CN109753897A - Based on memory unit reinforcing-time-series dynamics study Activity recognition method - Google Patents
Based on memory unit reinforcing-time-series dynamics study Activity recognition method Download PDFInfo
- Publication number
- CN109753897A CN109753897A CN201811569882.8A CN201811569882A CN109753897A CN 109753897 A CN109753897 A CN 109753897A CN 201811569882 A CN201811569882 A CN 201811569882A CN 109753897 A CN109753897 A CN 109753897A
- Authority
- CN
- China
- Prior art keywords
- memory unit
- video
- feature
- follows
- recurrent neural
- 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
Abstract
The invention discloses a kind of Activity recognition method based on memory unit reinforcing-time-series dynamics study, the technical issues of the practicability is poor for solving existing Activity recognition method.Technical solution is the timing structure information of video sequence when modeling long using the recurrent neural network of fusion memory unit, each video frame of video sequence is classified as associated frame and noise frame by discretization memory unit read-write controller module, the information write-in memory unit of associated frame is ignored into noise frame information simultaneously.This method can filter a large amount of noise information in non-editing video, the recurrent neural network of fusion memory unit realizes the connection of large span sequential organization, pass through the autonomous training study of data-driven, to complicated personage's behavior it is long when sequential organization mode model, solve background technique to it is long when, the motor pattern of non-editing video it is complicated, background changes more problems, improves the robustness of personage's Activity recognition method, and has reached average 94.8% recognition accuracy.
Description
Technical field
It is the present invention relates to a kind of Activity recognition method, in particular to a kind of based on memory unit reinforcing-time-series dynamics study
Activity recognition method.
Background technique
Document " L.Wang, Y.Xiong, Z.Wang, Y.Qiao, D.Lin, X.Tang, and L.V.Gool.Temporal
Segment Networks:Towards Good Practices for Deep Action Recognition,In
Proceedings of European Conference on Computer Vision, pp.20-36,2016. " discloses one
Personage Activity recognition method of the kind based on double-current convolutional neural networks Yu temporal sequence network.This method utilizes two independent volumes
Product neural network solves Activity recognition task, wherein and space flow network extracts the appearance features of target from video frame, and when
Sequence flow network then extracts the motion feature of target from corresponding light stream field data, is exported by merging the two networks and is gone
For recognition result.Meanwhile this method propose temporal sequence network come model video sequence it is long when timing structure information, the network
By the supervised learning of sparse Temporal Sampling strategy and sequence scale, efficient effective study of entire neural network is realized, and
It is yielded good result on extensive public data collection.Document the method is more coarse to the time series modeling in video,
So that network often sequential correlation of override feature in learning process；When video sequence is longer and non-editing, the party
Unrelated noise information can be incorporated final recognition result, the accuracy rate of reduction personage's Activity recognition, while noise information by method
It is added, the training study of entire neural network can also be made to become difficult.
Summary of the invention
In order to overcome the shortcomings of existing Activity recognition method, the practicability is poor, and the present invention provides a kind of strong based on memory unit
Change-time-series dynamics study Activity recognition method.This method regards when long using the recurrent neural network modeling of fusion memory unit
The timing structure information of frequency sequence is divided each video frame of video sequence by discretization memory unit read-write controller module
Class is associated frame and noise frame, the information write-in memory unit of associated frame is ignored noise frame information simultaneously, this method can filter
Fall a large amount of noise information in non-editing video, promotes the accuracy rate of subsequent act identification.In addition, the recurrence of fusion memory unit
The connection of large span sequential organization may be implemented in neural network, is learnt by the autonomous training of data-driven, to complicated personage's row
For it is long when sequential organization mode modeled, and then solve existing Activity recognition method to it is long when, non-editing video
Motor pattern is complicated, and background changes more problems, improves the robustness of personage's Activity recognition method, and reaches average
94.8%, 71.8% recognition accuracy.
