CN110033473A - Motion target tracking method based on template matching and depth sorting network - Google Patents
Motion target tracking method based on template matching and depth sorting network Download PDFInfo
- Publication number
- CN110033473A CN110033473A CN201910297980.9A CN201910297980A CN110033473A CN 110033473 A CN110033473 A CN 110033473A CN 201910297980 A CN201910297980 A CN 201910297980A CN 110033473 A CN110033473 A CN 110033473A
- Authority
- CN
- China
- Prior art keywords
- template
- target
- image
- network
- residual error
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
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
-
- 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 invention discloses a kind of motion target tracking method based on template matching and depth sorting network, mainly solves that prior art target detection speed is slow, and appearance deformation occurs in target, blocks and constantly tracks the problem of inaccuracy.Its implementation are as follows: 1) build double residual error depth sorting networks, and it is trained；2) template network and detection network are extracted in double residual error depth sorting networks；3) template characteristic is extracted using template network；4) detection feature is extracted using detection network；5) template characteristic is subjected to template matching in detection feature, obtains template matching figure；6) target position is determined according to template matching figure；7) template characteristic is updated according to target position；8) judge whether present frame is last frame, if so, target end tracks, otherwise, using updated template characteristic as the template characteristic of next frame, return 4).Tracking velocity of the present invention is fast, and accuracy rate is high, for drastic mechanical deformation, illumination variation video frequency object tracking.
Description
Technical field
The invention belongs to technical field of image processing, further relate to a kind of motion target tracking method, can be used for pair
The video frequency object tracking of these types such as drastic mechanical deformation, camera lens shake, dimensional variation, illumination variation.
Background technique
The main task of motion target tracking requires the study in the case where just knowing that target to be tracked initial frame information to arrive
One tracker, so that tracker can be with the position of Accurate Prediction target to be tracked next frame in the video sequence.With people
To the understanding that deepens continuously of computer vision field, motion target tracking is used widely and is developed in the field, due to depth
Degree learns the continuous application in image classification, image segmentation field, and deep learning method is also gradually applied to target following neck
Domain.Compared to the manual extraction characterization method for the priori knowledge for excessively relying on designer in traditional tracking, deep learning side
Method can use the advantage of big data, by the training of mass data, neural network can automatic learning characteristic, deposited at present
Motion target tracking is realized in a large amount of track algorithm.But due to blocking, background is mixed and disorderly, appearance deformation, illumination variation,
The influence of the objective factors such as visual angle change, so that being accurately tracked by target still suffers from great challenge.
A kind of patent document " anti-method for tracking target blocked the " (patent application of Nanjing Aero-Space University in its application
Numbers 201610818828.7, publication number 106408591A) in disclose a kind of target following side based on detection, tracking and study
Method.What this method was realized comprises the concrete steps that, firstly, determining mesh target area according to initial picture frame, tracker passes through the mesh
It marks region and forms initial target template；Secondly, initialization cascade detectors parameter；Then, the testing mechanism blocked is added, and real
Shi Gengxin threshold value；Then, tracker and detector are calculated separately to the tracking creditability of target and detection confidence level；Finally, root
Tracking result to be integrated according to confidence level, fails as tracker tracks, is initialized with testing result, tracking result passes through study module,
Detector relevant parameter is updated.Shortcoming existing for this method is to utilize the weighted results of target template and background template
As confidence value, failing the fluctuation situation for reflecting target response to be tracked, the classifier recognition capability that training obtains is not strong enough,
It cannot achieve when strong illumination variation occurs for target, target quickly moves and be accurately tracked by for a long time.
Agricultural University Of South China is in patent document " method for tracking target based on local feature learning " (patent Shen of its application
Please numbers 201610024953.0, publication number 108038435A) in disclose it is a kind of using local feature learning to moving target with
Track method.What this method was realized comprises the concrete steps that a large amount of local units are resolved into target area and background area by (1), uses
The mode training of deep learning, building apparent model；(2) confidence that each regional area of next frame image belongs to target is calculated
Degree obtains the confidence level figure positioned for target；(3) threshold values T is setposAnd Tneg, threshold values is greater than TposRegional area be added
Threshold values is less than T by target sample collectionnegRegional area be added background sample collection, update apparent model.Existing for this method not
Foot place is the sample type for needing to judge by setting threshold values each regional area of image, when target to be tracked generation is larger
When degree is blocked, it can cause updated model that can not continue to accurately track target target sample or background sample mistake point.
Summary of the invention
It is a kind of based on template matching and depth sorting net the purpose of the present invention is in view of the above shortcomings of the prior art, proposing
The motion target tracking method of network, to realize the standard to target in the case where target generates deformation, dimensional variation or blocks
Really, it effectively tracks.
