CN110110686A - Based on the human motion recognition methods for losing double-current convolutional neural networks more - Google Patents
Based on the human motion recognition methods for losing double-current convolutional neural networks more Download PDFInfo
- Publication number
- CN110110686A CN110110686A CN201910400344.4A CN201910400344A CN110110686A CN 110110686 A CN110110686 A CN 110110686A CN 201910400344 A CN201910400344 A CN 201910400344A CN 110110686 A CN110110686 A CN 110110686A
- Authority
- CN
- China
- Prior art keywords
- network
- neural networks
- convolutional neural
- double
- action recognition
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 45
- 238000000034 method Methods 0.000 title claims abstract description 19
- 230000009471 action Effects 0.000 claims abstract description 71
- 239000012530 fluid Substances 0.000 claims abstract description 13
- 238000012549 training Methods 0.000 claims abstract description 13
- 238000011084 recovery Methods 0.000 claims abstract description 12
- 239000000284 extract Substances 0.000 claims abstract description 9
- 230000011218 segmentation Effects 0.000 claims abstract description 7
- 230000006872 improvement Effects 0.000 claims abstract description 5
- 238000005070 sampling Methods 0.000 claims description 8
- 230000004927 fusion Effects 0.000 claims description 7
- 230000007774 longterm Effects 0.000 claims description 5
- 238000012512 characterization method Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000003475 lamination Methods 0.000 claims description 3
- 230000000717 retained effect Effects 0.000 claims description 3
- 238000012360 testing method Methods 0.000 claims description 3
- 238000004422 calculation algorithm Methods 0.000 claims description 2
- 230000029087 digestion Effects 0.000 claims 1
- 238000013100 final test Methods 0.000 claims 1
- 238000005516 engineering process Methods 0.000 description 4
- 230000001537 neural effect Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses based on the human motion recognition methods for losing double-current convolutional neural networks, it belongs to action recognition technical field more, solves the problems, such as that traditional binary-flow network movement detailed information loses and can not extract space-time characteristic.The present invention is to be constituted to the improvement of timing segmentation network by losing spatial network and time network more, and from the point of view of architecture angle, the double fluid convolutional neural networks that lose are made of more three branches: action recognition, movement are restored and difference punishment.Movement, which is restored, joined recovery loss, and reservation acts the motion characteristic information that detailed information and balance are extracted.Difference punishment is classified using external appearance characteristic calculating action feature, to obtain effective space-time characteristic.It is lose the training study in a manner of end to end of double fluid convolutional neural networks more, and extract video abundant using action recognition loss, recovery loss and difference loss auxiliary movement identification module and express, it can preferably improve the accuracy rate of action recognition.
Description
Technical field
The present invention relates to computer vision and area of pattern recognition, especially relate to lose double-current convolutional Neural based on more
The human motion recognition method of network belongs to action recognition field.
Background technique
Action recognition technology is identified to the movement of the people in one section of video.With the day of internet and digital equipment
Benefit is universal, and processing and the analysis especially video actions identification of video are widely studied in computer vision direction, it can be wide
It is general to be applied to every field, for example, intelligent video monitoring, human-computer interaction and human behavior analysis etc..Due to convolutional Neural net
Network achieves huge success in picture classification task, and the action recognition technology based on video can be counted as a kind of classification
Task, therefore action identification method is no longer limited to traditional hand-made characterization method, but based on convolutional neural networks
Action identification method.But the research direction still has many challenges, for example, the movement of camera, the influence of shooting background
The accuracy of action recognition can all be impacted with the variation of light etc..
In recent years, action recognition technology focuses primarily upon the static and dynamic of fusion video in the progress of video field
State information.Due to validity of the convolutional neural networks in Computer Vision Task, naturally it is applied in action recognition
Space characteristics extract network in.However, only capturing video static state action message is not fill to complicated action recognition task
Point.So a kind of input mode of the light stream as complementation, captures the multidate information of video in time network, show to dynamic
Make the powerful validity of identification mission.The double-current convolutional network that Karen Simonyan et al. is proposed, combines spatial network
And time network, become one of the mainstream of action recognition technology, but it is confined to the input of single picture and light stream figure, not
There is the sequence problem in view of video.Timing segmentation network is proposed as Limin Wang et al. is improved, when being based on long-term
Between structural modeling thought, sparse time sampling strategy is utilized and measure of supervision based on videl stage carries out action recognition, energy
It is enough efficiently to learn entire action video.But timing segmentation network also has certain difficulty on identifying similar movement, and
And in the detailed information for being easily lost movement and space-time characteristic in video is not accounted for.Therefore the present invention is divided in timing
Improved on the basis of network, introduce recovery loss and difference loss can auxiliary movement identification module extract space-time
Feature and reservation act detailed information.
