CN108038435A - A kind of feature extraction and method for tracking target based on convolutional neural networks - Google Patents
A kind of feature extraction and method for tracking target based on convolutional neural networks Download PDFInfo
- Publication number
- CN108038435A CN108038435A CN201711262806.8A CN201711262806A CN108038435A CN 108038435 A CN108038435 A CN 108038435A CN 201711262806 A CN201711262806 A CN 201711262806A CN 108038435 A CN108038435 A CN 108038435A
- Authority
- CN
- China
- Prior art keywords
- network
- target
- training
- feature extraction
- tracking
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of feature extraction based on convolutional neural networks and method for tracking target, first by the method for pre-training under line, improves performance of the network in feature extraction and foreground segmentation；The first frame of video sequence that calibration has is put into network into training on line afterwards, the parameter of network model is finely adjusted, so as to lift disposal ability of the convolutional neural networks in particular problem.Method by increasing random two-dimentional mask and successive ignition, the problem of both having improved the accuracy rate of neural network forecast, while it also avoid over-fitting.Multiple dimensioned washability also causes this method that the scale size of target can be adaptive selected in object tracking process.In terms of network parameter renewal, it is updated in real time by way of threshold value is set, improves the precision and robustness of method for tracking target.
Description
Technical field
The present invention relates to computer realm, more particularly, to a kind of feature extraction based on convolutional neural networks and mesh
Mark tracking.
Background technology
Target Tracking Problem is an important component in computer vision field.Its basic task is to provide one
Section video sequence, and the original state (such as position, size) of target is calibrated in the first frame, by a series of algorithm certainly
It is dynamic to estimate state of the target among subsequent frame.Current track algorithm can be divided into production (generative model)
With two kinds of discriminate (discriminative model).Production method with generation model by describing the apparent spy of target
Sign, search candidate target minimize reconstructed error, representative algorithm such as sparse coding (sparse coding),
Line density estimation (online density estimation), principal component analysis (PCA, principal component
Analysis) etc.；Discriminate method distinguishes target and background by training aids, realizes the real-time tracking to target, common calculation
Method includes multi-instance learning method (multiple instance learning), structure SVM algorithm (structured SVM)
Deng.Compared with production method, discriminate method can more significantly distinguish the foreground and background information in image, show
The more feature of robust, therefore become main stream approach at present, most deep learning methods also using discriminate frame.
Conventional method is when extracting characteristics of image, the method often extracted using hand-designed feature, as SIFT is special
Sign (Scale-invariant feature transform, scale invariant feature conversion), HOG features ((Histogram of
Oriented Gradient, histograms of oriented gradients), LBP features (Local Binary Patterns, local binary patterns)
Etc..The method of manual extraction feature excessively relies on the priori of designer, it is difficult to using the advantage of big data, is carried in feature
The influence of various factors is also limited by terms of the effect taken.The method of deep learning and the maximum difference of conventional method are, lead to
The training of mass data is crossed, neutral net being capable of automatic learning characteristic.In the case where training data is sufficiently large, deep learning obtains
To feature extraction to be far superior to manual method extraction feature.
In recent years, as the continuous development of deep learning, deep learning method are regarded in computers such as target detection, identifications
The key areas of feel has shown very outstanding performance, but general in target tracking domain performance.Main cause just exists
Training set provided in target following is too small, usually only includes the original state demarcated in video sequence in the first frame, difficult
To have given play to the advantage of deep learning method, deep learning method still has very big in the efficiency of tracking, accuracy rate etc.
Room for promotion.
The content of the invention
The present invention is overcomes at least one defect described in the above-mentioned prior art, there is provided a kind of based on convolutional neural networks
Feature extraction and method for tracking target, first by the method for pre-training under line, improve network in feature extraction and foreground segmentation
When performance；The first frame of video sequence that calibration has groundtruth is put into network into training on line, to net afterwards
The parameter of network model is finely adjusted, so as to lift disposal ability of the convolutional neural networks in particular problem.It is random by increasing
The method of two-dimentional mask and successive ignition, the problem of both having improved the accuracy rate of neural network forecast, while it also avoid over-fitting.It is more
The washability of scale also causes this method that the scale size of target can be adaptive selected in object tracking process.In network
In terms of parameter renewal, it is updated in real time by way of threshold value is set, improves the precision and robustness of method for tracking target.
