CN109460855A - A kind of throughput of crowded groups prediction model and method based on focus mechanism - Google Patents
A kind of throughput of crowded groups prediction model and method based on focus mechanism Download PDFInfo
- Publication number
- CN109460855A CN109460855A CN201811150902.8A CN201811150902A CN109460855A CN 109460855 A CN109460855 A CN 109460855A CN 201811150902 A CN201811150902 A CN 201811150902A CN 109460855 A CN109460855 A CN 109460855A
- Authority
- CN
- China
- Prior art keywords
- throughput
- feature
- crowded
- crowded groups
- prediction model
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Abstract
The invention discloses a kind of throughput of crowded groups prediction model and method based on focus mechanism, the model includes: continuity Characteristics study module, for flowing machine (Attentive Crowd Flow Machines using focusing group, ACFM) to the characteristic pattern Sequence Learning continuity Characteristics expression according to time-sequencing, continuity Characteristics figure is obtained；Periodic feature study module, for obtaining periodic feature figure using group stream machine ACFM is focused to the characteristic pattern Sequence Learning periodic feature expression according to time-sequencing；Fusion Module with time change guides the continuity Characteristics figure and the fusion of periodic feature figure for introducing external information, and the present invention passes through the dynamic representation that learning data changes in the time domain, to be inferred to the trend in throughput of crowded groups future.
Description
Technical field
The present invention relates to the fields such as security monitoring, urban traffic control, computer vision, more particularly to one kind based on poly-
The throughput of crowded groups prediction model and method of burnt mechanism.
Background technique
Throughput of crowded groups analysis is an important task.Since it has huge dive in many intelligent applications
Power, therefore attracted a large amount of research interest.As shown in Figure 1a, in city management, the target of whole city's throughput of crowded groups analysis is
It will be flowed in and out according to the future in the GPS signal predicted city area of current slot, and in security monitoring, such as Fig. 1 b institute
Show, the purpose of crowd's flow analysis is quantity and the position by predicting them using their current positions in video.
Although the region scale that the throughput of crowded groups of different field is analyzed is widely different, it is aobvious that their key problem is all how to excavate
The spatial information of work, and temporal correlation is constructed to obtain more accurate prediction.
In academic research, scholar proposes a large amount of work for space-time modeling, but in complex scene still
There are the challenges of some restricted population flow points analysis performance.Firstly, group's flow data can vary widely in time series,
This dynamic change is captured to be very important；Secondly, some periodic rules (for example, traffic flow due to rush hour or
Influence before vacation changes suddenly) variation that group is flowed can be greatly influenced, which increase learn group from data
The difficulty of flow characteristics.In addition, in different environments (for example, predicting number under a specific monitoring scene or making
With GPS data to citywide taxi volume forecasting), the prediction of group's stream usually requires the neural network of building differentiation.
Summary of the invention
In order to overcome the deficiencies of the above existing technologies, purpose of the present invention is to provide a kind of groups based on focus mechanism
Body flux prediction model and method, with the dynamic representation changed in the time domain by learning data, to be inferred to throughput of crowded groups
Following trend.
In view of the above and other objects, the present invention proposes a kind of throughput of crowded groups prediction model based on focus mechanism, comprising:
Continuity Characteristics study module, for flowing machine ACFM to the feature graphic sequence according to time-sequencing using focusing group
Continuity Characteristics expression is practised, continuity Characteristics figure is obtained；
Periodic feature study module, for flowing machine ACFM to the feature graphic sequence according to time-sequencing using focusing group
Periodic feature expression is practised, periodic feature figure is obtained；
Fusion Module with time change guides the continuity Characteristics figure and periodically spy for introducing external information
Sign figure fusion.
Preferably, the continuity Characteristics study module is the son that time span is less than the periodic feature study module
Machine ACFM submodule flows in network, including n focusing group, and using the hidden state of its last one ACFM submodule output as
The input of one convolutional layer obtains final continuity Characteristics figure.
Preferably, the periodic feature study module includes m focusing group stream machine ACFM submodule, and it is last
Input of the hidden state of one ACFM submodule output as a convolutional layer, obtains final periodic feature figure.
Preferably, the continuity Characteristics study module list entries length for extracting continuity Characteristics is longer than extracting cycle
The periodic feature study module list entries length of property feature.
Preferably, the focusing group stream machine ACFM is including two shot and long term memory networks and for focusing Weight prediction
Convolutional layer, the shot and long term memory network is connected with the convolutional layer for focusing Weight prediction, first shot and long term memory net
Network flows feature insertion by original group and is connected to one come simulated time dependence, output hidden state and current stream feature
It rises, and is entered into the convolutional layer inferred for weight mapping, second shot and long term memory network and first shot and long term
Memory network structure is identical, using the group's stream feature weighted again as input in each time step, and recycles
Studying space-temporal characteristics is flowed so as to further group and is predicted.
Preferably, the shot and long term memory network is convolution shot and long term memory network, the hidden state that each step generates
Dimension with input dimension be consistent.
Preferably, the convolutional layer for focusing Weight prediction does not change the dimension of input feature vector, exports corresponding former special
The attention that sign schemes each pixel is tried hard to, and the attention that former characteristic pattern and convolutional layer generate is tried hard to do matrix Hadamard product, output dimension
Spend constant input as second shot and long term memory network.
In order to achieve the above objectives, the throughput of crowded groups prediction technique based on focus mechanism that the present invention also provides a kind of, including such as
Lower step:
Step S1 constructs the throughput of crowded groups prediction model based on timing information Yu spatial information feature extraction, and the model is to go through
The sequence of grey level and external information that the multiple period throughput of crowded groups of history are converted to are used as input, corresponding sequence subsequent period
The corresponding grayscale image of practical throughput of crowded groups is exported as target
Step S2 learns the parameter of the throughput of crowded groups prediction model, updates each layer of throughput of crowded groups prediction model
Parameter；
Step S3 joins the model that training obtains in the throughput of crowded groups prediction model frame constructed in step S1 and step S2
Fallout predictor of the number as throughput of crowded groups, inputs the throughput of crowded groups sequence of grey level of continuous time, predicts group's stream of subsequent period
Measure grayscale image.
Preferably, step S1 further comprises:
Step S100, building continuity Characteristics study module and periodic feature study module are predicted as the throughput of crowded groups
Two sub-networks of model, difference output continuity characteristic pattern and periodic feature figure；
Step S101 constructs the Fusion Module with time change, introduces external information, the output to each sub-network
It distributes reasonable weight and is merged.
Preferably, step S101 further comprises:
Step S101a connects continuity Characteristics figure, periodic feature figure and external information characteristic pattern in port number dimension
It connects；
The connection features figure that step S100a is obtained is inputted the module being made of two layers of full articulamentum by step S101b；
Step S101c activates the output scalar that step S101b is obtained to obtain each feature weight shared in fusion,
Two kinds of feature representations are merged according to weight proportion and fusion results dimensionality reduction is obtained into final prediction result with linear change.
Compared with prior art, a kind of throughput of crowded groups prediction model and method based on focus mechanism of the present invention passes through building
Throughput of crowded groups prediction model based on timing information Yu spatial information feature extraction, and the group is updated using back-propagation algorithm
The parameter that each layer of body flux prediction model, the model parameter that the throughput of crowded groups prediction model frame of building and training are obtained
As the fallout predictor of throughput of crowded groups, the throughput of crowded groups sequence of grey level of continuous time is inputted, prediction obtains the group of subsequent period
Flow grayscale image, to infer the trend in throughput of crowded groups future.
Detailed description of the invention
Fig. 1 a and Fig. 1 b is the legend of initial data of the present invention；
Fig. 2 is a kind of system architecture diagram of the throughput of crowded groups prediction model based on focus mechanism of the present invention；
Fig. 3 is the structural schematic diagram for focusing group in the specific embodiment of the invention and flowing machine ACFM；
Fig. 4 is the structural schematic diagram of the residual unit of characteristic extracting module in the specific embodiment of the invention；
Fig. 5 is a kind of step flow chart of the throughput of crowded groups prediction technique based on focus mechanism of the present invention.
Specific embodiment
Below by way of specific specific example and embodiments of the present invention are described with reference to the drawings, those skilled in the art can
Understand further advantage and effect of the invention easily by content disclosed in the present specification.The present invention can also pass through other differences
Specific example implemented or applied, details in this specification can also be based on different perspectives and applications, without departing substantially from
Various modifications and change are carried out under spirit of the invention.
Fig. 2 is a kind of system architecture diagram of the throughput of crowded groups prediction model based on focus mechanism of the present invention.As shown in Fig. 2,
A kind of throughput of crowded groups prediction model based on focus mechanism of the present invention includes two sub-networks and a weight fusion layer, and this two
Sub-network is respectively continuity Characteristics study module 201, periodic feature study module 202, and weight fusion layer is with the time
The Fusion Module 203 of variation.
Wherein, continuity Characteristics study module 201, for flowing machine (Attentive Crowd Flow using focusing group
Machines, ACFM) to the characteristic pattern Sequence Learning continuity Characteristics expression according to time-sequencing, obtain continuity Characteristics figure；Week
Phase property feature learning module 202, for flowing machine (ACFM) to the characteristic pattern Sequence Learning week according to time-sequencing using focusing group
Phase property feature representation obtains periodic feature figure；Fusion Module 203 with time change, for introducing external information guidance
The fusion of continuity Characteristics figure and periodic feature figure, the i.e. Fusion Module 203 with time change connect external information and each
From feature representation learn weight r, adaptively merge the different characteristic of each module, and combined characteristic pattern is fed to one
Additional convolutional layer, for group's stream prediction.That is, continuity Characteristics study module 201 and periodic feature learn mould
The input data of 202 two sub-networks of block is all the feature graphic sequence according to time-sequencing, the time of two sub- network inputs sequences
Span is different, sub-network (continuity Characteristics study module 201) the study continuity Characteristics expression of smaller time span, the time
Span less than one day (such as one hour or half an hour)；The sub-network (periodic feature study module 202) of larger time span
It is responsible for learning cycle feature representation, time span is more than or equal to one day (such as one day, one week or one month).
In the specific embodiment of the invention, continuity Characteristics study module 201 is adopted with periodic feature study module 202
With group's stream machine (ACFM) is focused to characteristic pattern Sequence Learning continuity/periodicity according to time-sequencing, first illustrate to focus below
Machine (ACFM) structure flows in group:
Focusing group stream machine ACFM is the unified neural network for being used for space-time feature learning.It can be filled
Divide the various context dependencies of ground capture group stream, for example, the consistency in space and time close the dependence of shot and long term
System.
As shown in figure 3, focusing group stream machine ACFM by two shot and long term memory networks and for focusing Weight prediction
Convolutional layer is formed, and specifically, which is two progressive convolution shot and long term memory networks
ConvLSTM network is connected with the convolutional layer for focusing Weight prediction, for the focusing Weight prediction of each step, wherein
One convolution shot and long term memory network ConvLSTM (LSTM of bottom) is embedded in by original group's stream feature and (is mentioned from CNN
Take) carry out simulated time dependence, output hidden state links together with current stream feature, and is entered into and reflects for weight
Penetrate the convolutional layer of deduction, second convolution shot and long term memory network ConvLSTM (LSTM at top) and first convolution length
The phase structure of memory network ConvLSTM is identical, but using the group's stream feature weighted again in each time step
As input, and study space-time feature is recycled, flows and predict so as to further group.Second shot and long term memory network
Hyper parameter is consistent with first shot and long term memory network, except the port number of hidden state and output state regards application scenarios complexity
Depending on, take the several times of former characteristic pattern port number.Here it should be noted that, the convolutional layer for focusing Weight prediction does not change input
The dimension of feature, the attention for exporting each pixel of corresponding former characteristic pattern are tried hard to, the attention that former characteristic pattern and convolutional layer generate
Figure does matrix Hadamard product, and output dimension is constant, the input as second shot and long term memory network.
In first convolution shot and long term memory network ConvLSTM, it is assumed that the input feature vector figure of i-th of period is Xi
∈Rw*h*c, wherein w, h, c, respectively represent width, height and the port number of characteristic pattern.Hidden stateIt can indicate
Are as follows:
WhereinIt is memory shape of first layer (a) the convolution shot and long term memory network ConvLSTM in (i-1)-th period
State.Hidden stateThen remain the dynamic modeling to previous population stream sequence.The present invention willAnd XiConnection generates intermediate become
AmountBy convolutional layer, (convolution kernel is 1 × 1, step-length 1, and convolution nuclear volume is c) to generate to pay attention to trying hard to Wi:
Wherein w indicates the parameter of convolutional layer, WiRepresent characteristic pattern XiWeight in each spatial position.The present invention is into one
Step is to XiAnd WiCharacteristic pattern after obtaining weight with matrix element multiplication operations, and remember as the second layer (a) convolution shot and long term
The input for recalling network C onvLSTM carries out feature learning, hidden stateIt can indicate are as follows:
The multiplication of matrix corresponding element is represented,Focus-sensitive content currently entered is encoded, and remembers the upper of preceding moment
Hereafter knowledge.The information of the entire group's stream sequence of the exports coding of the last one hidden state, and be used as flow graph future and drill
Change the space-time characteristic of analysis.
Compared with traditional shot and long term memory network, in convolution shot and long term memory network ConvLSTM of the invention input with
Interdepartmental connection is convolution operation, and the dependence between state and state is also to have convolution algorithm realization, is realized
Principle can be indicated by following equation:
it=σ (Wxi*Xt+Whi*Ht-1+Wci Ct-1+bi)
ft=σ (Wxf*Xt+Whf*Ht-1+Wcf Ct-1+bf)
Ct=ft Ct-1+it tanh(Wxc*Xt+Whc*Ht-1+bc)
ot=σ (Wxo*Xt+Who*Ht-1+Wco Ct+bo)
Ht=ot tanh(Ct)
Wherein * represents convolution operation, represents the multiplication of matrix corresponding element, it, ft, Ct, ot, HtIt respectively represents input gate, forget
Remember door, cell state, out gate, hidden state, and its dimension is three-dimensional.
After having described above-mentioned ACFM structure, the principle of each module of the present invention will be explained in detail below:
1, continuity Characteristics study module 201
In the specific embodiment of the invention, continuity Characteristics study module 201 includes that machine ACFM submodule flows in n focusing group
Block, input of the hidden state of the last one ACFM submodule output as a convolutional layer, obtains final continuity Characteristics
Figure.Generally, sub-network (continuity Characteristics study module 201) the list entries length for extracting continuity Characteristics is longer than extraction week
The sub-network (periodic feature study module 202) of phase property feature.
Enabling successional character representation is S={ Si| i=1 ..., n }, n is the relevant time interval length of continuity, Si
It is the connection of group the stream feature and external factor feature of i-th of period.Characteristic sequence S is sequentially input and focuses group's stream machine
ACFM learns continuity Characteristics.
Each of S element is circuited sequentially above and below as inputting and remembering as shown in Fig. 2, focusing group stream machine ACFM
Literary dependence, the hidden state of last time circulation output will input the characteristic pattern S of one layer of convolutional layer generation w*h*c dimensionf。
2, periodic feature study module 202
In the specific embodiment of the invention, continuity Characteristics study module 201 includes that machine ACFM submodule flows in m focusing group
Block, input of the hidden state of the last one ACFM submodule output as a convolutional layer, obtains final periodic feature
Figure,
It is similar with continuity Characteristics study module 201, enable P={ Pj| j=1 ..., m }, representing time interval is one day,
The periodic feature that length is m days, PjIt is the connection of QianjTian group stream feature and external factor feature.As shown in Fig. 2, P makees
To focus the input learning cycle feature representation that machine ACFM flows in group, the hidden state of last time circulation passes through one layer of convolution
Layer generates characteristic pattern Pf∈Rw*h*c。
3, the Fusion Module 203 with time change
In the specific embodiment of the invention, 203 module of Fusion Module with time change can be adaptive by different proportion
Merge the continuity Characteristics expression and periodic feature expression acquired front with answering,
In view of weather condition or special event may also will affect the weight proportion of two kinds of feature representations, used here as even
Continuous property feature Sf, periodic feature PfAnd comprehensive external factor feature EfConnection calculate weight proportion, wherein EfIt is right
Answer the result that all external factor characteristic elements are added in the period.
As shown in Fig. 2, first by Sf, Pf, EfConnect and input two layers of full articulamentum (in the specific embodiment of the invention,
One layer contains 512 neurons, and the second layer contains 1 neuron), output result finally obtains scalar r by the activation of sigmoid function
∈ [0,1], represents continuity Characteristics SfImportance；Opposite, 1-r represents periodic feature PfWeight proportion.According to weight
Ratio merges two kinds of feature representations and with linear change by fusion results dimensionality reduction to two port numbers:
Xf=T (r*Sf+(1-r)*Pf) (formula 4)
Wherein T represents linear transformation (convolution kernel as 1 × 1 convolutional layer).
Finally, group's stream of prediction subsequent time
Tanh function guarantees the range of final output between -1 to 1.
Preferably, the throughput of crowded groups prediction model based on focus mechanism of the present invention further includes characteristic extracting module, is used for
Group's stream feature and external factor feature are extracted from initial data, group's stream feature here is continuity Characteristics study mould
Input data S={ the S of block 201i| i=1 ..., n with the input data P={ P of periodic feature study module 202j| j=
1 ..., m }, external factor feature is to input the external information of the Fusion Module 203 with time change
In the specific embodiment of the invention, characteristic pattern is flowed for the group of i-th of periodCharacteristic extracting module passes through
Light-weighted ResNet network is to original density figure FiExtraction obtains, wherein light-weighted ResNet network do not include it is any under
Sampling operation (is illustrated in figure 4 classical residual unit schematic diagram, which is designed as H (x) by N number of residual unit
=F (x)+x, F (x) are that a residual error about identical x maps, and H (x) is any one ideal mapping, and it will not be described here)
Composition is stacked, the output channel number of each unit is 16 (depending on the complexities of data set), and convolution kernel size is 3 × 3.
For the external factor (such as weather, red-letter day, event) of i-th of period, one-dimensional vector is spread out into first
And normalize, the full articulamentum and reshape function then stacked with two layers converts it into high dimensional feature Ei f, wherein Ei f
Dimension and Fi fIt is consistent.
Fig. 5 is a kind of step flow chart of the throughput of crowded groups prediction technique based on focus mechanism of the present invention.As shown in figure 5,
A kind of throughput of crowded groups prediction technique based on focus mechanism of the present invention, includes the following steps:
Step S1, construct the throughput of crowded groups prediction model based on timing information Yu spatial information feature extraction, the model with
The sequence of grey level and external information (weather, festivals or holidays etc.) that the multiple period throughput of crowded groups of history are converted to are used as input,
The corresponding grayscale image of practical throughput of crowded groups of corresponding sequence subsequent period is exported as target.
Specifically, step S1 further comprises:
Step S100, building continuity Characteristics study module and periodic feature study module are predicted as the throughput of crowded groups
Two sub-networks of model, difference output continuity characteristic pattern and periodic feature figure.Described two sub-networks successively include n, m
A ACFM submodule carries out Context-dependent modeling on scale in different times, wherein as continuity Characteristics study module
Sub-network learns continuity Characteristics expression, the sub-network learning cycle feature representation as periodic feature study module；
Step S101 constructs the Fusion Module with time change, introduces external information, the output to each sub-network
It distributes reasonable weight and is merged.
Specifically, step S101 further comprises:
Step S101a connects continuity Characteristics figure, periodic feature figure and external information characteristic pattern in port number dimension
It connects.
The connection features figure that step S100a is obtained is inputted the module being made of two layers of full articulamentum by step S101b,
The middle full articulamentum of first layer contains 512 neurons (quantity is depending on characteristic pattern complexity), and the full articulamentum of the second layer contains 1 mind
Through member.
The output scalar that step S101b is obtained is activated to obtain weight r by step S101c by sigmoid function, according to
Weight proportion merges two kinds of feature representations and fusion results dimensionality reduction is obtained final prediction result with linear change.Wherein, it weighs
Weight r represents the continuity Characteristics specific gravity shared in fusion, and 1-r represents the specific gravity of periodic feature, and the implementation of fusion is adopted
Linear transformation is done with the convolutional layer that one layer of convolution kernel is 1 × 1, and result is activated to obtain final prediction by tanh function
As a result.
Step S2 learns the parameter of the throughput of crowded groups prediction model, the group is updated by back-propagation algorithm
The parameter that each layer of flux prediction model.
Step S3 joins the model that training obtains in the throughput of crowded groups prediction model frame constructed in step S1 and step S2
Number inputs the throughput of crowded groups sequence of grey level of continuous time, passes through network propagated forward together as the fallout predictor of throughput of crowded groups
Predict the throughput of crowded groups grayscale image of subsequent period.
In conclusion when a kind of throughput of crowded groups prediction model and method based on focus mechanism of the present invention is based on by building
The throughput of crowded groups prediction model of sequence information and spatial information feature extraction, and the throughput of crowded groups is updated using back-propagation algorithm
The parameter that each layer of prediction model, the model parameter that the throughput of crowded groups prediction model frame of building and training are obtained is as group
The fallout predictor of body flow, inputs the throughput of crowded groups sequence of grey level of continuous time, and prediction obtains the throughput of crowded groups ash of subsequent period
Degree figure, to infer the trend in throughput of crowded groups future.
The invention has the following beneficial effects:
One, the present invention is concerned with the analysis developed to group's stream, rather than flow graph calculates, and can push away from history flow graph
Duan Chu group current density figure, without observing any current data on flows.
Two, the focus mechanism that the present invention will contain space-time modeling is integrated into ACFM module, and have a model can
The explanatory high accuracy with prediction result.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.Any
Without departing from the spirit and scope of the present invention, modifications and changes are made to the above embodiments by field technical staff.Therefore,
The scope of the present invention, should be as listed in the claims.
Claims (10)
1. a kind of throughput of crowded groups prediction model based on focus mechanism, comprising:
Continuity Characteristics study module, for being connected using focusing group stream machine ACFM to according to the characteristic pattern Sequence Learning of time-sequencing
Continuous property feature representation, obtains continuity Characteristics figure；
Periodic feature study module, for flowing machine ACFM to the characteristic pattern Sequence Learning week according to time-sequencing using focusing group
Phase property feature representation obtains periodic feature figure；
Fusion Module with time change guides the continuity Characteristics figure and periodic feature figure for introducing external information
Fusion.
2. a kind of throughput of crowded groups prediction model based on focus mechanism as described in claim 1, it is characterised in that: described continuous
Property feature learning module be sub-network that time span is less than the periodic feature study module, including n focusing group flows
Machine ACFM submodule, and using the hidden state of its last one ACFM submodule output as the input of a convolutional layer, it obtains
Final continuity Characteristics figure.
3. a kind of throughput of crowded groups prediction model based on focus mechanism as claimed in claim 2, it is characterised in that: the period
Property feature learning module include m focusing group stream machine ACFM submodule, and by the hidden of its last one ACFM submodule output
Input of the hiding state as a convolutional layer obtains final periodic feature figure.
4. a kind of throughput of crowded groups prediction model based on focus mechanism as claimed in claim 3, it is characterised in that: extract continuous
The continuity Characteristics study module list entries length of property feature is longer than the periodic feature of extracting cycle feature
Study module list entries length.
5. a kind of throughput of crowded groups prediction model based on focus mechanism as described in claim 1, it is characterised in that: the focusing
Group stream machine ACFM includes two shot and long term memory networks and the convolutional layer for focusing Weight prediction, the shot and long term memory
Network is connected with the convolutional layer for focusing Weight prediction, and first shot and long term memory network is embedding by original group's stream feature
Enter simulated time dependence, output hidden state links together with current stream feature, and is entered into and reflects for weight
The convolutional layer of deduction is penetrated, second shot and long term memory network is identical as first shot and long term memory network structure, at each
All using the group's stream feature weighted again as input in time step, and study space-time feature is recycled, so as into one
Prediction is flowed by the group of step.
6. a kind of throughput of crowded groups prediction model based on focus mechanism as claimed in claim 5, it is characterised in that: the length
Phase memory network is convolution shot and long term memory network, and the dimension and input dimension for the hidden state that each step generates keep one
It causes.
7. a kind of throughput of crowded groups prediction model based on focus mechanism as claimed in claim 6, it is characterised in that: described to be used for
The convolutional layer for focusing Weight prediction does not change the dimension of input feature vector, exports the attention of each pixel of corresponding former characteristic pattern
The attention that figure, former characteristic pattern and convolutional layer generate is tried hard to do matrix Hadamard product, and output dimension is constant, as second shot and long term
The input of memory network.
8. a kind of throughput of crowded groups prediction technique based on focus mechanism, includes the following steps:
Step S1 constructs the throughput of crowded groups prediction model based on timing information Yu spatial information feature extraction, and the model is more with history
The sequence of grey level and external information that a period throughput of crowded groups is converted to are as input, the reality of corresponding sequence subsequent period
The corresponding grayscale image of throughput of crowded groups is exported as target
Step S2 learns the parameter of the throughput of crowded groups prediction model, updates the ginseng of each layer of the throughput of crowded groups prediction model
Number；
Step S3 makees the model parameter that training obtains in the throughput of crowded groups prediction model frame constructed in step S1 and step S2
For the fallout predictor of throughput of crowded groups, the throughput of crowded groups sequence of grey level of continuous time is inputted, predicts the throughput of crowded groups ash of subsequent period
Degree figure.
9. a kind of throughput of crowded groups prediction technique based on focus mechanism as claimed in claim 8, which is characterized in that step S1 into
One step includes:
Step S100, building continuity Characteristics study module and periodic feature study module are as the throughput of crowded groups prediction model
Two sub-networks, output continuity characteristic pattern and periodic feature figure respectively；
Step S101 constructs the Fusion Module with time change, introduces external information, distributes the output of each sub-network
Reasonable weight is simultaneously merged.
10. a kind of throughput of crowded groups prediction technique based on focus mechanism as claimed in claim 9, which is characterized in that step
S101 further comprises:
Step S101a connects continuity Characteristics figure, periodic feature figure and external information characteristic pattern in port number dimension；
The connection features figure that step S100a is obtained is inputted the module being made of two layers of full articulamentum by step S101b；
Step S101c activates the output scalar that step S101b is obtained to obtain each feature weight shared in fusion, according to
Weight proportion merges two kinds of feature representations and fusion results dimensionality reduction is obtained final prediction result with linear change.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811150902.8A CN109460855A (en) | 2018-09-29 | 2018-09-29 | A kind of throughput of crowded groups prediction model and method based on focus mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811150902.8A CN109460855A (en) | 2018-09-29 | 2018-09-29 | A kind of throughput of crowded groups prediction model and method based on focus mechanism |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109460855A true CN109460855A (en) | 2019-03-12 |
Family
ID=65607236
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811150902.8A Pending CN109460855A (en) | 2018-09-29 | 2018-09-29 | A kind of throughput of crowded groups prediction model and method based on focus mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109460855A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110610143A (en) * | 2019-08-27 | 2019-12-24 | 汇纳科技股份有限公司 | Crowd counting network method, system, medium and terminal for multi-task joint training |
| CN111079998A (en) * | 2019-12-03 | 2020-04-28 | 华东师范大学 | Flow prediction method based on long and short time sequence correlation attention mechanism model |
| CN111275971A (en) * | 2020-02-18 | 2020-06-12 | 山西交通控股集团有限公司 | Expressway traffic flow prediction method |
| CN111340543A (en) * | 2020-02-25 | 2020-06-26 | 上海大学 | Fine-grained taxi taking demand prediction method based on attention mechanism |
| CN111524348A (en) * | 2020-04-14 | 2020-08-11 | 长安大学 | Long-short term traffic flow prediction model and method |
| CN111611878A (en) * | 2020-04-30 | 2020-09-01 | 杭州电子科技大学 | Method for crowd counting and future people flow prediction based on video image |
| CN111710008A (en) * | 2020-05-29 | 2020-09-25 | 北京百度网讯科技有限公司 | People stream density generation method and device, electronic device and storage medium |
| CN111833605A (en) * | 2020-07-10 | 2020-10-27 | 北京嘀嘀无限科技发展有限公司 | Road condition prediction method, road condition prediction model training device and storage medium |
| CN112434630A (en) * | 2020-12-01 | 2021-03-02 | 深圳先进技术研究院 | Training method of continuous motion information prediction model and training device and equipment thereof |
| CN113807334A (en) * | 2021-11-22 | 2021-12-17 | 成都航空职业技术学院 | Residual error network-based multi-scale feature fusion crowd density estimation method |
| CN115150840A (en) * | 2022-05-18 | 2022-10-04 | 西安交通大学 | Mobile network flow prediction method based on deep learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103218816A (en) * | 2013-04-18 | 2013-07-24 | 中山大学 | Crowd density estimation method and pedestrian volume statistical method based on video analysis |
| CN107451552A (en) * | 2017-07-25 | 2017-12-08 | 北京联合大学 | A kind of gesture identification method based on 3D CNN and convolution LSTM |
| CN108446794A (en) * | 2018-02-25 | 2018-08-24 | 西安电子科技大学 | One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques |
-
2018
- 2018-09-29 CN CN201811150902.8A patent/CN109460855A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103218816A (en) * | 2013-04-18 | 2013-07-24 | 中山大学 | Crowd density estimation method and pedestrian volume statistical method based on video analysis |
| CN107451552A (en) * | 2017-07-25 | 2017-12-08 | 北京联合大学 | A kind of gesture identification method based on 3D CNN and convolution LSTM |
| CN108446794A (en) * | 2018-02-25 | 2018-08-24 | 西安电子科技大学 | One kind being based on multiple convolutional neural networks combination framework deep learning prediction techniques |
Non-Patent Citations (1)
| Title |
|---|
| LINGBO LIU等: "Attentive Crowd Flow Machines", 《ARXIV》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110610143B (en) * | 2019-08-27 | 2023-05-12 | 汇纳科技股份有限公司 | Crowd counting network method, system, medium and terminal for multi-task combined training |
| CN110610143A (en) * | 2019-08-27 | 2019-12-24 | 汇纳科技股份有限公司 | Crowd counting network method, system, medium and terminal for multi-task joint training |