Realizing the object of the invention technical solution is, firstly, choosing the instruction of training under line aiming at the problem that lack of training samples
Practice mechanism；Secondly, constructing template network and detection network using Resnet 50, the spy of template image is carried out using template network
Sign is extracted, and the feature extraction of image to be detected is carried out using detection network；Finally, the template characteristic of extraction is mentioned with detection template
The feature taken is matched, and determines target position, and specific steps include the following:
(1) double residual error depth sorting network models are built:
(1a) is using two depth residual error neural network ResNet50 as the front end net of double residual error depth sorting network models
The input layer parameter of network, the two depth residual error neural networks is different, and the parameter of other layers is identical.
(1b) builds back-end network of two 3 layers of the fully-connected network as double residual error depth sorting network models, each
The first layer of fully-connected network is input layer, and the second layer is hidden layer, and third layer is output layer, the of the two fully-connected networks
One layer parameter is different, and the second layer, the parameter of third layer are identical；
(2) ImageNet categorized data set is input to double residual error depth sorting network models, uses stochastic gradient descent
Method updates the weight of each node in double residual error depth sorting network models, obtains trained double residual error depth sorting nets
Network model；
(3) reciprocal the of depth residual error network ResNet50 is deleted in trained double residual error depth sorting network models
All layers after two hidden layers obtain template network model and detection network model；
(4) template characteristic figure is extracted using template network:
(4a) input is containing first frame image in the sequence of video images of target to be tracked, in target initial position to be tracked
Center at, determine a rectangle frame with one times of length and width of target to be tracked；
(4b) intercepts target image from rectangle frame, and adjustment image size is 224 × 224 × 3 pixels, obtains Prototype drawing
Picture；
Template image is input in template network by (4c), extracts the feature of image, and all features are formed characteristics of image
Figure exports 2048 7 × 7 template characteristic figures in the last layer of template network；
(5) detection characteristic pattern is extracted using detection network:
(5a) input contain target to be tracked image to be detected, at the center of target initial position to be tracked, with to
Twice of length and width of tracking target determine a rectangle frame；
(5b) intercepts target image from rectangle frame, and adjustment image size is 448 × 448 × 3 pixels, obtains detection figure
Picture；
(5c) will test image and be input in detection network, extract the feature of image, and all features are formed characteristics of image
Figure exports 2048 14 × 14 detection characteristic patterns in the last layer of detection network；
(6) template matching:
(6a) corresponds the detection characteristic pattern of 2048 template characteristic figures and 2048, forms 2048 moulds
Plate detects feature pair；
(6b) in every a pair of of template detection feature pair, by 7 × 7 template characteristic figure on 14 × 14 detection characteristic pattern
The convolution for carrying out sliding sash mode, obtains 2048 14 × 14 template matching figures；
(6c) corresponds 14 × 14 pixels in 2048 template matching figures, and by the matching value in corresponding points
Summation operation is carried out, one 14 × 14 characteristic response figure is obtained.
(7) target position is determined:
Response in 14 × 14 characteristic response figures is ranked up by (7a) from big to small, and it is corresponding to choose preceding 10 responses
Normalized coordinate, seek its average normalized coordinate value；
(7b) is calculate by the following formula position of the tracking target in video frame images according to average normalized coordinate value；
X '=x × m+a-w, y '=y × n+b-h
Wherein, in x ' expression video frame first pixel in the target image upper left corner abscissa value, x indicate it is average normalized
Abscissa, a indicate the abscissa value of target initial position to be tracked, and w indicates that the width of template image, m indicate detection image
Width, the ordinate value of first pixel in the target image upper left corner in y ' expression video frame, y indicate average normalized ordinate, b
Indicate that the ordinate value of target initial position to be tracked, h indicate template image height, n indicates the height of detection image.
(8) tracking target signature is extracted according to position of the tracking target in video frame images, it is special according to tracking target
Sign figure, updates template characteristic figure: Z=η Z1+(1-η)Z2, wherein Z indicates updated template characteristic figure, Z1Indicate previous frame figure
Template characteristic figure as in, η indicate the learning rate of template renewal, wherein | η |≤1, Z2Indicate the tracking mesh in current video frame
Mark characteristic pattern；
(9) judge current frame video image whether be sequence of video images to be tracked last frame video image, if so,
Then terminate the tracking to target to be tracked is moved, otherwise, using updated template characteristic figure as next frame target to be tracked
Template characteristic figure returns (5), completes target following.
The present invention has the advantage that compared with prior art
First, since the present invention is using the mechanism of training categorized data set under line, overcome in the prior art in training net
Iterated when network using first frame image, be easy over-fitting, and when target to be tracked generate largely deformation when, cause with
The problem of track inaccuracy allows the invention to more accurately track target when target to be tracked generates larger deformation.