In order to solve the above technical problems, technical scheme is as follows：
A kind of feature extraction and method for tracking target based on convolutional neural networks, comprise the following steps：
S1：Build simultaneously pre-training network model；
S2：According to video sequence, training network model on line；
S3：Input video sequence, calculates tracking result；
S4：The tracking result of previous frame in video sequence is assessed, positive sample result is chosen and is put into iteration in network
To update network parameter；When being updated to network, given threshold θ, works as LsDuring less than threshold value, result is put into network middle school
Practise and update network parameter；Work as LsDuring more than threshold value, then retain tracking result, network is not updated.
In a kind of preferable scheme, step S1 can be divided into following three steps and perform：
S11：Obtain video sequence used in the data set for training foreground segmentation network and target following；
S12：The network model needed for foreground segmentation is built, network model parameter is initialized；
S13：The data that Utilization prospects segmentation network data is concentrated are trained network model, until result restrains.
In a kind of preferable scheme, training set has been used under line in step S11 in advance to the network prospect of convolutional Neural
Segmentation module is trained, and data set used in the training foreground segmentation network is ILSVRC 2014Object
Detection Dataset, can obtain from Imagenet；The data set is the standard data set for target identification, contains 200
Tens of thousands of photos of a variety of different classes of objects, enumerate block, rotate, the different disturbing factor such as deformation.Used in test
Data set is OTB50, which is target following frequently-used data collection, contains 50 different video sequences of length.These
Video sequence shooting is contained and quickly move, blocks, rotating, under illumination variation, the difference scenes such as deformation in different scenes
Follow the trail of result；Video sequence used in the target following is Object Tracking Benchmark 50.
In a kind of preferable scheme, the network model needed for foreground segmentation is built in step S12 and includes 11 for one
The feature extraction network of a convolutional layer, its parameter are as shown in table 1.
In a kind of preferable scheme, with ILSVRC 2014Object Detection Dataset data sets
Image is put into feature extraction network and is trained, until result restrains as label by bounding box；The network
Loss function be：
Wherein pxyFor the predicted value at (x, y) place, txyFor the true value at (x, y) place.This method is according to where target in image
Bounding box location positions train label, t in bounding box locationsxy1 is set to, other regions are set to 0；
The parameter of the feature extraction network of 1 11 convolutional layers of table
In a kind of preferable scheme, step S2 can be divided into following two steps and perform：
S21：The first frame in video sequence is put into network, the tracking object position provided using groundtruth
Put and network model is finely adjusted；
S22：Extract the characteristic extraction part and parameter in training network, reconstruct target positioning network, by video sequence the
One frame is put into the network to reconstruct and iterates, training network model.
Training result in step 2 in step S1, remains Conv4 and its preceding 8 layer of volume in feature extraction network
Lamination and model parameter, have reformulated target positioning network.Foreground extraction knot among original image is used as using the output of Conv4
Fruit, is put into a network comprising two layers of convolutional layer, to obtain center where tracking target.Its parameter is as shown in table 2：
The parameter of 2 one networks comprising two layers of convolutional layer of table
Dimensional Gaussian using the first frame in OTB50 video sequences as training sample, in groundtruth calibration ranges
It is distributed as label and is put into present networks to be trained.Random binary mask successive ignition is added to sample in the training process,
Improve the robustness of network.The loss function of the network is：
Wherein G0For the standard two-dimensional Gaussian Profile in groundtruth calibration ranges, ε is number of channels, XiI-th logical
The distribution output in road, γiFor this multi-channel parametric, the parameters such as convolution nuclear parameter, bias term are contained.After successive ignition, the network
Loss function can slowly restrain；
In a kind of preferable scheme, step S3 can be divided into following two steps and perform：
S31：Image in selecting video sequence successively, on the basis of target location in previous frame, determines tracking target institute
Approximate range, be put into target following network, according to obtained probability distribution matrix after processing, determine in sequence
The center of current tracking target；
S32：Centered on currently definite target's center position, the confidence corresponding to each scale in multiscale space is calculated
Value, chooses scale of the scale corresponding to response maximum as tracking target, and multiscale space is updated.
In a kind of preferable scheme, step S3 is selected in the scope where choosing target according to following scope：
Wherein (x0, y0) demarcating the top left co-ordinate of position by groudtruth, w is the width of bounding box, h
For the height of bounding box.
In a kind of preferable scheme, step S3 calculates confidence value in the following way when determining target scale：
The scope that wherein S is demarcated by scale.When current scale can not obtain higher value, it is updated；
Compared with prior art, the beneficial effect of technical solution of the present invention is：A kind of feature based on convolutional neural networks
Extraction and method for tracking target, first by the method for pre-training under line, improve network in feature extraction and foreground segmentation
Performance；The first frame of video sequence that calibration has groundtruth is put into network into training on line, to network mould afterwards
The parameter of type is finely adjusted, so as to lift disposal ability of the convolutional neural networks in particular problem.With convolutional neural networks
Characteristics of image is extracted, the characteristics of image of extraction has more robustness, while also has preferably characterization to target detail；
Convolution operation is carried out using the small convolution collecting image of multilayer, improves the completeness of extraction feature；On line training when add with
Machine binary mask, carries out successive ignition to the image in training set, had both improved the robustness of network, while also effectively avoid
Over-fitting；Threshold value has been preset during tracking, to network model real-time update, has improved the accuracy rate of tracking.
Brief description of the drawings
Fig. 1 is the step flow diagram of the embodiment of the present invention 1.
Fig. 2 is the overall effect flow diagram of the embodiment of the present invention 1.
Embodiment
Attached drawing is only for illustration, it is impossible to is interpreted as the limitation to this patent；
In order to more preferably illustrate the present embodiment, some components of attached drawing have omission, zoom in or out, and do not represent actual product
Size；
To those skilled in the art, it is to be appreciated that some known features and its explanation, which may be omitted, in attached drawing
's.
Technical scheme is described further with reference to the accompanying drawings and examples.
Embodiment 1
As shown in Figure 1, a kind of feature extraction and method for tracking target based on convolutional neural networks, comprise the following steps：
S1：Build simultaneously pre-training network model；
S2：According to video sequence, training network model on line；
S3：Input video sequence, calculates tracking result；
S4：The tracking result of previous frame in video sequence is assessed, positive sample result is chosen and is put into iteration in network
To update network parameter；
In specific implementation process, step S1 can be divided into following three steps and perform：
S11：Obtain video sequence used in the data set for training foreground segmentation network and target following；
S12：The network model needed for foreground segmentation is built, network model parameter is initialized；
S13：The data that Utilization prospects segmentation network data is concentrated are trained network model, until result restrains.
In specific implementation process, step S2 can be divided into following two steps and perform：
S21：The first frame in video sequence is put into network, the tracking object position provided using groundtruth
Put and network model is finely adjusted；
S22：Extract the characteristic extraction part and parameter in training network, reconstruct target positioning network, by video sequence the
One frame is put into the network to reconstruct and iterates, training network model.
In specific implementation process, step S3 can be divided into following two steps and perform：
S31：Image in selecting video sequence successively, on the basis of target location in previous frame, determines tracking target institute
Approximate range, be put into target following network, according to obtained probability distribution matrix after processing, determine in sequence
The center of current tracking target；
S32：Centered on currently definite target's center position, the confidence corresponding to each scale in multiscale space is calculated
Value, chooses scale of the scale corresponding to response maximum as tracking target, and multiscale space is updated.
In specific implementation process, S11：Obtain training set and test set.Used training set is ILSVRC
2014Object Detection Dataset data sets, which is the standard data set for target identification, is contained
Tens of thousands of photos of the different classes of object of kind more than 200, enumerate block, rotate, the different disturbing factor such as deformation.Used in test
Data set be OTB50, which is target following frequently-used data collection, contains 50 different video sequences of length.This
A little video sequences shootings are contained and quickly move, block, rotating, under illumination variation, the difference scenes such as deformation in different scenes
Tracking result.
In specific implementation process, S12：Build foreground segmentation network model.Extracted to improve convolutional neural networks
Performance during feature (foreground segmentation), this method construct one 11 layers of foreground segmentation network, and network model is carried out with
Machine initializes, its design parameter is as shown in table 1.Convolutional layer has all used the small convolution kernel of 3x3, and it is thin both to have improved network extraction
The performance of portion's feature, while the parameter of network model is also greatly reduced, accelerate the process of convolution algorithm.
The parameter of the feature extraction network of 1 11 convolutional layers of table
In specific implementation process, S13:Training network model.Training, which has been used in training set, under line amounts to 200 classes
Tens of thousands of pictures, improve the performance of network model.The problem of a large amount of training datas bring over-fitting in order to prevent, is being instructed
For example pure background image of part negative sample etc. is also added when practicing, improves the robustness of network model.
According to the groundtruth information provided in data set during training, using the center of groundtruth calibration as
Center, extracts the block of pixels in scope to be detected and is put into foreground segmentation network.To be detected piece of size is as follows：
Wherein (x0, y0) demarcating the top left co-ordinate of position by groudtruth, w is the width of bounding box, h
For the height of bounding box.Size to be detected piece whole is less than 280x280 pixel sizes.Detected and pass through convolutional layer
Afterwards, the eigenmatrix of 1 35x35 will be exported, each numerical representation method in matrix is carried to the pixel regions of former 8x8 block sizes
Take result.Label used is what is be manually set during training, we are using the region that bounding box are outlined as target area, structure
The 0-1 matrixes of 35x35 sizes are built, the numerical value corresponding to bounding box in region is set to 1, and other regions are all set to 0.Before
The loss function of scape segmentation network is shown below：
Wherein pxyFor the predicted value at (x, y) place, txyFor the true value at (x, y) place.After largely training, the network model
To gradually it restrain.
In specific implementation process, S21:Trim network model.By way of training on line, to foreground segmentation network
It is finely adjusted.The first two field picture with groundtruth is put into foreground segmentation network, as training sample iteration 50
Secondary, trim network model, improves foreground segmentation performance under special scenes.
In specific implementation process, S22:Reconstruct target positioning network, training network model parameter.Take foreground segmentation net
Network Conv4 and its preceding network structure, and retain the network model parameter after training, two layers after network model is carried out random first
Beginningization, its parameter are as shown in table 2：
The parameter of 2 one networks comprising two layers of convolutional layer of table
The input of target positioning network is identical with foreground segmentation network, is to be detected piece of 280x280 sizes.Carry out
During training, the first frame of selecting video sequence demarcates the standard two-dimensional Gaussian Profile in region as training set, groundtruth
For label.The probability distributing density function of the label can be obtained by following formula：
Wherein (μ1, μ2) it is that groudtruth marks center point coordinate, σ1=σ2=1.
Target positioning network losses function is as follows defined in training process：
Wherein G0For the standard two-dimensional Gaussian Profile in groundtruth calibration ranges, ε is number of channels, XiI-th logical
The distribution output in road, γiFor this multi-channel parametric, the parameters such as convolution nuclear parameter, bias term are contained.
To solve the problems, such as that training sample is excessively rare, random binary mask behaviour is added when carrying out convolution operation
Make：
Wherein FcExported for convolutional layer, i.e., the characteristic image of one K channel；ωcFor convolution kernel, XkFor last layer convolutional layer
Output as a result, bcFor bias term.Here we assume that input picture is a c channel image, McFor a c passage two into
Mask processed, its parameter are random initializtions.After adding random binary masking operations, make training objective positioning network on line
Training set quantity greatly increases, and improves the robustness of network.It the experiment proved that, when training on line, pass through 150 times or so
Repetitive exercise, the loss function of the network are gradually restrained, the accuracy rate highest of target positioning.
In specific implementation process, S31:Video sequence is read in, determines tracking position of object in present frame.By video sequence
Image in row is put into target positioning network, and corresponding target position will be returned to after a series of convolution operation most
Coordinate value at maximum probability.
