Coke, (c) Couple, (d) Panda, (e) Shaking, (f) Soccer6 test video tracking result visualize block diagram.
Fig. 6 then gives in terms of accuracy curve and success rate curve two of the invention whole with other 5 kinds of trackers OPE curve graph under body performance and the different challenge factors.
Either Fig. 5 or Fig. 6, the present invention realize effective tracking effect, it can be seen that, mesh proposed by the present invention Tracking is marked under self comparison and comparison with existing algorithm, the performance of tracking is promoted, and effect is also all well and good.

Claims (3)

1. a kind of multi-channel feature and the method for tracking target preferentially updated parallel, comprising the following steps:
Step 1: pre-training network
Initialization using the weight parameter of VGG-M network as training network, by using VOT video sequence to the nerve net Network carries out pre-training, determines parameter；The parameter includes learning rate 0.0001, the learning rate 0.001, momentum of full articulamentum Coefficient 0.9 and weight attenuation coefficient 0.002；
Step 2: it constructs branch up and down and generates target signature collection
Upper and lower branch includes upper branch depth characteristic tracking branch and lower branch craft signature tracking branch；Upper and lower branch is according to the Basic image pattern is obtained sample set X by cyclic shift by target position known to one frame；For upper branch by sample set X It is input in the good neural network of above-mentioned pre-training, the CNNs feature for extracting the last one convolutional layer constitutes depth characteristic collection, together The lower branch that is input to sample set X of sample extracts the manual feature set of CN and HOG feature composition respectively；
Step 3: training stage
After obtaining the feature of step 2, upper and lower branch passes through classifier f (X)=Xw respectively and finds respective the smallest weight Coefficient w, so that the sample of each branch and its recurrence the smallest square error of label, the least cost function of training regression function Are as follows:
Wherein w is also referred to as target template, and λ is punishment term coefficient, and y is recurrence label corresponding with X；When training data is linear When, target template expression formula are as follows:
Wherein, x is the feature templates of upper and lower branch, and F is discrete fourier calculating, ()*Indicate that conjugate matrices, ☉ indicate dot product The problem of calculating, most of sample characteristics collection is linearly inseparable, can solve feature set linearly inseparable using geo-nuclear tracin4, Inversion operation influences algorithm speed, is obtained using circular matrix diagonalization and Fourier transformation:
Wherein, what K was indicated is to assess calculation, K between two functionsxXThe first row of nuclear matrix K all elements, so far train with Track device goes to the optimal α of searching from optimal w is found, and acquires the target template w of training stage on this basis；
Step 4: positioning stage
Since the 2nd frame, when image to be detected of input is m, according to obtaining candidate samples collection after detection range cyclic shift Matrix M；Candidate samples collection matrix M is sent to two branches up and down respectively and presses step by the progress feature templates extraction of step 2 method Under the premise of rapid three find out optimal objective template, the sample set to be detected of input is calculated, upper and lower branch is respectively obtained and waits for The response matrix of detection image are as follows:
F=F-1((F(KXM)☉F(α)) (4)
Taking position corresponding to f maximum value is the target position of optimal tracking；
Step 5: more new stage
After the input of present frame picture, the feature templates x of upper and lower branch present frame is acquired according to first four stepdeepAnd xhand, And upper and lower branch Filtering Template wdeep, whand, after the input of next frame picture, acquire the best response of upper and lower branch fdeepAnd fhand, take the corresponding coordinate of the maximum response of the two as present frame target optimum position, in more two branch of new stage More new capital of road next frame is updated using optimal branch parameters, formula are as follows:
wdt=(1- ξ) wt-1+ξwdeept (5)
xdt=(1- ξ) xt-1+ξxdeept (6)
wht=(1- ξ) wt-1+ξwhandt (7)
xht=(1- ξ) xt-1+ξxhandt (8)
ξ is the model learning factor, w in formulat-1It is the previous frame optimum position target template that a branch relatively obtains later above and below, wdeept, whandtRespectively indicate the target template of the upper and lower branch of present frame acquisition, wdt, whtIt respectively indicates upper and lower for detecting The target template of branch subsequent frame, xt-1For the best features template that upper and lower branch previous frame relatively obtains later, xdeept, xhandtIt is then the feature templates for the upper and lower branch that present frame obtains, xdt, xhtIndicate the spy for detecting upper and lower branch subsequent frame Levy template；Step 2~step 5 is repeated, each frame Filtering Template and feature templates are updated, until terminating.
2. multi-channel feature and the method for tracking target preferentially updated parallel according to claim 1, wherein in step 5, ξ Value takes 0.02.
3. multi-channel feature according to claim 1 or claim 2 and the method for tracking target preferentially updated parallel, wherein step 5 In, the corresponding depth characteristic collection extracted of feature templates in upper branch；The corresponding manual feature set extracted of lower branch feature templates.
CN201910451153.0A 2019-05-28 2019-05-28 A kind of multi-channel feature and the target following preferentially updated parallel Pending CN110189361A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910451153.0A CN110189361A (en) 2019-05-28 2019-05-28 A kind of multi-channel feature and the target following preferentially updated parallel

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910451153.0A CN110189361A (en) 2019-05-28 2019-05-28 A kind of multi-channel feature and the target following preferentially updated parallel

Publications (1)

Publication Number Publication Date
CN110189361A true CN110189361A (en) 2019-08-30

Family

ID=67718198

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910451153.0A Pending CN110189361A (en) 2019-05-28 2019-05-28 A kind of multi-channel feature and the target following preferentially updated parallel

Country Status (1)

Country Link
CN (1) CN110189361A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112634344A (en) * 2020-12-15 2021-04-09 西安理工大学 Method for detecting center position of cold-rolled strip coil shaft hole based on machine vision

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
胡昭华 等: "多通道特征和择优并行更新的核相关滤波跟踪", 《HTTP://KNS.CNKI.NET/KCMS/DETAIL/11.2127.TP.20190308.1454.032.HTML》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112634344A (en) * 2020-12-15 2021-04-09 西安理工大学 Method for detecting center position of cold-rolled strip coil shaft hole based on machine vision

Similar Documents

Publication Publication Date Title
CN110322423B (en) Multi-modal image target detection method based on image fusion
CN108510532A (en) Optics and SAR image registration method based on depth convolution GAN
CN111709902A (en) Infrared and visible light image fusion method based on self-attention mechanism
CN107346434A (en) A kind of plant pest detection method based on multiple features and SVMs
CN107145846A (en) A kind of insulator recognition methods based on deep learning
CN106407903A (en) Multiple dimensioned convolution neural network-based real time human body abnormal behavior identification method
CN105740758A (en) Internet video face recognition method based on deep learning
CN108304826A (en) Facial expression recognizing method based on convolutional neural networks
CN110443827A (en) A kind of UAV Video single goal long-term follow method based on the twin network of improvement
CN109741318A (en) The real-time detection method of single phase multiple dimensioned specific objective based on effective receptive field
CN105160310A (en) 3D (three-dimensional) convolutional neural network based human body behavior recognition method
CN109086754A (en) A kind of human posture recognition method based on deep learning
CN107423702A (en) Video target tracking method based on TLD tracking systems
CN108520202A (en) Confrontation robustness image characteristic extracting method based on variation spherical projection
CN108921019A (en) A kind of gait recognition method based on GEI and TripletLoss-DenseNet
CN110175649A (en) It is a kind of about the quick multiscale estimatiL method for tracking target detected again
CN109785366A (en) It is a kind of for the correlation filtering method for tracking target blocked
CN103295221B (en) The waterborne target method for testing motion of simulation compound eye visual mechanism and polarization imaging
CN106997475A (en) A kind of insect image-recognizing method based on parallel-convolution neutral net
CN107945210A (en) Target tracking algorism based on deep learning and environment self-adaption
CN105893947A (en) Bi-visual-angle face identification method based on multi-local correlation characteristic learning
CN110048978A (en) A kind of signal modulate method
CN107463954A (en) A kind of template matches recognition methods for obscuring different spectrogram picture
CN108682022A (en) Based on the visual tracking method and system to anti-migration network
CN106991428A (en) Insect image-recognizing method based on adaptive pool model

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20190830

WD01 Invention patent application deemed withdrawn after publication