Second, since the present invention constructs double residual error depth sorting network models, extracted using template network and detection network
Characteristics of image matched, the position of target to be tracked is judged using response, is overcome in the prior art when mesh to be tracked
When mark generation is largely blocked, easily by positive negative sample mistake point, cause updated model that can not continue to accurately track target
Problem allows the invention to more accurately track target when target to be tracked is generated and largely blocked.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is analogous diagram of the invention.
Specific embodiment
The embodiment of the present invention and effect are further described with reference to the accompanying drawing.
Referring to Fig.1, to the specific steps of the present invention are as follows.
Step 1, double residual error depth sorting network models are built.
1.1) front network is set:
The input layer parameter of existing two depth residual error neural network ResNet50 is adjusted, wherein first
The neuron number of network input layer is set as 224 × 224 × 3, and the neuron number setting 448 of second network input layer ×
448 × 3, other each layer parameters remain unchanged, and using the two depth residual error neural networks as double residual error depth sorting networks
The front network of model；
1.2) back-end network is set:
Back-end network of two three layers of the fully-connected network as double residual error depth sorting network models is built, is each connected entirely
The first layer for connecing network is input layer, and the second layer is hidden layer, and third layer is output layer, the first layer of the two fully-connected networks
Parameter is different, and the second layer, the parameter of third layer are identical, wherein the parameter of each layer is as follows in two fully-connected networks:
The number of the first layer neuron of first network is set as 1 × 1 × 2048, the first layer nerve of second network
The number of member is set as 2 × 2 × 2048；
The number of the neuron of the second layer of the two networks is set as 1024 simultaneously, and activation primitive is set as correcting simultaneously
Linear unit ReLU function；
The neuron number of the third layer of the two networks is set as 1000 simultaneously, and activation primitive is set as simultaneously
Softmax function.
Step 2, the double residual error depth sorting network models of training.
ImageNet categorized data set is input in double residual error depth sorting network models that step 1 is built, using with
Machine gradient descent method updates the weight of each node in double residual error depth sorting network models, obtains trained double residual errors
Depth sorting network model:
(2a) selects a number at random in (0,0.1) range, uses the number as every in double residual error depth sorting network models
The initial weight of a node；
(2b) is using the initial weight of each node as in residual error depth sorting network models double in first time iterative process
The current weight of each node；
(2c) randomly selects 2 from ImageNet classification data concentrationnA sample image is in double residual error depth sorting network moulds
Forward-propagating in type, the output layer output 2 of double residual error depth sorting network modelsnThe classification results of a sample image, wherein 3≤
n≤7；
(2d) calculates the average log penalty values of classification results according to the following formula according to the classification results of sample image:
Wherein, the average log penalty values of L presentation class result, N indicate the sum of the sample image randomly selected, i table
Show the serial number of input sample image, yiIndicate the classification of i-th of input sample image, the y of positive class sampleiValue is 1, negative class sample
This yiValue is 0, piDouble residual error depth sorting network model output valves of i-th of sample image in presentation class result；
(2e) seeks local derviation with current weight of the average log penalty values to each node in double residual error depth sorting networks,
Obtain the gradient value Δ w of each node current weight in double residual error depth sorting network modelsk；
(2f) calculates each node updates in double residual error depth sorting network models according to the gradient value of node current weight
Weight afterwards:
Wherein,Weight after indicating double residual error depth sorting k-th of node updates of network model, wkIndicate that double residual errors are deep
The current weight of k-th of node of sorter network model is spent, α indicates learning rate, and value range is (0,1)；
Whether the sample image in (2g) training of judgement data set is all selected, if so, obtaining trained double residual
Poor depth sorting network model, otherwise, using the weight after each node updates as current weight after, execute (3c).
Step 3, template network model and detection network model are extracted.
In the trained double residual error depth sorting network models obtained in step 2, two depth networks are deleted respectively
The 49th layer after network layer, remaining network become new network.
According to input layer parameter extraction template network model and detection network model from remaining network, i.e., by input layer
The rest network that parameter is 224 × 224 × 3 will input remaining net of the layer parameter for 448 × 448 × 3 as template network model
Network is as detection network model.
Step 4, template characteristic figure is extracted using template network.
(4a) input is containing first frame image in the sequence of video images of target to be tracked, in target initial position to be tracked
Center at, with one rectangle frame of a double-length of target to be tracked and wide determination；
(4b) intercepts target image from rectangle frame, and adjustment image size is 224 × 224 × 3 pixels, obtains Prototype drawing
Picture；
Template image is input in template network obtained in step 3 and carries out feature extraction to template image by (4c), and
After the feature composition characteristic figure of extraction, export 2048 7 × 7 characteristic patterns in the last layer of template network, by this 2048
A 7 × 7 characteristic pattern is as template characteristic figure.