The technical solution adopted by the present invention to solve the technical problems: one kind being based on memory unit reinforcing-time-series dynamics
The Activity recognition method of habit, its main feature is that the following steps are included:
Step 1: calculating video frame IaOptic flow information, wherein the Optic flow information of each pixel is by bivector (Δ x, Δ
Y) it indicates and saves as light stream figure Im.Respective higher-dimension semantic feature is extracted using two independent thinking convolutional neural networks:
xa=CNNa(Ia；wa) (1)
xm=CNNm(Im；wm) (2)
Wherein, CNNa、CNNmApparent convolutional neural networks and movement convolutional neural networks are respectively represented, to extract video
Frame IaWith light stream figure ImHigh dimensional feature.xa、xmRespectively 2048 dimensional vectors, represent that convolutional neural networks extract it is apparent with
Motion feature.wa、wmIndicate that the inside of two convolutional neural networks can training parameter.Indicate that convolutional neural networks extract using x
High dimensional feature.
Step 2: initialization memory unit M is sky, it is expressed as M0.Assuming that when t video frame, memory unit MtIt is not sky,
It wherein include Nt> 0 element, is expressed asSo, the memory module read operation at corresponding moment is as follows:
Wherein, the mh read outtRepresent the historical information of t moment before video.
Step 3: extracting the contextual feature in short-term of video content using section type recurrent neural network.In terms of step 1
For obtained higher-dimension semantic feature x as input, feature when corresponding to t video frame is denoted as xt.The long mind of recurrence in short-term of initialization
Hidden state h through network (LSTM)0、c0It is zero, then the contextual feature in short-term of t moment calculates as follows:
Wherein, EMD () indicates long recurrent neural network in short-term, ht-1,ct-1Indicate the hidden of recurrent neural network previous moment
State.AndContextual feature in short-term as video content is used for subsequent calculating.
Step 4: for each video frame, Step 1: two, the three higher-dimension semantic feature x being calculatedt, memory unit goes through
History information mhtAnd contextual feature in short-termMemory unit controller is inputted, binaryzation memory unit write instruction is calculated
st∈ { 0,1 }, specific as follows:
at=σ (qt) (6)
st=τ (at) (7)
Wherein, vTFor the row vector parameter that can learn, Wf、Wc、WmFor the weight parameter that can learn, bsFor offset parameter.
Sigmoid function σ () is by the result q of linear weighted functiontIt normalizes between 0,1, i.e. at∈(0,1)。atIt is input to threshold restriction
Binaryzation function τ () obtain binaryzation memory unit write instruction st。
Step 5: being based on binaryzation memory unit write instruction st, update memory unit and section type recurrent neural network.
For each video frame, memory unit MtMore new strategy it is as follows:
Wherein, WwFor that can learn weight matrix, which passes through multiplying for higher-dimension semantic feature xtIt is single to be converted to memory
Element Indicating willMemory unit M is writtent-1, form new memory unit Mt.In addition, section type recurrence
The hidden state h of neural networkt, ctIt updates as follows:
Wherein,The result being calculated for formula (4).
Step 6: carrying out behavior classification using memory unit.Assuming that video overall length is T, entire video is remembered when processing terminate
Recalling unit is MT, wherein there is NTA element, then the character representation f of entire video are as follows:
Wherein, f is D dimensional vector, represents the information of behavior classification in video.This feature inputs full link sort layer and is gone
It is specific as follows for category score y:
Y=soft max (Wf) (12)
Wherein, W ∈ RC×D, the identifiable behavior classification sum of C expression.The y being calculated indicates system to each classification
Classification score, the higher expression of score are more likely to be the class behavior.Assuming that ya、ymIt respectively indicates and is apparently obtained with kinesitherapy nerve network
The score arrived, then final score yfIt is as follows:
yf=ya+ym (13)
Wherein, yfIndicate final personage's Activity recognition result.
The beneficial effects of the present invention are: video sequence when this method is long using the recurrent neural network modeling of fusion memory unit
Each video frame of video sequence is classified as by the timing structure information of column by discretization memory unit read-write controller module
The information write-in memory unit of associated frame is ignored noise frame information simultaneously by associated frame and noise frame, and this method can filter not
A large amount of noise information in editing video promotes the accuracy rate of subsequent act identification.In addition, the recurrent neural of fusion memory unit
The connection of large span sequential organization may be implemented in network, is learnt by the autonomous training of data-driven, to complicated personage's behavior
Sequential organization mode is modeled when long, so solve existing Activity recognition method to it is long when, non-editing video movement
Mode is complicated, and background changes more problems, improves the robustness of personage's Activity recognition method, and reach average 94.8%,
71.8% recognition accuracy.