| CN111079998A (en) * | 2019-12-03 | 2020-04-28 | 华东师范大学 | Flow prediction method based on long and short time sequence correlation attention mechanism model |
| CN111275971A (en) * | 2020-02-18 | 2020-06-12 | 山西交通控股集团有限公司 | Expressway traffic flow prediction method |
| CN111340543A (en) * | 2020-02-25 | 2020-06-26 | 上海大学 | Fine-grained taxi taking demand prediction method based on attention mechanism |
| CN111340543B (en) * | 2020-02-25 | 2024-03-22 | 上海大学 | Fine granularity taxi taking demand prediction method based on attention mechanism |
| CN111524348A (en) * | 2020-04-14 | 2020-08-11 | 长安大学 | Long-short term traffic flow prediction model and method |
| CN111611878A (en) * | 2020-04-30 | 2020-09-01 | 杭州电子科技大学 | Method for crowd counting and future people flow prediction based on video image |
| CN111710008A (en) * | 2020-05-29 | 2020-09-25 | 北京百度网讯科技有限公司 | People stream density generation method and device, electronic device and storage medium |
| CN111710008B (en) * | 2020-05-29 | 2023-07-11 | 北京百度网讯科技有限公司 | Method and device for generating people stream density, electronic equipment and storage medium |
| CN111833605A (en) * | 2020-07-10 | 2020-10-27 | 北京嘀嘀无限科技发展有限公司 | Road condition prediction method, road condition prediction model training device and storage medium |
| CN111833605B (en) * | 2020-07-10 | 2022-04-26 | 北京嘀嘀无限科技发展有限公司 | Road condition prediction method, road condition prediction model training device and storage medium |
| CN112434630A (en) * | 2020-12-01 | 2021-03-02 | 深圳先进技术研究院 | Training method of continuous motion information prediction model and training device and equipment thereof |
| CN113807334A (en) * | 2021-11-22 | 2021-12-17 | 成都航空职业技术学院 | Residual error network-based multi-scale feature fusion crowd density estimation method |
| CN115150840A (en) * | 2022-05-18 | 2022-10-04 | 西安交通大学 | Mobile network flow prediction method based on deep learning |
| CN115150840B (en) * | 2022-05-18 | 2024-03-12 | 西安交通大学 | Mobile network flow prediction method based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109460855A (en) | A kind of throughput of crowded groups prediction model and method based on focus mechanism | |
| Wang et al. | Predrnn: A recurrent neural network for spatiotemporal predictive learning | |
| Smagulova et al. | Overview of long short-term memory neural networks | |
| CN109190795A (en) | A kind of interregional Travel Demand Forecasting method and device | |
| Choi et al. | Real-time significant wave height estimation from raw ocean images based on 2D and 3D deep neural networks | |
| CN110163299A (en) | A kind of vision answering method based on bottom-up attention mechanism and memory network | |
| Yao et al. | Learning crowd behavior from real data: A residual network method for crowd simulation | |
| CN110516539A (en) | Remote sensing image building extracting method, system, storage medium and equipment based on confrontation network | |
| CN115438732B (en) | Cross-domain recommendation method for cold start user based on classified preference migration | |
| Shen et al. | Machine learning assisted urban filling | |
| He et al. | Prediction of crime rate in urban neighborhoods based on machine learning | |
| Zhao et al. | A data-driven crowd simulation model based on clustering and classification | |
| Adate et al. | A survey on deep learning methodologies of recent applications | |
| CN111259779A (en) | Video motion detection method based on central point trajectory prediction | |
| Kumar et al. | Deep Learning as a Frontier of Machine Learning: A | |
| CN110889450A (en) | Method and device for super-parameter tuning and model building | |
| CN110490320A (en) | Deep neural network structural optimization method based on forecasting mechanism and Genetic Algorithm Fusion | |
| Chen et al. | Pedestrian behavior prediction model with a convolutional LSTM encoder–decoder | |
| Cheng et al. | Analysis and forecasting of the day-to-day travel demand variations for large-scale transportation networks: a deep learning approach | |
| CN112116137A (en) | Student class dropping prediction method based on mixed deep neural network | |
| CN115719036A (en) | Space-time process simulation method and system based on stacked space-time memory unit | |
| CN115376317A (en) | Traffic flow prediction method based on dynamic graph convolution and time sequence convolution network | |
| Xiong et al. | DCAST: a spatiotemporal model with DenseNet and GRU based on attention mechanism | |
| Jia et al. | Physics-guided graph diffusion network for combining heterogeneous simulated data: An application in predicting stream water temperature | |
| CN112381148B (en) | Semi-supervised image classification method based on random regional interpolation |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190312 |