Step 5, detection characteristic pattern is extracted using detection network.
(5a) input contain target to be tracked image to be detected, at the center of target initial position to be tracked, with to
Twice of length and width of tracking target determine a rectangle frame.
(5b) intercepts target image from rectangle frame, and adjustment image size is 448 × 448 × 3 pixels, obtains detection figure
Picture；
(5c) will test image and be input in the detection network that step 3 obtains to detection image progress feature extraction, and will
After the feature composition characteristic figure of extraction, export 2048 14 × 14 characteristic patterns in the last layer of detection network, by this 2048
A 14 × 14 characteristic pattern is as detection characteristic pattern.
Step 6, template matching.
(6a) is by 2048 obtained in 2048 obtained in step 4 template characteristic figures and step 5 detection features
Figure is corresponded, and 2048 template detection features pair are formed；
(6b) is step with 1 pixel to detect the upper left corner of characteristic pattern as starting point in every a pair of of template detection feature pair
It is long, corresponding template characteristic figure is successively moved to the upper right corner, the lower right corner, the lower left corner of detection characteristic pattern, finally translates Hui Zuo
Upper angle carries out convolution algorithm, obtains 2048 14 × 14 template matching figures；
(6c) corresponds 14 × 14 pixels in 2048 template matching figures, and by the matching value in corresponding points
Summation operation is carried out, one 14 × 14 characteristic response figure is obtained.
Step 7, target position is determined.
(7a) according to being ranked up from big to small, selects the response in 14 × 14 characteristic response figures obtained in step 6
The corresponding normalized coordinate of preceding 10 responses is taken, this 10 normalized coordinate values are averaging operation, are obtained average normalized
Coordinate value (x, y)；
(7b) is calculate by the following formula position of the tracking target in video frame images according to average normalized coordinate value:
X '=x × m+a-w,
Y '=y × n+b-h,
Wherein, in x ' expression video frame first pixel in the target image upper left corner abscissa value, a indicates target to be tracked
The abscissa value of initial position, w indicate that the width of template image, m indicate the width of detection image, target in y ' expression video frame
The ordinate value of first pixel in the image upper left corner, b indicate that the ordinate value of target initial position to be tracked, h indicate Prototype drawing
Image height degree, n indicate the height of detection image.
Step 8, template renewal.
Detection characteristic pattern in, centered on position of the tracking target acquired in step 7 in video frame images, with
The initial size of track target is that size carries out shearing manipulation, obtains tracking target signature, according to tracking target signature, is updated
Template characteristic figure:
Z=η Z1+(1-η)Z2,
Wherein, Z indicates updated template characteristic figure, Z1Indicate that the template characteristic figure in previous frame image, η indicate template
The learning rate of update, | η |≤1, Z2Indicate the tracking target signature in current video frame.
Step 9, judge current frame video image whether be sequence of video images to be tracked last frame video image, if
The tracking then terminated to target to be tracked is moved, otherwise, using the updated template characteristic figure of step 8 as next frame wait for
The template characteristic figure of track target, return step 5 complete target following.
Effect of the invention is described further below with reference to emulation experiment.
1. emulation experiment condition:
The hardware test platform of emulation experiment of the present invention is: CPU be intel Core i5-6500, dominant frequency 3.2GHz,
Memory 8GB, GPU are NVIDIA TITAN Xp；Software platform is: 16.04 LTS of Ubuntu, 64 bit manipulation systems, python
3.6.5。
2. emulation content and result:
With the method for the present invention to the Duan Yiming acquired from 2015 database of Object tracking benchmark
The sequence of video images that man on the way walks about carries out motion target tracking emulation experiment, which shares 252 frames
Video image, the result of emulation experiment is as shown in figure (2), in which:
Fig. 2 (a) is the 1st frame image of the sequence of video images of emulation experiment acquisition, and white line rectangle frame indicates in Fig. 2 (a)
The initial position of target to be tracked.
Fig. 2 (b) is in emulation experiment of the present invention, to the sequence of video images of acquisition, carry out target following a frame wait for
The tracking result of appearance deformation and video image when target occlusion occurs for track target, wherein gray line rectangle frame mark be to
The predicted position of target is tracked, what white line rectangle frame marked is the actual position of target to be tracked, it can be seen from the graph that mesh to be tracked
Appearance deformation and target occlusion has occurred compared with the target to be tracked in Fig. 2 (a) in mark.
Fig. 2 (c) is in emulation experiment of the present invention, and the frame for carrying out target following to the sequence of video images of acquisition is to be tracked
The tracking result of appearance deformation and video image when illumination variation occurs for target, wherein gray line rectangle frame mark be to
The predicted position of track target, what white line rectangle frame marked is the position of target to be tracked.It can be seen from the graph that target to be tracked and figure
Target to be tracked in 2 (a) is compared, and appearance deformation and illumination enhancing has occurred.