It elaborates with reference to the accompanying drawings and detailed description to the present invention.
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow charts of memory unit reinforcing-time-series dynamics study Activity recognition method.
Specific embodiment
Referring to Fig.1.The present invention is based on memory unit reinforcing-time-series dynamics study Activity recognition method specific steps such as
Under:
Contain the higher-dimension of semantic information apparently and motion feature Step 1: extracting.Firstly, calculating video frame IaLight stream letter
Breath, wherein the Optic flow information of each pixel is by bivector, (Δ x, Δ y) expression simultaneously save as light stream figure Im.Then, two are utilized
A independent thinking convolutional neural networks extract respective higher-dimension semantic feature:
xa=CNNa(Ia；wa) (1)
xm=CNNm(Im；wm) (2)
Wherein CNNa、CNNmApparent convolutional neural networks and movement convolutional neural networks are respectively represented, to extract video frame
IaWith light stream figure ImHigh dimensional feature.xa、xmRespectively 2048 dimensional vectors represent the apparent and fortune that convolutional neural networks extract
Dynamic feature.wa、wmIndicate that the inside of two convolutional neural networks can training parameter.Due to apparent neural network and kinesitherapy nerve net
The subsequent operation of network is completely the same, to make label simply clear, indicates that the higher-dimension that convolutional neural networks extract is special using x
Sign.
Step 2: initialization memory unit M is sky, it is expressed as M0.Assuming that when t video frame, memory unit MtIt is not sky,
It wherein include Nt> 0 element, is expressed asSo, the memory module read operation at corresponding moment is as follows:
The mh wherein read outtThe historical information of t moment before video is represented, while the historical information affects this moment
The analysis and understanding of video content.
Step 3: extracting the contextual feature in short-term of video content using section type recurrent neural network.In terms of step 1
For obtained higher-dimension semantic feature x as input, feature when corresponding to t video frame is denoted as xt.Firstly, initialization length is passed in short-term
Return the hidden state h of neural network (LSTM)0、c0It is zero, then the contextual feature in short-term of t moment calculates as follows:
Wherein EMD () indicates long recurrent neural network in short-term, ht-1,ct-1Indicate the hidden shape of recurrent neural network previous moment
State.AndContextual feature in short-term as video content is used for subsequent calculating.
Step 4: discretization memory unit writing controller.For each video frame, step 1,2,3 height being calculated
Tie up semantic feature xt, memory unit historical information mhtAnd contextual feature in short-termMemory unit controller is inputted, is calculated
To binaryzation memory unit write instruction st∈ { 0,1 }, specific as follows:
at=σ (qt) (6)
st=τ (at) (7)
Wherein vTFor the row vector parameter that can learn, Wf、Wc、WmFor the weight parameter that can learn, bsFor offset parameter.By upper
It can be seen that, sigmoid function σ () is by the result q of linear weighted functiontIt normalizes between 0,1, i.e. at∈(0,1).Secondly, atIt is defeated
The binaryzation function τ () entered to threshold restriction obtains binaryzation memory unit write instruction st。
Step 5: being based on binaryzation memory unit write instruction st, update memory unit and section type recurrent neural network.
For each video frame, memory unit MtMore new strategy it is as follows:
Wherein WwFor that can learn weight matrix, which passes through multiplying for higher-dimension semantic feature xtIt is single to be converted to memory
Element Indicating willMemory unit M is writtent-1, form new memory unit Mt.In addition, section type recurrence
The hidden state h of neural networkt, ctIt updates as follows:
WhereinThe result being calculated for formula (4).
Step 6: carrying out behavior classification using memory unit.Assuming that video overall length is T, entire video is remembered when processing terminate
Recalling unit is MT, wherein there is NTA element, then the character representation f of entire video are as follows:
Wherein f is D dimensional vector, represents the information of behavior classification in video.Then, this feature inputs full link sort layer
Behavior category score y is obtained, specific as follows:
Y=soft max (Wf) (12)
Wherein W ∈ RC×D, the identifiable behavior classification sum of C expression.The y being calculated indicates system to each classification
Classification score, the higher expression of score are more likely to be the class behavior.Assuming that ya、ymIt respectively indicates and is apparently obtained with kinesitherapy nerve network
The score arrived, then final score yfIt is as follows:
yf=ya+ym (13)
Wherein yfIndicate final personage's Activity recognition result.
Effect of the invention is described further by following emulation experiment.
1. simulated conditions.
The present invention is to be in central processing unitXeon E5-2697A 2.6GHz CPU, video card NVIDIA K80,
In 7 operating system of memory 16G, Centos, with the emulation of PyTorch software progress.
Data used in emulation are the data in two open test data set UCF101/HMDB51, wherein video camera
Movement changes greatly, and background is complex.Experimental data includes 13320/6766 section of video altogether, can be divided into according to behavior classification
101/51 class.The wherein non-editing mostly of the video data in HMDB51 data set includes more noise.
2. emulation content.
In order to prove effectiveness of the invention, emulation experiment strengthens to memory unit proposed by the present invention and time-series dynamics
Learning method has carried out comparative experiments.Specifically, as comparison algorithm of the invention, emulation experiment has selected accuracy rate highest double
Flow network framework (TSN) and L.Sun et al. are in document " L.Sun, K.Jia, K.Chen, D.Yeung, B.Shi and
S.Savarese.Lattice Long Short-Term Memory for Human Action Recognition,In
Proceedings of IEEE Conference on Computer Vision and Pattern Recognition,
The method (Lattice-LSTM) of the long recurrent neural network in short-term of lattice is proposed in pp.2166-2175,2011. ".Three algorithms
Same parameter is set, its average AUC numerical value on UCF101/HMDB51 data set is calculated.Comparing result is as shown in table 1.
Table 1
| Method | TSN | Lattice-LSTM | OUR |
| AUC(UCF101) | 93.6% | 94.0% | 94.8% |
| AUC(HMDB51) | 66.2% | 68.5% | 71.8% |
As seen from Table 1, recognition accuracy of the invention is higher than existing Activity recognition method significantly.Specifically, algorithm TSN
Accuracy rate be lower than algorithm Lattice-LSTM and OUR, reason is that TSN algorithm does not account for the timing variations mould of video content
Formula, and Lattice-LSTM and OUR use recurrent neural network and are modeled to the timing variations mode of video, thus
Demonstrate the validity of the time-series dynamics learning method proposed by the present invention based on recurrent neural network.In addition, in HMDB51 number
According on collection, algorithm OUR is substantially better than Lattice-LSTM, passs this is because memory unit proposed by the present invention can effectively be strengthened
Return neural network to it is long when, non-editing video processing capacity.Therefore, in order to which memory unit has recurrent neural network reinforcing
Effect property, emulation experiment is on UCF101 data set by all kinds of recurrent neural network LSTM, ALSTM and VideoLSTM and this hair
Bright algorithm has carried out comparative experiments, and the results are shown in Table 2.
Table 2
| Method | LSTM | ALSTM | VideoLSTM | Ours |
| AUC | 88.3% | 77.0% | 89.2% | 91.03% |
As seen from Table 2, the result that the present invention merges is higher than all kinds of recurrent neural network result accuracys rate, and reason exists
In memory unit intensifying method of the invention can effectively extract the effective information in video, and then model the timing in video
Changing pattern.In contrast, influence of the simple recurrent neural network method vulnerable to noise, therefore accuracy rate is reduced instead.
Therefore, effectiveness of the invention can be verified by the above emulation experiment.