In specific implementation process, S32:Dynamitic scales.For improve under multiple dimensioned change condition target with
The order of accuarcy of track, determines to track the ruler corresponding to target in present frame method proposes a kind of dynamitic scales method
Degree.The change of scale vector of a 7 degree of freedom is preset, its parameter is as follows：
Scale_step=1.02
Scale_sigma=1.4361
On the basis of the scale size in previous frame, using Scale_step as step-length establish change of scale vector, take pair
Answer confidence value the maximum：
The scope that wherein S is demarcated by scale.Here given threshold is 0.6.When corresponding maximum letter in change of scale vector
When center value is less than threshold value, then according to Scale_sigma step-lengths renewal change of scale vector.
In specific implementation process, S4:Tracking result is assessed, with new network model.To improve the accurate of target positioning
Rate, this method is during target following to network model parameter real-time update.Selected threshold θ=Lθ=0.2, present frame
When loss function is less than threshold value, the renewal of random binary mask successive ignition, the same S22 of its specific steps are added；It is on the contrary then retain
Tracking result, is not updated network model.
As shown in Fig. 2, a large amount of training set training network models are used under online in advance, and by the first frame in video sequence
It is put into network and is iterated fine setting, improves the performance of neutral net foreground segmentation；Then net is recombinated using two layers of convolutional layer
Network, realizes the positioning to target location, and the shape of tracking target in the current frame is finally obtained after adding adaptive change of scale
State.
The same or similar label correspond to the same or similar components；
The terms describing the positional relationship in the drawings are only for illustration, it is impossible to is interpreted as the limitation to this patent；
Obviously, the above embodiment of the present invention is only intended to clearly illustrate example of the present invention, and is not pair
The restriction of embodiments of the present invention.For those of ordinary skill in the field, may be used also on the basis of the above description
To make other variations or changes in different ways.There is no necessity and possibility to exhaust all the enbodiments.It is all this
All any modification, equivalent and improvement made within the spirit and principle of invention etc., should be included in the claims in the present invention
Protection domain within.
Claims (7)
1. a kind of feature extraction and method for tracking target based on convolutional neural networks, it is characterised in that comprise the following steps：
S1：Build simultaneously pre-training network model；
S2：According to video sequence, training network model on line；
S3：Input video sequence, calculates tracking result；
S4：The tracking result of previous frame in video sequence is assessed, positive sample result is chosen and is put into network iteration with more
New network parameter.
2. feature extraction and method for tracking target according to claim 1 based on convolutional neural networks, it is characterised in that
Step S1 can be divided into following three steps and perform：
S11：Obtain video sequence used in the data set for training foreground segmentation network and target following；
S12：The network model needed for foreground segmentation is built, network model parameter is initialized；
S13：The data that Utilization prospects segmentation network data is concentrated are trained network model, until result restrains.
3. feature extraction and method for tracking target according to claim 2 based on convolutional neural networks, it is characterised in that
Training set under line has been used to be trained the network foreground segmentation module of convolutional Neural in step S11 in advance, the training
Data set used in foreground segmentation network is 2014 Object Detection Dataset of ILSVRC；The target following institute
It is Object Tracking Benchmark 50 with video sequence.
4. feature extraction and method for tracking target according to claim 2 based on convolutional neural networks, it is characterised in that
The network model built in step S12 needed for foreground segmentation is a feature extraction network for including 11 convolutional layers.
5. feature extraction and method for tracking target according to claim 3 based on convolutional neural networks, it is characterised in that
Using the bounding box in 2014 Object Detection Dataset data sets of ILSVRC as label, by image
It is put into feature extraction network and is trained, until result restrains.
6. feature extraction and method for tracking target according to claim 3 based on convolutional neural networks, it is characterised in that
Step S2 can be divided into following two steps and perform：
S21：The first frame in video sequence is put into network, the tracking object positions pair provided using groundtruth
Network model is finely adjusted；
S22：Extract the characteristic extraction part and parameter in training network, reconstruct target positioning network, by the first frame of video sequence
It is put into the network to reconstruct and iterates, training network model.
7. feature extraction and method for tracking target according to claim 6 based on convolutional neural networks, it is characterised in that
Step S3 can be divided into following two steps and perform：
S31：Image in selecting video sequence successively, on the basis of target location in previous frame, where definite tracking target
Approximate range, is put into target following network in sequence, according to obtained probability distribution matrix after processing, determines current
Track the center of target；
S32：Centered on currently definite target's center position, the confidence value corresponding to each scale in multiscale space is calculated,
Scale of the scale corresponding to response maximum as tracking target is chosen, and multiscale space is updated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711262806.8A CN108038435B (en) | 2017-12-04 | 2017-12-04 | Feature extraction and target tracking method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711262806.8A CN108038435B (en) | 2017-12-04 | 2017-12-04 | Feature extraction and target tracking method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108038435A true CN108038435A (en) | 2018-05-15 |
| CN108038435B CN108038435B (en) | 2022-01-04 |
Family
ID=62095161
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711262806.8A Active CN108038435B (en) | 2017-12-04 | 2017-12-04 | Feature extraction and target tracking method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108038435B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109064493A (en) * | 2018-08-01 | 2018-12-21 | 北京飞搜科技有限公司 | A kind of method for tracking target and device based on meta learning |
| CN109446804A (en) * | 2018-09-27 | 2019-03-08 | 桂林电子科技大学 | A kind of intrusion detection method based on Analysis On Multi-scale Features connection convolutional neural networks |
| CN109671102A (en) * | 2018-12-03 | 2019-04-23 | 华中科技大学 | A kind of composite type method for tracking target based on depth characteristic fusion convolutional neural networks |
| CN109829936A (en) * | 2019-01-29 | 2019-05-31 | 青岛海信网络科技股份有限公司 | A kind of method and apparatus of target tracking |
| CN109840477A (en) * | 2019-01-04 | 2019-06-04 | 苏州飞搜科技有限公司 | Face identification method and device are blocked based on eigentransformation |
| CN110827292A (en) * | 2019-10-23 | 2020-02-21 | 中科智云科技有限公司 | Video instance segmentation method and device based on convolutional neural network |
| CN111325338A (en) * | 2020-02-12 | 2020-06-23 | 暗物智能科技(广州)有限公司 | Neural network structure evaluation model construction and neural network structure search method |
| CN112037173A (en) * | 2020-08-04 | 2020-12-04 | 湖南自兴智慧医疗科技有限公司 | Chromosome detection method and device and electronic equipment |
| CN112489056A (en) * | 2020-12-01 | 2021-03-12 | 叠境数字科技（上海）有限公司 | Real-time human body matting method suitable for mobile terminal |
| CN116468984A (en) * | 2023-03-10 | 2023-07-21 | 衡阳师范学院 | Construction method of movable object detection model, detection model and detection method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106204642A (en) * | 2016-06-29 | 2016-12-07 | 四川大学 | A kind of cell tracker method based on deep neural network |
| CN106709936A (en) * | 2016-12-14 | 2017-05-24 | 北京工业大学 | Single target tracking method based on convolution neural network |
-
2017
- 2017-12-04 CN CN201711262806.8A patent/CN108038435B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106204642A (en) * | 2016-06-29 | 2016-12-07 | 四川大学 | A kind of cell tracker method based on deep neural network |
| CN106709936A (en) * | 2016-12-14 | 2017-05-24 | 北京工业大学 | Single target tracking method based on convolution neural network |
Non-Patent Citations (3)
| Title |
|---|
| LIJUN WANG ET AL: "STCT: Sequentially Training Convolutional Networks for Visual Tracking", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| NAIYAN WANG ET AL: "Transferring Rich Feature Hierarchies for Robust Visual Tracking", 《ARXIV:1501.04587V2 [CS.CV]》 * |
| 袁大龙 等: "协同运动状态估计的多目标跟踪算法", 《计算机科学》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109064493A (en) * | 2018-08-01 | 2018-12-21 | 北京飞搜科技有限公司 | A kind of method for tracking target and device based on meta learning |
| CN109446804A (en) * | 2018-09-27 | 2019-03-08 | 桂林电子科技大学 | A kind of intrusion detection method based on Analysis On Multi-scale Features connection convolutional neural networks |
| CN109446804B (en) * | 2018-09-27 | 2022-02-01 | 桂林电子科技大学 | Intrusion detection method based on multi-scale feature connection convolutional neural network |
| CN109671102A (en) * | 2018-12-03 | 2019-04-23 | 华中科技大学 | A kind of composite type method for tracking target based on depth characteristic fusion convolutional neural networks |
| CN109671102B (en) * | 2018-12-03 | 2021-02-05 | 华中科技大学 | Comprehensive target tracking method based on depth feature fusion convolutional neural network |
| CN109840477A (en) * | 2019-01-04 | 2019-06-04 | 苏州飞搜科技有限公司 | Face identification method and device are blocked based on eigentransformation |
| CN109829936B (en) * | 2019-01-29 | 2021-12-24 | 青岛海信网络科技股份有限公司 | Target tracking method and device |
| CN109829936A (en) * | 2019-01-29 | 2019-05-31 | 青岛海信网络科技股份有限公司 | A kind of method and apparatus of target tracking |
| CN110827292A (en) * | 2019-10-23 | 2020-02-21 | 中科智云科技有限公司 | Video instance segmentation method and device based on convolutional neural network |
| CN111325338A (en) * | 2020-02-12 | 2020-06-23 | 暗物智能科技(广州)有限公司 | Neural network structure evaluation model construction and neural network structure search method |
| CN112037173A (en) * | 2020-08-04 | 2020-12-04 | 湖南自兴智慧医疗科技有限公司 | Chromosome detection method and device and electronic equipment |
| CN112037173B (en) * | 2020-08-04 | 2024-04-05 | 湖南自兴智慧医疗科技有限公司 | Chromosome detection method and device and electronic equipment |
| CN112489056A (en) * | 2020-12-01 | 2021-03-12 | 叠境数字科技（上海）有限公司 | Real-time human body matting method suitable for mobile terminal |
| CN116468984A (en) * | 2023-03-10 | 2023-07-21 | 衡阳师范学院 | Construction method of movable object detection model, detection model and detection method |
| CN116468984B (en) * | 2023-03-10 | 2023-10-27 | 衡阳师范学院 | Construction method of movable object detection model, detection model and detection method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108038435B (en) | 2022-01-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108038435A (en) | A kind of feature extraction and method for tracking target based on convolutional neural networks | |
| CN110738207B (en) | Character detection method for fusing character area edge information in character image | |
| Lee et al. | Deep saliency with encoded low level distance map and high level features | |
| CN110176027A (en) | Video target tracking method, device, equipment and storage medium | |
| CN107103326B (en) | Collaborative significance detection method based on super-pixel clustering | |
| CN105389550B (en) | It is a kind of based on sparse guide and the remote sensing target detection method that significantly drives | |
| CN107833213B (en) | Weak supervision object detection method based on false-true value self-adaptive method | |
| CN110084249A (en) | The image significance detection method paid attention to based on pyramid feature | |
| CN110569738B (en) | Natural scene text detection method, equipment and medium based on densely connected network | |
| CN107633226B (en) | Human body motion tracking feature processing method | |
| CN110334762B (en) | Feature matching method based on quad tree combined with ORB and SIFT | |
| CN110175613A (en) | Street view image semantic segmentation method based on Analysis On Multi-scale Features and codec models | |
| CN106570874B (en) | Image marking method combining image local constraint and object global constraint | |
| CN111008618B (en) | Self-attention deep learning end-to-end pedestrian re-identification method | |
| CN109101981B (en) | Loop detection method based on global image stripe code in streetscape scene | |
| CN106778852A (en) | A kind of picture material recognition methods for correcting erroneous judgement | |
| CN105868706A (en) | Method for identifying 3D model based on sparse coding | |
| CN113223042B (en) | Intelligent acquisition method and equipment for remote sensing image deep learning sample | |
| CN105046714A (en) | Unsupervised image segmentation method based on super pixels and target discovering mechanism | |
| CN110827312A (en) | Learning method based on cooperative visual attention neural network | |
| CN110147841A (en) | The fine grit classification method for being detected and being divided based on Weakly supervised and unsupervised component | |
| CN110852327A (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| CN114092487A (en) | Target fruit instance segmentation method and system | |
| CN104282025A (en) | Biomedical image feature extraction method | |
| Yang et al. | An improved algorithm for the detection of fastening targets based on machine vision |
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 |