What the target and white line rectangle frame that the gray line rectangle frame in figure it can be seen from Fig. 2 (b) and Fig. 2 (c) outlines outlined
Target is consistent, illustrates that the present invention can be accurate, high when target to be tracked generates deformation, illumination variation, blocks in video image
Effect ground tracking target.
Claims (5)
1. a kind of motion target tracking method based on template matching and depth sorting network, which is characterized in that include the following:
(1) double residual error depth sorting network models are built:
(1a) using two depth residual error neural network ResNet50 as the front network of double residual error depth sorting network models, this
The input layer parameter of two depth residual error neural networks is different, and the parameter of other layers is identical.
(1b) builds back-end network of two 3 layers of the fully-connected network as double residual error depth sorting network models, each connects entirely
The first layer for connecing network is input layer, and the second layer is hidden layer, and third layer is output layer, the first layer of the two fully-connected networks
Parameter is different, and the second layer, the parameter of third layer are identical；
(2) ImageNet categorized data set is input to double residual error depth sorting network models, using stochastic gradient descent method, more
The weight of each node, obtains trained double residual error depth sorting network moulds in new double residual error depth sorting network models
Type；
(3) penultimate of depth residual error network ResNet50 is deleted in trained double residual error depth sorting network models
All layers after hidden layer obtain template network model and detection network model；
(4) template characteristic figure is extracted using template network:
(4a) input is containing first frame image in the sequence of video images of target to be tracked, in target initial position to be tracked
At the heart, a rectangle frame is determined with one times of length and width of target to be tracked；
(4b) intercepts target image from rectangle frame, and adjustment image size is 224 × 224 × 3 pixels, obtains template image；
Template image is input in template network by (4c), extracts the feature of image, and all features are formed characteristics of image figure,
The last layer of template network exports 2048 7 × 7 template characteristic figures；
(5) detection characteristic pattern is extracted using detection network:
(5a) input contains image to be detected of target to be tracked, at the center of target initial position to be tracked, with to be tracked
Twice of length and width of target determine a rectangle frame；
(5b) intercepts target image from rectangle frame, and adjustment image size is 448 × 448 × 3 pixels, obtains detection image；
(5c) will test image and be input in detection network, extract the feature of image, and all features are formed characteristics of image figure,
The last layer for detecting network exports 2048 14 × 14 detection characteristic patterns；
(6) template matching:
(6a) corresponds the detection characteristic pattern of 2048 template characteristic figures and 2048, forms 2048 template inspections
Survey feature pair；
(6b) carries out 7 × 7 template characteristic figure in every a pair of of template detection feature pair on 14 × 14 detection characteristic pattern
The convolution of sliding sash mode obtains 2048 14 × 14 template matching figures；
(6c) corresponds 14 × 14 pixels in 2048 template matching figures, and the matching value in corresponding points is carried out
Summation operation obtains one 14 × 14 characteristic response figure.
(7) target position is determined:
Response in 14 × 14 characteristic response figures is ranked up by (7a) from big to small, and preceding 10 responses of selection are corresponding to return
One changes coordinate, seeks its average normalized coordinate value；
(7b) is calculate by the following formula position of the tracking target in video frame images according to average normalized coordinate value；
X '=x × m+a-w, y '=y × n+b-h,
Wherein, in x ' expression video frame first pixel in the target image upper left corner abscissa value, x indicates average normalized horizontal seat
Mark, a indicate the abscissa value of target initial position to be tracked, and w indicates that the width of template image, m indicate the width of detection image,
The ordinate value of first pixel in the target image upper left corner in y ' expression video frame, y indicate that average normalized ordinate, b indicate
The ordinate value of target initial position to be tracked, h indicate template image height, and n indicates the height of detection image.
(8) tracking target signature is extracted according to position of the tracking target in video frame images, according to tracking target signature,
Update template characteristic figure: Z=η Z1+(1-η)Z2, wherein Z indicates updated template characteristic figure, Z1It indicates in previous frame image
Template characteristic figure, η indicate template renewal learning rate, wherein | η |≤1, Z2Indicate that the tracking target in current video frame is special
Sign figure；
(9) judge whether current frame video image is the last frame video image of sequence of video images to be tracked, if so, tying
Tracking of the beam to target to be tracked is moved, otherwise, using updated template characteristic figure as the template of next frame target to be tracked
Characteristic pattern returns (5), completes target following.
2. the method according to claim 1, wherein two depth residual error neural network ResNet50 in (1a)
The number of input layer, neuron is respectively set to 224 × 224 × 3 and 448 × 448 × 3.