Summary of the invention
The purpose of the present invention is to solve can not be extracted existing for traditional double-current convolutional neural networks space-time characteristic and
Shortage movement details and the problem that causes accuracy of identification low.
The technical solution adopted by the present invention to solve the above technical problem is:
S1. video V in data set is equally divided into K sections of S1,S2,…,SK(K be empirical value K=3), from each subsegment with
Machine samples a frame picture and light stream figure as the inputs for losing double-current convolutional neural networks more.
S2. building loses double-current convolutional neural networks framework.
S3. the picture acquired in step S1 and light stream figure lose in double-current convolutional neural networks is input to more to instruct
Practice, so that loss function is minimum.
S4. by test sample picture and light stream figure be input to it is above-mentioned trained complete more loss double fluid convolutional Neurals
It is tested in network, then carries out double-current fusion, finally complete the human action identification based on video.
Specifically, the building double fluid convolutional neural networks that lose include the following contents more:
The double fluid convolutional neural networks that lose are the improvement to timing segmentation network, the net of spatial network and time network more
Network structure is identical (input mode is different, is picture and light stream figure respectively), the spaces for losing double-current convolutional neural networks more
The network structure of network and time network is divided into three branches: action recognition, movement are restored and difference punishment.
(1) action recognition
Action recognition branch selects the network structure based on BN-Inception, in order to simulate long-term time knot
Structure, the present invention carry out sparse sampling in entire video and assemble segment characterizations progress action recognition.
(2) movement is restored
The recovery to input data is carried out in the output of the last layer convolutional layer of action recognition branch, present invention employs
Four layers of warp lamination and four layers of jump articulamentum are restored, and poor using the recovery of Euclidean distance costing bio disturbance, in order to guarantee
Partial act detailed information can be retained in action recognition network.
(3) difference is punished
Difference punishes that branch and action recognition and movement restore branch's sharing feature coding network, it is in action recognition branch
The last layer convolutional layer after carry out difference punishment operation, present invention utilizes the feature differences between adjacent segment to carry out movement knowledge
Not, auxiliary movement identification network can extract space-time characteristic abundant.
Specifically, the training methods for losing double-current convolutional neural networks are to do pre-training using ImageNet data set more
Model, training action identification module are trained whole network on the basis of action recognition network training is completed, and are utilized
Stochastic gradient descent algorithm optimizes.
The loss function calculation formula for losing double-current convolutional neural networks as follows more:
(1) action recognition
Entire video V is equally divided into K sections of { S1,S2,…,SK, the one frame { I of stochastical sampling from each segment1,I2,…,
IKInput as network, it finally obtains video and predicts fractional function in each movement class are as follows:
R(I1,I2,…,IK)=P (h (C (I1；W),C(I2；W),…,C(IK；W))) (1)
Wherein W is parameter, function C (Ik；W it) calculates each input and passes through the class score that network is exported, k ∈ 1,2 ...,
K, function h represent the output of K segment of fusion, obtain final movement class hypothesis score, and function P is Softmax operation.
Therefore, the loss function of action recognition module are as follows:
Wherein n is movement class total number, yiBeing really is label, Hr=h (C (I1；W),C(I2；W),…,C(IK；W)), i.e. Hi
=h (Ci(I1),Ci(I2),…,Ci(IK)) indicate that the same i of K segment prediction acts the score of class.
(2) movement is restored
This module optimizes training, loss function using Euclidean distance loss are as follows:
WhereinIt is the characteristic pattern that k-th of segment is exported by restoring network, IkFor primitive character figure.
(3) difference is punished
It is defeated to last one layer of convolutional layer by action recognition network to first have to calculate adjacent two picture either light stream figure
Feature (f outk, fk+1) difference:
dk=fk+1-fk (4)
Then by feature dkCarry out action recognition, the i.e. loss function of difference punishment are as follows:
In conclusion the total losses functions for losing double-current convolutional neural networks more are as follows:
L=Lr(y,Hr)+Lg+Ld(y,Hd) (6)
In general, through the invention it is contemplated above technical scheme is compared with the prior art, have below beneficial to effect
Fruit:
(1) present invention employs the sparse samplings on entire video, to obtain long term time video presentation.
(2) present invention adds restoring to lose, reduce the loss of movement detailed information to a certain extent, and balance
The video presentation information of extraction.
(3) the invention proposes difference penalty term, subtract each other to obtain feature difference using adjacent segment feature, recycle difference
Feature carries out action recognition, and auxiliary movement identifies that network extracts space-time characteristic.
(4) present invention utilizes multiple losses to optimize action recognition network, with action recognition loss, restores loss and difference
It loses auxiliary movement identification module and extracts better space-time video expression characteristic, increase substantially the precision of action recognition.
Detailed description of the invention
Fig. 1 is more loss double fluid convolutional neural networks structural schematic diagrams that the embodiment of the present invention uses.
Fig. 2 is that timing divides network and the Structure Comparison for losing double-current convolutional neural networks provided in an embodiment of the present invention more
Figure.
Specific embodiment
The attached figures are only used for illustrative purposes and cannot be understood as limitating the patent.
Below in conjunction with drawings and examples, the present invention is further elaborated.
Fig. 1 is more loss double fluid convolutional neural networks structural schematic diagrams that the embodiment of the present invention uses.As shown in Figure 1, should
Method the following steps are included:
S1. video V in data set is equally divided into K sections of S1,S2,…,SK(K be empirical value K=3), from each subsegment with
Machine samples a frame picture and light stream figure as the inputs for losing double-current convolutional neural networks more.
S2. building loses double-current convolutional neural networks framework.
S3. the picture acquired in step S1 and light stream figure lose in double-current convolutional neural networks is input to more to instruct
Practice, so that loss function is minimum.
S4. by test sample picture and light stream figure be input to it is above-mentioned trained complete more loss double fluid convolutional Neurals
It is tested in network, then carries out double-current fusion, finally complete the human action identification based on video.
Specifically, the building double fluid convolutional neural networks that lose include the following contents more:
The double fluid convolutional neural networks that lose are the improvement to timing segmentation network, the net of spatial network and time network more
Network structure is identical (input mode is different, is picture and light stream figure respectively), the spaces for losing double-current convolutional neural networks more
The network structure of network and time network is divided into three branches: action recognition, movement are restored, difference is punished.
(1) action recognition
Action recognition branch selects the network structure based on BN-Inception, in order to simulate long-term time knot
Structure, the present invention carry out sparse sampling in entire video and assemble segment characterizations progress action recognition.
(2) movement is restored
The recovery to input data is carried out in the output of the last layer convolutional layer of action recognition branch, present invention employs
Four layers of warp lamination and four layers of jump articulamentum are restored, and poor using the recovery of Euclidean distance costing bio disturbance, in order to guarantee
Partial act detailed information can be retained in action recognition network.
(3) difference is punished
Difference punishes that branch and action recognition and movement restore branch's sharing feature coding network, it is in action recognition branch
The last layer convolutional layer after carry out difference punishment operation, present invention utilizes the feature differences between adjacent segment to carry out movement knowledge
Not, auxiliary movement identification network can extract space-time characteristic abundant.
The loss function calculation formula for losing double-current convolutional neural networks as follows more:
(1) action recognition
Entire video V is equally divided into K sections of { S1,S2,…,SK, the one frame { I of stochastical sampling from each segment1,I2,…,
IKInput as network, it finally obtains video and predicts fractional function in each movement class are as follows:
R(I1,I2,…,IK)=P (h (C (I1；W),C(I2；W),…,C(IK；W))) (1)
Wherein W is parameter, function C (Ik；W it) calculates each input and passes through the class score that network is exported, k ∈ 1,2 ...,
K, function h represent the output of K segment of fusion, obtain final movement class hypothesis score, and function P is Softmax operation.
Therefore, the loss function of action recognition module are as follows:
Wherein n is movement class total number, yiBeing really is label, Hr=h (C (I1；W),C(I2；W),…,C(IK；W)), i.e. Hi
=h (Ci(I1),Ci(I2),…,Ci(IK)) indicate that the same i of K segment prediction acts the score of class.
(2) movement is restored
This module optimizes training, loss function using Euclidean distance loss are as follows:
WhereinIt is the characteristic pattern that k-th of segment is exported by restoring network, IkFor primitive character figure.
(3) difference is punished
It is defeated to last one layer of convolutional layer by action recognition network to first have to calculate adjacent two picture either light stream figure
Feature (f outk, fk+1) difference:
dk=fk+1-fk (4)
Then by feature dkCarry out action recognition, the i.e. loss function of difference punishment are as follows:
In conclusion the total losses functions for losing double-current convolutional neural networks more are as follows:
L=Lr(y,Hr)+Lg+Ld(y,Hd) (6)
Fig. 2 is that timing divides network and the Structure Comparison for losing double-current convolutional neural networks provided in an embodiment of the present invention more
Figure.As shown in Fig. 2, Fig. 2-(a) is that timing divides network, Fig. 2-(b) be lose double-current convolutional neural networks, the present invention when
Recovery loss and difference penalty term are increased on the basis of sequence segmentation network, it is special that the video that multiple loss optimizations are extracted is utilized
Sign, robustness and accuracy with higher.
In this work, the present invention provides a kind of new methods to identify skill to complete the human action based on video
Art.Compared with the existing methods, the present invention is utilized recovery loss and difference is punished on the basis of traditional timing divides network
It penalizes item to carry out auxiliary optimization to action recognition module, realizes the video expression characteristic for extracting and there is space-time, and reduce movement
The loss of detailed information, so that action recognition precision is greatly improved.
Finally, the details of the above embodiment of the present invention is only to illustrate examples of the invention, for this field
Technical staff, any modification, improvement and replacement etc. to above-described embodiment, should be included in the protection model of the claims in the present invention
Within enclosing.
Claims (4)
1. based on the human motion recognition methods for losing double-current convolutional neural networks more, which is characterized in that the method includes with
Lower step:
S1. video V in data set is equally divided into K sections of S1,S2,…,SK(K is empirical value K=3), adopts at random from each subsegment
One frame picture of sample and light stream figure as the inputs for losing double-current convolutional neural networks more.
S2. building loses double-current convolutional neural networks framework.
S3. the picture acquired in step S1 and light stream figure are input to lose in double-current convolutional neural networks to be trained more, are made
It is minimum to obtain loss function.
S4. by test sample picture and light stream figure be input to it is above-mentioned trained complete more loss double fluid convolutional neural networks
In tested, then carry out double-current fusion, finally complete the human action identification based on video.
2. according to claim 1 based on the human motion recognition methods for losing double-current convolutional neural networks, feature more
It is, the detailed process of the S2 are as follows:
The double fluid convolutional neural networks that lose are the improvement to timing segmentation network, the network knot of spatial network and time network more
Structure is identical (input mode is different, is picture and light stream figure respectively), the spatial networks for losing double-current convolutional neural networks more
Be divided into three branches with the network structure of time network: action recognition, movement are restored and difference punishment.
(1) action recognition
Action recognition branch selects the network structure based on BN-Inception, in order to simulate long-term time structure, this
Invention carries out sparse sampling in entire video and assembles segment characterizations progress action recognition.
(2) movement is restored
The recovery to input data is carried out in the output of the last layer convolutional layer of action recognition branch, present invention employs four layers
Warp lamination and four layers of jump articulamentum are restored, and poor using the recovery of Euclidean distance costing bio disturbance, in order to guarantee to act
Partial act detailed information can be retained in identification network.
(3) difference is punished
Difference punishes branch and action recognition and movement recovery branch's sharing feature coding network, it action recognition branch most
Difference punishment operation is carried out after later layer convolutional layer, present invention utilizes the feature differences between adjacent segment to carry out action recognition, auxiliary
Help action recognition network that can extract space-time characteristic abundant.
3. according to claim 1 based on the human motion recognition methods for losing double-current convolutional neural networks, feature more
It is, the detailed process of the S3 are as follows:
Specifically, the training methods for losing double-current convolutional neural networks are to do pre-training mould using ImageNet data set more
Type, training action identification module, action recognition network training complete on the basis of be trained whole network, and be utilized with
Machine gradient descent algorithm optimizes.
The loss function calculation formula for losing double-current convolutional neural networks as follows more:
(1) action recognition
Entire video V is equally divided into K sections of { S1,S2,…,SK, the one frame { I of stochastical sampling from each segment1,I2,…,IKMake
For the input of network, finally obtains video and predicts fractional function in each movement class are as follows:
R(I1,I2,…,IK)=P (h (C (I1；W),C(I2；W),…,C(IK；W))) (1)
Wherein W is parameter, function C (Ik；W the class score that each input is exported by network, k ∈ 1,2 ..., K, function) are calculated
H represents the output of K segment of fusion, obtains final movement class hypothesis score, and function P is Softmax operation.
Therefore, the loss function of action recognition module are as follows:
Wherein n is movement class total number, yiBeing really is label, Hr=h (C (I1；W),C(I2；W),…,C(IK；W)),
That is Hi=h (Ci(I1),Ci(I2),…,Ci(IK)) indicate that the same i of K segment prediction acts the score of class.
(2) movement is restored
This module optimizes training, loss function using Euclidean distance loss are as follows:
WhereinIt is the characteristic pattern that k-th of segment is exported by restoring network, IkFor primitive character figure.
(3) difference is punished
It first has to calculate what adjacent two picture either light stream figure was exported by action recognition network to last one layer of convolutional layer
Feature (fk, fk+1) difference:
dk=fk+1-fk (4)
Then by feature dkCarry out action recognition, the i.e. loss function of difference punishment are as follows:
In conclusion the total losses functions for losing double-current convolutional neural networks more are as follows:
L=Lr(y,Hr)+Lg+Ld(y,Hd) (6) 。
4. according to claim 1 based on the human motion recognition methods for losing double-current convolutional neural networks, feature more
It is, more loss double fluid convolutional neural networks that training is completed are tested in the S4, and each video uses a picture either
Light stream figure carrys out the score of prediction action identification as the input of more loss double-stream digestions, finally merges spatial network and time network
The score of output as the final test scores for losing double-current convolutional neural networks more.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910400344.4A CN110110686A (en) | 2019-05-14 | 2019-05-14 | Based on the human motion recognition methods for losing double-current convolutional neural networks more |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910400344.4A CN110110686A (en) | 2019-05-14 | 2019-05-14 | Based on the human motion recognition methods for losing double-current convolutional neural networks more |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110110686A true CN110110686A (en) | 2019-08-09 |
Family
ID=67490072
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910400344.4A Pending CN110110686A (en) | 2019-05-14 | 2019-05-14 | Based on the human motion recognition methods for losing double-current convolutional neural networks more |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110110686A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598598A (en) * | 2019-08-30 | 2019-12-20 | 西安理工大学 | Double-current convolution neural network human behavior identification method based on finite sample set |
| CN110852273A (en) * | 2019-11-12 | 2020-02-28 | 重庆大学 | Behavior identification method based on reinforcement learning attention mechanism |
| CN110969191A (en) * | 2019-11-07 | 2020-04-07 | 吉林大学 | Glaucoma prevalence probability prediction method based on similarity maintenance metric learning method |
| CN111027377A (en) * | 2019-10-30 | 2020-04-17 | 杭州电子科技大学 | Double-flow neural network time sequence action positioning method |
| CN111539290A (en) * | 2020-04-16 | 2020-08-14 | 咪咕文化科技有限公司 | Video motion recognition method and device, electronic equipment and storage medium |
| CN112258381A (en) * | 2020-09-29 | 2021-01-22 | 北京达佳互联信息技术有限公司 | Model training method, image processing method, device, equipment and storage medium |
| CN113139467A (en) * | 2021-04-23 | 2021-07-20 | 西安交通大学 | Hierarchical structure-based fine-grained video action identification method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006133866A (en) * | 2004-11-02 | 2006-05-25 | Advanced Telecommunication Research Institute International | Robot device and position memory unit |
| CN108427713A (en) * | 2018-02-01 | 2018-08-21 | 宁波诺丁汉大学 | A kind of video summarization method and system for homemade video |
| CN108764128A (en) * | 2018-05-25 | 2018-11-06 | 华中科技大学 | A kind of video actions recognition methods based on sparse time slice network |
| CN108805078A (en) * | 2018-06-11 | 2018-11-13 | 山东大学 | Video pedestrian based on pedestrian's average state recognition methods and system again |
| CN109492581A (en) * | 2018-11-09 | 2019-03-19 | 中国石油大学(华东) | A kind of human motion recognition method based on TP-STG frame |
-
2019
- 2019-05-14 CN CN201910400344.4A patent/CN110110686A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006133866A (en) * | 2004-11-02 | 2006-05-25 | Advanced Telecommunication Research Institute International | Robot device and position memory unit |
| CN108427713A (en) * | 2018-02-01 | 2018-08-21 | 宁波诺丁汉大学 | A kind of video summarization method and system for homemade video |
| CN108764128A (en) * | 2018-05-25 | 2018-11-06 | 华中科技大学 | A kind of video actions recognition methods based on sparse time slice network |
| CN108805078A (en) * | 2018-06-11 | 2018-11-13 | 山东大学 | Video pedestrian based on pedestrian's average state recognition methods and system again |
| CN109492581A (en) * | 2018-11-09 | 2019-03-19 | 中国石油大学(华东) | A kind of human motion recognition method based on TP-STG frame |
Non-Patent Citations (1)
| Title |
|---|
| 张晶晶: "面向步态识别的显著前景分割", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598598A (en) * | 2019-08-30 | 2019-12-20 | 西安理工大学 | Double-current convolution neural network human behavior identification method based on finite sample set |
| CN111027377A (en) * | 2019-10-30 | 2020-04-17 | 杭州电子科技大学 | Double-flow neural network time sequence action positioning method |
| CN110969191A (en) * | 2019-11-07 | 2020-04-07 | 吉林大学 | Glaucoma prevalence probability prediction method based on similarity maintenance metric learning method |
| CN110852273A (en) * | 2019-11-12 | 2020-02-28 | 重庆大学 | Behavior identification method based on reinforcement learning attention mechanism |
| CN111539290A (en) * | 2020-04-16 | 2020-08-14 | 咪咕文化科技有限公司 | Video motion recognition method and device, electronic equipment and storage medium |
| CN111539290B (en) * | 2020-04-16 | 2023-10-20 | 咪咕文化科技有限公司 | Video motion recognition method and device, electronic equipment and storage medium |
| CN112258381A (en) * | 2020-09-29 | 2021-01-22 | 北京达佳互联信息技术有限公司 | Model training method, image processing method, device, equipment and storage medium |
| CN112258381B (en) * | 2020-09-29 | 2024-02-09 | 北京达佳互联信息技术有限公司 | Model training method, image processing method, device, equipment and storage medium |
| CN113139467A (en) * | 2021-04-23 | 2021-07-20 | 西安交通大学 | Hierarchical structure-based fine-grained video action identification method |
| CN113139467B (en) * | 2021-04-23 | 2023-04-25 | 西安交通大学 | Fine granularity video action recognition method based on hierarchical structure |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110110686A (en) | Based on the human motion recognition methods for losing double-current convolutional neural networks more | |
| CN107330362B (en) | Video classification method based on space-time attention | |
| Zhang et al. | Real-time action recognition with enhanced motion vector CNNs | |
| Simonyan et al. | Two-stream convolutional networks for action recognition in videos | |
| CN106096568B (en) | A kind of pedestrian's recognition methods again based on CNN and convolution LSTM network | |
| CN108830252A (en) | A kind of convolutional neural networks human motion recognition method of amalgamation of global space-time characteristic | |
| Biswas et al. | Structural recurrent neural network (SRNN) for group activity analysis | |
| CN109101896A (en) | A kind of video behavior recognition methods based on temporal-spatial fusion feature and attention mechanism | |
| CN110188637A (en) | A kind of Activity recognition technical method based on deep learning | |
| CN110147743A (en) | Real-time online pedestrian analysis and number system and method under a kind of complex scene | |
| CN111626171B (en) | Group behavior identification method based on video segment attention mechanism and interactive relation activity diagram modeling | |
| CN110119703A (en) | The human motion recognition method of attention mechanism and space-time diagram convolutional neural networks is merged under a kind of security protection scene | |
| CN109815785A (en) | A kind of face Emotion identification method based on double-current convolutional neural networks | |
| CN109829443A (en) | Video behavior recognition methods based on image enhancement Yu 3D convolutional neural networks | |
| CN109891897A (en) | Method for analyzing media content | |
| CN105138953B (en) | A method of action recognition in the video based on continuous more case-based learnings | |
| CN105574510A (en) | Gait identification method and device | |
| CN109214285A (en) | Detection method is fallen down based on depth convolutional neural networks and shot and long term memory network | |
| CN107025420A (en) | The method and apparatus of Human bodys' response in video | |
| CN106650694A (en) | Human face recognition method taking convolutional neural network as feature extractor | |
| CN106529477A (en) | Video human behavior recognition method based on significant trajectory and time-space evolution information | |
| CN110348364A (en) | A kind of basketball video group behavior recognition methods that Unsupervised clustering is combined with time-space domain depth network | |
| CN113239801B (en) | Cross-domain action recognition method based on multi-scale feature learning and multi-level domain alignment | |
| CN107392131A (en) | A kind of action identification method based on skeleton nodal distance | |
| Xu et al. | Scene image and human skeleton-based dual-stream human action recognition |
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 | ||
| AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20231215 |
|
| AD01 | Patent right deemed abandoned |