Claims (1)
1. a kind of based on memory unit reinforcing-time-series dynamics study Activity recognition method, it is characterised in that the following steps are included:
Step 1: calculating video frame IaOptic flow information, wherein the Optic flow information of each pixel is by bivector (Δ x, Δ y) table
Show and saves as light stream figure Im；Respective higher-dimension semantic feature is extracted using two independent thinking convolutional neural networks:
xa=CNNa(Ia；wa) (1)
xm=CNNm(Im；wm) (2)
Wherein, CNNa、CNNmApparent convolutional neural networks and movement convolutional neural networks are respectively represented, to extract video frame Ia
With light stream figure ImHigh dimensional feature；xa、xmRespectively 2048 dimensional vectors represent the apparent and movement that convolutional neural networks extract
Feature；wa、wmIndicate that the inside of two convolutional neural networks can training parameter；The height that convolutional neural networks extract is indicated using x
Dimensional feature；
Step 2: initialization memory unit M is sky, it is expressed as M0；Assuming that when t video frame, memory unit MtIt is not sky, wherein
Include Nt> 0 element, is expressed as m1,m2,...mNt；So, the memory module read operation at corresponding moment is as follows:
Wherein, the mh read outtRepresent the historical information of t moment before video；
Step 3: extracting the contextual feature in short-term of video content using section type recurrent neural network；It is calculated with step 1
For the higher-dimension semantic feature x arrived as input, feature when corresponding to t video frame is denoted as xt；The long recurrent neural net in short-term of initialization
The hidden state h of network (LSTM)0、c0It is zero, then the contextual feature in short-term of t moment calculates as follows:
Wherein, EMD () indicates long recurrent neural network in short-term, ht-1,ct-1Indicate the hidden state of recurrent neural network previous moment；
AndContextual feature in short-term as video content is used for subsequent calculating；
Step 4: for each video frame, Step 1: two, the three higher-dimension semantic feature x being calculatedt, memory unit history letter
Cease mhtAnd contextual feature in short-termMemory unit controller is inputted, binaryzation memory unit write instruction s is calculatedt∈
{ 0,1 }, specific as follows:
at=σ (qt) (6)
st=τ (at) (7)
Wherein, vTFor the row vector parameter that can learn, Wf、Wc、WmFor the weight parameter that can learn, bsFor offset parameter；sigmoid
Function σ () is by the result q of linear weighted functiontIt normalizes between 0,1, i.e. at∈(0,1)；atIt is input to the binaryzation of threshold restriction
Function τ () obtains binaryzation memory unit write instruction st；
Step 5: being based on binaryzation memory unit write instruction st, update memory unit and section type recurrent neural network；For
Each video frame, memory unit MtMore new strategy it is as follows:
Wherein, WwFor that can learn weight matrix, which passes through multiplying for higher-dimension semantic feature xtBe converted to memory unit member
Element Indicating willMemory unit M is writtent-1, form new memory unit Mt；In addition, section type recurrent neural
The hidden state h of networkt, ctIt updates as follows:
Wherein,The result being calculated for formula (4)；
Step 6: carrying out behavior classification using memory unit；Assuming that video overall length is T, entire video is remembered single when processing terminate
Member is MT, wherein there is NTA element, then the character representation f of entire video are as follows:
Wherein, f is D dimensional vector, represents the information of behavior classification in video；This feature inputs full link sort layer and obtains behavior class
Other score y, specific as follows:
Y=softmax (Wf) (12)
Wherein, W ∈ RC×D, the identifiable behavior classification sum of C expression；The y being calculated indicates classification of the system to each classification
Score, the higher expression of score are more likely to be the class behavior；Assuming that ya、ymIt respectively indicates and is apparently obtained with kinesitherapy nerve network
Score, then final score yfIt is as follows:
yf=ya+ym (13)
Wherein, yfIndicate final personage's Activity recognition result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811569882.8A CN109753897B (en) | 2018-12-21 | 2018-12-21 | Behavior recognition method based on memory cell reinforcement-time sequence dynamic learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811569882.8A CN109753897B (en) | 2018-12-21 | 2018-12-21 | Behavior recognition method based on memory cell reinforcement-time sequence dynamic learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109753897A true CN109753897A (en) | 2019-05-14 |
| CN109753897B CN109753897B (en) | 2022-05-27 |
Family
ID=66403877
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811569882.8A Active CN109753897B (en) | 2018-12-21 | 2018-12-21 | Behavior recognition method based on memory cell reinforcement-time sequence dynamic learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109753897B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135345A (en) * | 2019-05-15 | 2019-08-16 | 武汉纵横智慧城市股份有限公司 | Activity recognition method, apparatus, equipment and storage medium based on deep learning |
| CN110348567A (en) * | 2019-07-15 | 2019-10-18 | 北京大学深圳研究生院 | A kind of memory network method integrated based on automatic addressing and recurrence information |
| CN110852273A (en) * | 2019-11-12 | 2020-02-28 | 重庆大学 | Behavior identification method based on reinforcement learning attention mechanism |
| CN111401149A (en) * | 2020-02-27 | 2020-07-10 | 西北工业大学 | Lightweight video behavior identification method based on long-short-term time domain modeling algorithm |
| CN111639548A (en) * | 2020-05-11 | 2020-09-08 | 华南理工大学 | Door-based video context multi-modal perceptual feature optimization method |
| CN112633260A (en) * | 2021-03-08 | 2021-04-09 | 北京世纪好未来教育科技有限公司 | Video motion classification method and device, readable storage medium and equipment |
| CN112926453A (en) * | 2021-02-26 | 2021-06-08 | 电子科技大学 | Examination room cheating behavior analysis method based on motion feature enhancement and long-term time sequence modeling |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106407889A (en) * | 2016-08-26 | 2017-02-15 | 上海交通大学 | Video human body interaction motion identification method based on optical flow graph depth learning model |
| CN106845351A (en) * | 2016-05-13 | 2017-06-13 | 苏州大学 | It is a kind of for Activity recognition method of the video based on two-way length mnemon in short-term |
| CN106934352A (en) * | 2017-02-28 | 2017-07-07 | 华南理工大学 | A kind of video presentation method based on two-way fractal net work and LSTM |
| US20170255832A1 (en) * | 2016-03-02 | 2017-09-07 | Mitsubishi Electric Research Laboratories, Inc. | Method and System for Detecting Actions in Videos |
| CN107330362A (en) * | 2017-05-25 | 2017-11-07 | 北京大学 | A kind of video classification methods based on space-time notice |
| CN108681712A (en) * | 2018-05-17 | 2018-10-19 | 北京工业大学 | A kind of Basketball Match Context event recognition methods of fusion domain knowledge and multistage depth characteristic |
| CN108805080A (en) * | 2018-06-12 | 2018-11-13 | 上海交通大学 | Multi-level depth Recursive Networks group behavior recognition methods based on context |
-
2018
- 2018-12-21 CN CN201811569882.8A patent/CN109753897B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170255832A1 (en) * | 2016-03-02 | 2017-09-07 | Mitsubishi Electric Research Laboratories, Inc. | Method and System for Detecting Actions in Videos |
| CN106845351A (en) * | 2016-05-13 | 2017-06-13 | 苏州大学 | It is a kind of for Activity recognition method of the video based on two-way length mnemon in short-term |
| CN106407889A (en) * | 2016-08-26 | 2017-02-15 | 上海交通大学 | Video human body interaction motion identification method based on optical flow graph depth learning model |
| CN106934352A (en) * | 2017-02-28 | 2017-07-07 | 华南理工大学 | A kind of video presentation method based on two-way fractal net work and LSTM |
| CN107330362A (en) * | 2017-05-25 | 2017-11-07 | 北京大学 | A kind of video classification methods based on space-time notice |
| CN108681712A (en) * | 2018-05-17 | 2018-10-19 | 北京工业大学 | A kind of Basketball Match Context event recognition methods of fusion domain knowledge and multistage depth characteristic |
| CN108805080A (en) * | 2018-06-12 | 2018-11-13 | 上海交通大学 | Multi-level depth Recursive Networks group behavior recognition methods based on context |
Non-Patent Citations (4)
| Title |
|---|
| LIMIN WANG ET AL.: "Temporal Segment Networks: Towards Good Practices for Deep Action Recognition", 《ARXIV》 * |
| LIN SUN ET AL.: "Lattice Long Short-Term Memory for Human Action Recognition", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION》 * |
| LINCHAO ZHU: "Bidirectional Multirate Reconstruction for Temporal Modeling in Videos", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 谯庆伟: "融合双重时空网络流和attention机制的人体行为识别", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135345A (en) * | 2019-05-15 | 2019-08-16 | 武汉纵横智慧城市股份有限公司 | Activity recognition method, apparatus, equipment and storage medium based on deep learning |
| CN110348567A (en) * | 2019-07-15 | 2019-10-18 | 北京大学深圳研究生院 | A kind of memory network method integrated based on automatic addressing and recurrence information |
| CN110348567B (en) * | 2019-07-15 | 2022-10-25 | 北京大学深圳研究生院 | Memory network method based on automatic addressing and recursive information integration |
| CN110852273A (en) * | 2019-11-12 | 2020-02-28 | 重庆大学 | Behavior identification method based on reinforcement learning attention mechanism |
| CN111401149A (en) * | 2020-02-27 | 2020-07-10 | 西北工业大学 | Lightweight video behavior identification method based on long-short-term time domain modeling algorithm |
| CN111401149B (en) * | 2020-02-27 | 2022-05-13 | 西北工业大学 | Lightweight video behavior identification method based on long-short-term time domain modeling algorithm |
| CN111639548A (en) * | 2020-05-11 | 2020-09-08 | 华南理工大学 | Door-based video context multi-modal perceptual feature optimization method |
| CN112926453A (en) * | 2021-02-26 | 2021-06-08 | 电子科技大学 | Examination room cheating behavior analysis method based on motion feature enhancement and long-term time sequence modeling |
| CN112926453B (en) * | 2021-02-26 | 2022-08-05 | 电子科技大学 | Examination room cheating behavior analysis method based on motion feature enhancement and long-term time sequence modeling |
| CN112633260A (en) * | 2021-03-08 | 2021-04-09 | 北京世纪好未来教育科技有限公司 | Video motion classification method and device, readable storage medium and equipment |
| CN112633260B (en) * | 2021-03-08 | 2021-06-22 | 北京世纪好未来教育科技有限公司 | Video motion classification method and device, readable storage medium and equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109753897B (en) | 2022-05-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109753897A (en) | Based on memory unit reinforcing-time-series dynamics study Activity recognition method | |
| CN107679526B (en) | Human face micro-expression recognition method | |
| Qi et al. | StagNet: An attentive semantic RNN for group activity and individual action recognition | |
| CN110428428B (en) | Image semantic segmentation method, electronic equipment and readable storage medium | |
| CN107506712B (en) | Human behavior identification method based on 3D deep convolutional network | |
| Hasani et al. | Spatio-temporal facial expression recognition using convolutional neural networks and conditional random fields | |
| Cui et al. | Efficient human motion prediction using temporal convolutional generative adversarial network | |
| CN106845499A (en) | A kind of image object detection method semantic based on natural language | |
| Hu et al. | Learning activity patterns using fuzzy self-organizing neural network | |
| CN110414498B (en) | Natural scene text recognition method based on cross attention mechanism | |
| Cui | Applying gradient descent in convolutional neural networks | |
| CN112784763B (en) | Expression recognition method and system based on local and overall feature adaptive fusion | |
| CN107480726A (en) | A kind of Scene Semantics dividing method based on full convolution and shot and long term mnemon | |
| CN111125358B (en) | Text classification method based on hypergraph | |
| CN108133188A (en) | A kind of Activity recognition method based on motion history image and convolutional neural networks | |
| CN107330362A (en) | A kind of video classification methods based on space-time notice | |
| CN108509839A (en) | One kind being based on the efficient gestures detection recognition methods of region convolutional neural networks | |
| CN110334589B (en) | High-time-sequence 3D neural network action identification method based on hole convolution | |
| CN106599941A (en) | Method for identifying handwritten numbers based on convolutional neural network and support vector machine | |
| CN110399821B (en) | Customer satisfaction acquisition method based on facial expression recognition | |
| CN110378208B (en) | Behavior identification method based on deep residual error network | |
| CN113469356A (en) | Improved VGG16 network pig identity recognition method based on transfer learning | |
| CN107657233A (en) | Static sign language real-time identification method based on modified single multi-target detection device | |
| CN112464865A (en) | Facial expression recognition method based on pixel and geometric mixed features | |
| CN110490136A (en) | A kind of human body behavior prediction method of knowledge based distillation |
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 |