3. the method according to claim 1, wherein the parameter setting of two each layers of fully-connected network is such as in (1b)
Under:
The number of first layer neuron is respectively 1 × 1 × 2048 and 2 × 2 × 2048；
The number of the neuron of the second layer is 1024, and activation primitive is set as correcting linear unit R eLU function；
The neuron number of third layer is 1000, and activation primitive is set as Softmax function.
4. it is deep to update double residual errors the method according to claim 1, wherein using stochastic gradient descent method in (2)
Spend the weight of each node in sorter network model, the specific steps are as follows:
(2a) selects a number at random in (0,0.1) range, uses the number as each section in double residual error depth sorting network models
The initial weight of point；
(2b) is using the initial weight of each node as each in residual error depth sorting network models double in first time iterative process
The current weight of node；
(2c) randomly selects 2 from ImageNet classification data concentrationnA sample image in double residual error depth sorting network models just
To propagation, wherein 3≤n≤7, the output layer output 2 of double residual error depth sorting network modelsnThe classification results of a sample image；
(2d) calculates the average log penalty values of classification results according to the following formula according to the classification results of sample image:
Wherein, the average log penalty values of L presentation class result, N indicate that the sum of the sample image randomly selected, i indicate defeated
Enter the serial number of sample image, yiIndicate the classification of i-th of input sample image, the y of positive class sampleiValue takes 1, the y of negative class samplei
Value takes 0, piDouble residual error depth sorting network model output valves of i-th of sample image in presentation class result；
(2e) seeks local derviation with current weight of the average log penalty values to each node in double residual error depth sorting networks, obtains
The gradient value Δ w of each node current weight in double residual error depth sorting network modelsk；
(2f) is calculated in double residual error depth sorting network models after each node updates according to the gradient value of node current weight
Weight:
Wherein,Weight after indicating double residual error depth sorting k-th of node updates of network model, wkIndicate double residual error depth point
The current weight of k-th of node of class network model, α indicate learning rate, and value range is (0,1)；
Whether the sample image in (2g) training of judgement data set is all selected, if so, it is deep to obtain trained double residual errors
Sorter network model is spent, otherwise, executes (2c) using the weight after each node updates as after current weight.
5. the method according to claim 1, wherein the sliding sash mode convolution in (6b), is with every a pair of of template
The upper left corner for detecting the detection characteristic pattern of feature centering is starting point, using 1 pixel as step-length, successively by corresponding template characteristic figure
The upper right corner, the lower right corner, the lower left corner for moving to detection characteristic pattern finally translate go back to the upper left corner and carry out convolution algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910297980.9A CN110033473B (en) | 2019-04-15 | 2019-04-15 | Moving target tracking method based on template matching and depth classification network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910297980.9A CN110033473B (en) | 2019-04-15 | 2019-04-15 | Moving target tracking method based on template matching and depth classification network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110033473A true CN110033473A (en) | 2019-07-19 |
| CN110033473B CN110033473B (en) | 2021-04-20 |
Family
ID=67238315
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910297980.9A Active CN110033473B (en) | 2019-04-15 | 2019-04-15 | Moving target tracking method based on template matching and depth classification network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110033473B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110647836A (en) * | 2019-09-18 | 2020-01-03 | 中国科学院光电技术研究所 | Robust single-target tracking method based on deep learning |
| CN110705479A (en) * | 2019-09-30 | 2020-01-17 | 北京猎户星空科技有限公司 | Model training method, target recognition method, device, equipment and medium |
| CN110766725A (en) * | 2019-10-31 | 2020-02-07 | 北京市商汤科技开发有限公司 | Template image updating method and device, target tracking method and device, electronic equipment and medium |
| CN110766724A (en) * | 2019-10-31 | 2020-02-07 | 北京市商汤科技开发有限公司 | Target tracking network training and tracking method and device, electronic equipment and medium |
| CN110930428A (en) * | 2020-02-19 | 2020-03-27 | 成都纵横大鹏无人机科技有限公司 | Target tracking method and device, electronic equipment and storage medium |
| CN111145215A (en) * | 2019-12-25 | 2020-05-12 | 北京迈格威科技有限公司 | Target tracking method and device |
| CN111461010A (en) * | 2020-04-01 | 2020-07-28 | 贵州电网有限责任公司 | Power equipment identification efficiency optimization method based on template tracking |
| CN111640136A (en) * | 2020-05-23 | 2020-09-08 | 西北工业大学 | Depth target tracking method in complex environment |
| CN112287906A (en) * | 2020-12-18 | 2021-01-29 | 中汽创智科技有限公司 | Template matching tracking method and system based on depth feature fusion |
| CN112818801A (en) * | 2021-01-26 | 2021-05-18 | 每步科技(上海)有限公司 | Motion counting method, recognition device, recognition system and storage medium |
| CN115100441A (en) * | 2022-08-23 | 2022-09-23 | 浙江大华技术股份有限公司 | Object detection method, electronic device, and storage medium |
| CN116596958A (en) * | 2023-07-18 | 2023-08-15 | 四川迪晟新达类脑智能技术有限公司 | Target tracking method and device based on online sample augmentation |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20020154820A1 (en) * | 2001-03-06 | 2002-10-24 | Toshimitsu Kaneko | Template matching method and image processing device |
| US6621929B1 (en) * | 1999-06-22 | 2003-09-16 | Siemens Corporate Research, Inc. | Method for matching images using spatially-varying illumination change models |
| US20070250898A1 (en) * | 2006-03-28 | 2007-10-25 | Object Video, Inc. | Automatic extraction of secondary video streams |
| CN101867699A (en) * | 2010-05-25 | 2010-10-20 | 中国科学技术大学 | Real-time tracking method of nonspecific target based on partitioning |
| CN103150572A (en) * | 2012-12-11 | 2013-06-12 | 中国科学院深圳先进技术研究院 | On-line type visual tracking method |
| CN105719292A (en) * | 2016-01-20 | 2016-06-29 | 华东师范大学 | Method of realizing video target tracking by adopting two-layer cascading Boosting classification algorithm |
| CN105787963A (en) * | 2016-02-26 | 2016-07-20 | 浪潮软件股份有限公司 | Video target tracking method and device |
| CN107689052A (en) * | 2017-07-11 | 2018-02-13 | 西安电子科技大学 | Visual target tracking method based on multi-model fusion and structuring depth characteristic |
| CN108694723A (en) * | 2018-05-11 | 2018-10-23 | 西安天和防务技术股份有限公司 | A kind of target in complex environment tenacious tracking method |
| CN109345559A (en) * | 2018-08-30 | 2019-02-15 | 西安电子科技大学 | Expand the motion target tracking method with depth sorting network based on sample |
-
2019
- 2019-04-15 CN CN201910297980.9A patent/CN110033473B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6621929B1 (en) * | 1999-06-22 | 2003-09-16 | Siemens Corporate Research, Inc. | Method for matching images using spatially-varying illumination change models |
| US20020154820A1 (en) * | 2001-03-06 | 2002-10-24 | Toshimitsu Kaneko | Template matching method and image processing device |
| US20070250898A1 (en) * | 2006-03-28 | 2007-10-25 | Object Video, Inc. | Automatic extraction of secondary video streams |
| WO2007126780A2 (en) * | 2006-03-28 | 2007-11-08 | Object Video, Inc. | Automatic extraction of secondary video streams |
| CN101867699A (en) * | 2010-05-25 | 2010-10-20 | 中国科学技术大学 | Real-time tracking method of nonspecific target based on partitioning |
| CN103150572A (en) * | 2012-12-11 | 2013-06-12 | 中国科学院深圳先进技术研究院 | On-line type visual tracking method |
| CN105719292A (en) * | 2016-01-20 | 2016-06-29 | 华东师范大学 | Method of realizing video target tracking by adopting two-layer cascading Boosting classification algorithm |
| CN105787963A (en) * | 2016-02-26 | 2016-07-20 | 浪潮软件股份有限公司 | Video target tracking method and device |
| CN107689052A (en) * | 2017-07-11 | 2018-02-13 | 西安电子科技大学 | Visual target tracking method based on multi-model fusion and structuring depth characteristic |
| CN108694723A (en) * | 2018-05-11 | 2018-10-23 | 西安天和防务技术股份有限公司 | A kind of target in complex environment tenacious tracking method |
| CN109345559A (en) * | 2018-08-30 | 2019-02-15 | 西安电子科技大学 | Expand the motion target tracking method with depth sorting network based on sample |
Non-Patent Citations (1)
| Title |
|---|
| TIAN, XIAOLIN 等: "Tracking with spatial constrained coding", 《IET COMPUTER VISION》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110647836A (en) * | 2019-09-18 | 2020-01-03 | 中国科学院光电技术研究所 | Robust single-target tracking method based on deep learning |
| CN110705479A (en) * | 2019-09-30 | 2020-01-17 | 北京猎户星空科技有限公司 | Model training method, target recognition method, device, equipment and medium |
| CN110766725B (en) * | 2019-10-31 | 2022-10-04 | 北京市商汤科技开发有限公司 | Template image updating method and device, target tracking method and device, electronic equipment and medium |
| CN110766725A (en) * | 2019-10-31 | 2020-02-07 | 北京市商汤科技开发有限公司 | Template image updating method and device, target tracking method and device, electronic equipment and medium |
| CN110766724A (en) * | 2019-10-31 | 2020-02-07 | 北京市商汤科技开发有限公司 | Target tracking network training and tracking method and device, electronic equipment and medium |
| CN110766724B (en) * | 2019-10-31 | 2023-01-24 | 北京市商汤科技开发有限公司 | Target tracking network training and tracking method and device, electronic equipment and medium |
| CN111145215A (en) * | 2019-12-25 | 2020-05-12 | 北京迈格威科技有限公司 | Target tracking method and device |
| CN111145215B (en) * | 2019-12-25 | 2023-09-05 | 北京迈格威科技有限公司 | Target tracking method and device |
| CN110930428A (en) * | 2020-02-19 | 2020-03-27 | 成都纵横大鹏无人机科技有限公司 | Target tracking method and device, electronic equipment and storage medium |
| CN111461010B (en) * | 2020-04-01 | 2022-08-12 | 贵州电网有限责任公司 | Power equipment identification efficiency optimization method based on template tracking |
| CN111461010A (en) * | 2020-04-01 | 2020-07-28 | 贵州电网有限责任公司 | Power equipment identification efficiency optimization method based on template tracking |
| CN111640136B (en) * | 2020-05-23 | 2022-02-25 | 西北工业大学 | Depth target tracking method in complex environment |
| CN111640136A (en) * | 2020-05-23 | 2020-09-08 | 西北工业大学 | Depth target tracking method in complex environment |
| CN112287906A (en) * | 2020-12-18 | 2021-01-29 | 中汽创智科技有限公司 | Template matching tracking method and system based on depth feature fusion |
| CN112818801A (en) * | 2021-01-26 | 2021-05-18 | 每步科技(上海)有限公司 | Motion counting method, recognition device, recognition system and storage medium |
| CN112818801B (en) * | 2021-01-26 | 2024-04-26 | 每步科技(上海)有限公司 | Motion counting method, recognition device, recognition system and storage medium |
| CN115100441A (en) * | 2022-08-23 | 2022-09-23 | 浙江大华技术股份有限公司 | Object detection method, electronic device, and storage medium |
| CN115100441B (en) * | 2022-08-23 | 2022-11-18 | 浙江大华技术股份有限公司 | Object detection method, electronic device, and storage medium |
| CN116596958A (en) * | 2023-07-18 | 2023-08-15 | 四川迪晟新达类脑智能技术有限公司 | Target tracking method and device based on online sample augmentation |
| CN116596958B (en) * | 2023-07-18 | 2023-10-10 | 四川迪晟新达类脑智能技术有限公司 | Target tracking method and device based on online sample augmentation |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110033473B (en) | 2021-04-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110033473A (en) | Motion target tracking method based on template matching and depth sorting network | |
| Yang et al. | Real-time face detection based on YOLO | |
| CN111259930B (en) | General target detection method of self-adaptive attention guidance mechanism | |
| CN108010067B (en) | A kind of visual target tracking method based on combination determination strategy | |
| CN107145908B (en) | A kind of small target detecting method based on R-FCN | |
| CN106709936A (en) | Single target tracking method based on convolution neural network | |
| CN110119728A (en) | Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network | |
| CN109102547A (en) | Robot based on object identification deep learning model grabs position and orientation estimation method | |
| CN104850865B (en) | A kind of Real Time Compression tracking of multiple features transfer learning | |
| CN108960086A (en) | Based on the multi-pose human body target tracking method for generating confrontation network positive sample enhancing | |
| CN109903312A (en) | A kind of football sportsman based on video multi-target tracking runs distance statistics method | |
| CN109613006A (en) | A kind of fabric defect detection method based on end-to-end neural network | |
| CN106529499A (en) | Fourier descriptor and gait energy image fusion feature-based gait identification method | |
| CN106204638A (en) | A kind of based on dimension self-adaption with the method for tracking target of taking photo by plane blocking process | |
| CN108304761A (en) | Method for text detection, device, storage medium and computer equipment | |
| CN110276269A (en) | A kind of Remote Sensing Target detection method based on attention mechanism | |
| CN109934115A (en) | Construction method, face identification method and the electronic equipment of human face recognition model | |
| CN106780546B (en) | The personal identification method of motion blur encoded point based on convolutional neural networks | |
| CN110084836A (en) | Method for tracking target based on the response fusion of depth convolution Dividing Characteristics | |
| CN107529650A (en) | The structure and closed loop detection method of network model, related device and computer equipment | |
| CN107507170A (en) | A kind of airfield runway crack detection method based on multi-scale image information fusion | |
| CN109509187A (en) | A kind of efficient check algorithm for the nibs in big resolution ratio cloth image | |
| CN108460790A (en) | A kind of visual tracking method based on consistency fallout predictor model | |
| CN110321811A (en) | Depth is against the object detection method in the unmanned plane video of intensified learning | |
| CN108648211A (en) | A kind of small target detecting method, device, equipment and medium based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |