CN110210603A - Counter model construction method, method of counting and the device of crowd - Google Patents
Counter model construction method, method of counting and the device of crowd Download PDFInfo
- Publication number
- CN110210603A CN110210603A CN201910497050.8A CN201910497050A CN110210603A CN 110210603 A CN110210603 A CN 110210603A CN 201910497050 A CN201910497050 A CN 201910497050A CN 110210603 A CN110210603 A CN 110210603A
- Authority
- CN
- China
- Prior art keywords
- crowd
- image
- training
- model
- obtains
- 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
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000010276 construction Methods 0.000 title claims abstract description 39
- 238000012549 training Methods 0.000 claims abstract description 220
- 238000012545 processing Methods 0.000 claims abstract description 61
- 238000004458 analytical method Methods 0.000 claims abstract description 56
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 38
- 238000006116 polymerization reaction Methods 0.000 claims abstract description 30
- 230000008569 process Effects 0.000 claims abstract description 23
- 230000003321 amplification Effects 0.000 claims abstract description 17
- 238000003199 nucleic acid amplification method Methods 0.000 claims abstract description 17
- 238000012360 testing method Methods 0.000 claims description 67
- 238000005457 optimization Methods 0.000 claims description 50
- 238000005520 cutting process Methods 0.000 claims description 35
- 238000003860 storage Methods 0.000 claims description 13
- 230000002776 aggregation Effects 0.000 claims description 11
- 238000004220 aggregation Methods 0.000 claims description 11
- 238000013507 mapping Methods 0.000 claims description 11
- 238000009825 accumulation Methods 0.000 claims description 9
- 239000012141 concentrate Substances 0.000 claims description 5
- 239000010410 layer Substances 0.000 description 32
- 238000010586 diagram Methods 0.000 description 10
- 230000006870 function Effects 0.000 description 9
- 230000000903 blocking effect Effects 0.000 description 6
- 230000007480 spreading Effects 0.000 description 5
- 238000003892 spreading Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 230000009021 linear effect Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000010339 dilation Effects 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000009434 installation Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06M—COUNTING MECHANISMS; COUNTING OF OBJECTS NOT OTHERWISE PROVIDED FOR
- G06M11/00—Counting of objects distributed at random, e.g. on a surface
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to counter model construction method, method of counting and the devices of a kind of crowd, the processing of number of people mark is carried out to crowd's image in pre-stored data set first, obtain crowd density figure, and be combined crowd's image and crowd density figure corresponding with crowd's image, obtain target data set;Target data is concentrated again and carries out data amplification processing according to the training set that the first preset ratio divides, obtains target training set;Finally using the training crowd image and training crowd's density map in target training set, to treated that Analysis On Multi-scale Features polymerization convolutional neural networks model is trained through parameter initialization, the counter model of crowd is obtained.Using technical solution of the present invention, it is trained in model and in application, both can solve the dimensional variation problem of personage in image, can also reduces operand, improve recognition efficiency, and do not need the applicability for improving model using perspective view in training process.
Description
Technical field
The present invention relates to computer vision fields, and in particular to a kind of counter model construction method, the method for counting of crowd
And device.
Background technique
It is that number of people number therein is calculated from image or video frame that crowd, which counts, and crowd density is crowd certain
Distribution situation in time certain space.Accurately estimate crowd's number and crowd density is to measure a security system quality
One of important indicator.It is extremely important accurately to estimate that crowd's number and crowd density have on public safety and traffic control
Effect, while by the crowd density distribution of analysis megastore, the purchase that can obtain customer is liked and is inclined to and excavates latent
Commercial value.
In to crowd's image processing process, the dimensional variation problem of personage is traditional based on single-row convolutional Neural in image
The algorithm of network is difficult to solve, and generallys use the processing of the multiple row network architecture, but multiple row network is bigger than the operand of single-row network,
Task training is heavy, and recognition efficiency is lower.And it is needed on Training scene and test scene during existing model training
Using perspective view, but perspective view is difficult to obtain in practical applications, so that the applicability of model is poor.
Therefore, operand how is reduced, the applicability for improving recognition efficiency and model is that those skilled in the art urgently solve
Certainly the technical issues of.
Summary of the invention
In view of this, the purpose of the present invention is to provide counter model construction method, method of counting and the dresses of a kind of crowd
It sets, to solve the problems, such as that operand is big in the prior art, recognition efficiency is lower, the applicability of model is poor.
In order to achieve the above object, the present invention adopts the following technical scheme:
A kind of counter model construction method of crowd, comprising:
The processing of number of people mark is carried out to crowd's image in pre-stored data set, obtains crowd density figure;
Crowd's image and crowd density figure corresponding with crowd's image are combined, target data is obtained
Collection;
The target data is concentrated and carries out data amplification processing according to the training set that the first preset ratio divides, is obtained
To target training set;
Using the training crowd image and training crowd's density map in the target training set, handle through parameter initialization
Analysis On Multi-scale Features polymerization convolutional neural networks model afterwards is trained, and obtains the counter model of crowd.
Further, in method described above, described concentrate to the target data divides according to the first preset ratio
Obtained training set carries out data amplification processing, obtains target training set, comprising:
By preset cutting mode, original trained crowd's image described to each of the training set and the original respectively
The corresponding original trained crowd's density map of training crowd's image that begins is cut, and obtains at least two cutting images and at least
Two cutting density maps;
The cutting image is subjected to overturning processing, obtains flipped image, and, the cutting density map is overturn
Processing obtains overturning density map;
According to the second preset ratio of original trained crowd's image, the cutting image and the flipped image are expanded
Greatly, enlarged image is obtained, and, the cutting density map and the overturning density map are expanded, obtain expanding density map;
By the enlarged image and expansion density map storage into the training set, target training set is obtained;
Wherein, trained crowd's image includes original trained crowd's image and the enlarged image;The training
Crowd density figure includes original trained crowd's density map and the expansion density map.
Further, in method described above, the training crowd image and instruction using in the target training set
Practice crowd density figure, is trained, obtains to through parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model
To the counter model of crowd, comprising:
By in the target training set trained crowd's image and trained crowd's density map be input to through parameter
The training that first level is carried out in Analysis On Multi-scale Features polymerization convolutional neural networks model after initialization process, obtains first
Training pattern;
The optimization for being carried out first level to the model parameter of first training pattern using Euclidean distance loss formula, is obtained
To the first Optimized model;
According to preset first the number of iterations, the process of the optimization of training and the first level to the first level
It is iterated execution, obtains first object model;
Trained crowd's image and trained crowd's density map are input in the first object model and carry out
The other training of second level, obtains the second training pattern;
Formula is lost using the Euclidean distance and opposite number loss formula joins the model of second training pattern
Number carries out the optimization of second level, obtains the second Optimized model;
According to preset secondary iteration number, the process of the optimization of training and the second level to the second level
It is iterated execution, obtains the second object module;
Using second object module as the counter model of the crowd.
Further, described using second object module as the count module of the crowd in method described above
Before type, further includes:
The test crowd in the test set divided according to first preset ratio is concentrated to scheme the target data
As being input to second object module, target group's density map is obtained;
According to the corresponding test crowd of the test crowd image in target group's density map and the test set
Density map determines the test accuracy rate of model;
Judge whether the test accuracy rate is greater than default accuracy rate;
It is accordingly, described using second object module as the counter model of the crowd, comprising:
If the test accuracy rate is greater than the default accuracy rate, using second object module as the meter of the crowd
Exponential model.
Further, in method described above, the Euclidean distance loses formula are as follows:
Wherein, Θ expression parameter model;F(Xi;Θ) indicate output model;XiIndicate the trained crowd of i-th of input
Image;FiIndicate the corresponding trained crowd's density map of trained crowd's image of i-th of input;N indicates input data
Number;
The opposite number loses formula are as follows:
Wherein, FD(Xi;Θ) indicate the number of prediction;DiIndicate true value number;The number of N expression input data;Denominator is
Di+ 1 is that denominator is zero in order to prevent;The training weight of the opposite number loss is L=1.0*L (Θ)+0.1*LD(Θ)。
The present invention also provides the method for counting of crowd a kind of, comprising:
Obtain crowd's image to be counted;
Based on the counter model of the crowd constructed in advance, crowd's image to be counted is inputted, it is close to obtain crowd to be counted
Degree figure;The counter model of the crowd is constructed by the counter model construction method of above-mentioned crowd;
Using accumulation calculating method, numerical value in the crowd density figure to be counted that adds up in each pixel obtains pixel
Total value, using the pixel total value as the number in crowd's image to be counted.
The present invention also provides the counter model construction devices of crowd a kind of, comprising:
First processing module obtains people for carrying out the processing of number of people mark to crowd's image in pre-stored data set
Group's density map;
Data combination module, for crowd's image and crowd density figure corresponding with crowd's image to be carried out group
It closes, obtains target data set;
Data expand module, for the target data concentrate the training set divided according to the first preset ratio into
Row data amplification processing, obtains target training set;
Second processing module, for utilizing training crowd image and training crowd's density map in the target training set,
It is trained to through parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model, obtains the count module of crowd
Type.
Further, in device described above, the Second processing module includes: the first training unit, the first optimization
Unit, iterative processing unit, the second training unit, the second optimization unit and model determination unit;
First training unit, for by the target training set trained crowd's image and the training of human
Group's density map, which is input to, is carried out the in parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model
The other training of level-one, obtains the first training pattern;
The first optimization unit, for the model parameter using Euclidean distance loss formula to first training pattern
The optimization for carrying out first level, obtains the first Optimized model;
The iterative processing unit, for the training and institute according to preset first the number of iterations, to the first level
The process for stating the optimization of first level is iterated execution, obtains first object model;
Second training unit, it is described for trained crowd's image and trained crowd's density map to be input to
The training that second level is carried out in first object model, obtains the second training pattern;
The second optimization unit, for losing formula and opposite number loss formula to described using the Euclidean distance
The model parameter of second training pattern carries out the optimization of second level, obtains the second Optimized model;
The iterative processing unit, is also used to according to preset secondary iteration number, training to the second level and
The process of the optimization of the second level is iterated execution, obtains the second object module;
Model determination unit, for using second object module as the counter model of the crowd.
Further, in device described above, the Analysis On Multi-scale Features polymerization convolutional neural networks model includes feature
Mapping block, Analysis On Multi-scale Features aggregation module and density map regression block;
The Feature Mapping module uses the preceding X convolutional layer of VGG-16 structure, and wherein X is 4 or 6;
The convolutional layer is the convolution kernel of 3*3;
The Analysis On Multi-scale Features aggregation module uses at least two scale branches;
The density map regression block uses B layers of A column of empty convolution, and wherein A and B is positive integer.
The present invention also provides the counting devices of crowd a kind of, comprising: image collection module, density map determining module and number
Statistical module;
Described image obtains module, for obtaining crowd's image to be counted;
The density map determining module inputs the people to be counted for the counter model based on the crowd constructed in advance
Group's image, obtains crowd density figure to be counted;The counter model of the crowd passes through the counter model construction method of above-mentioned crowd
Building;
The demographics module, for utilizing accumulation calculating method, each pixel in the crowd density figure to be counted that adds up
Numerical value in point obtains pixel total value, using the pixel total value as the number in crowd's image to be counted.
Counter model construction method, method of counting and the device of crowd of the invention, first in pre-stored data set
Crowd's image carry out the processing of number of people mark, obtain crowd density figure, and by crowd's image and crowd corresponding with crowd's image
Density map is combined, and obtains target data set, wherein this programme is labeled to the number of people in crowd's image, to body
Blocking for body does not influence as a result, so not needing to use perspective view on Training scene and test scene;Again to target data set
In the training set that divides according to the first preset ratio carry out data amplification processing, obtain target training set;Finally utilize mesh
The training crowd image and training crowd's density map in training set are marked, is polymerize to through parameter initialization treated Analysis On Multi-scale Features
Convolutional neural networks model is trained, and obtains the counter model of crowd.Wherein, Analysis On Multi-scale Features polymerize convolutional neural networks mould
Type is the single-row network architecture with Analysis On Multi-scale Features polymerizable functional.Using technical solution of the present invention, it is trained in model
With in application, both can solve the dimensional variation problem of personage in image, can also reduce operand, improve recognition efficiency, and instruct
The applicability that model is improved using perspective view is not needed during practicing.
It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not
It can the limitation present invention.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is the flow chart of the counter model construction method embodiment one of crowd of the invention;
Fig. 2 is the structural schematic diagram of Analysis On Multi-scale Features aggregation module in Analysis On Multi-scale Features polymerization convolutional neural networks model;
Fig. 3 is the structural schematic diagram for stacking pond;
Fig. 4 is the flow chart of the counter model construction method embodiment two of crowd of the invention;
Fig. 5 is the flow chart of the method for counting embodiment of crowd of the invention;
Fig. 6 is the structural schematic diagram of the counter model construction device embodiment one of crowd of the invention;
Fig. 7 is the structural schematic diagram of the counter model construction device embodiment two of crowd of the invention;
Fig. 8 is the structural schematic diagram of the counting device embodiment of crowd of the invention.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, technical solution of the present invention will be carried out below
Detailed description.Obviously, described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Base
Embodiment in the present invention, those of ordinary skill in the art are obtained all without making creative work
Other embodiment belongs to the range that the present invention is protected.
Fig. 1 is the flow chart of the counter model construction method embodiment one of crowd of the invention.As shown in Figure 1, this implementation
The counter model construction method of the crowd of example can specifically include following steps:
S101, the processing of number of people mark is carried out to crowd's image in pre-stored data set, obtains crowd density figure;
The present embodiment carries out the processing of number of people mark to crowd's image in pre-stored data set first, wherein data set
For the pre-stored set comprising many crowd's images.It handles to obtain the corresponding crowd of crowd's image by following formula again close
Degree figure:
Wherein, F (x) indicates crowd density figure;M indicates the sum of head mark in crowd's image;xiIndicate pixel;δ
(x-xi) indicate i-th of head part position;Indicate Gaussian kernel;Indicate the head position of i-th of people to k arest neighbors
The average distance of people.
S102, crowd's image and crowd density figure corresponding with crowd's image are combined, obtain target data set;
Through the above steps, after obtaining the corresponding crowd density figure of crowd's image, all groups in data set are schemed
Picture and the corresponding crowd density figure of crowd's image are combined and store, and obtain new data set, i.e. target data set.
S103, the training set progress data amplification processing divided according to the first preset ratio is concentrated to target data,
Obtain target training set;
Through the above steps, after obtaining target data set, target data set is divided into instruction according to the first preset ratio
Practice collection and test set, then training set is subjected to data amplification processing, to obtain target training set.In the present embodiment, first is pre-
If ratio is preferably 7:3, i.e. the ratio of the target data set training set being divided into and test intensive data is 7:3, in addition, generally
In the case of, the data in training set that target data set is divided into are more than the data in test set.
S104, using in target training set training crowd image and training crowd's density map, at through parameter initialization
Analysis On Multi-scale Features polymerization convolutional neural networks model after reason is trained, and obtains the counter model of crowd.
Through the above steps, after obtaining target training set, using in target training set training crowd image and the training
The corresponding trained crowd's density map of crowd's image polymerize convolutional neural networks to through parameter initialization treated Analysis On Multi-scale Features
Model is trained.Wherein, parameter initialization processing is to carry out Initialize installation to the parameter of model, and the present embodiment uses
Tensorflow implementation model code, in model parameter Initialize installation, initial learning rate is preferably 10-4, attenuation coefficient is excellent
It is selected as 0.9, the rate of decay is preferably 20.After polymerizeing convolutional neural networks model training to Analysis On Multi-scale Features, crowd just can be obtained
Counter model.
It include front end, middle-end and rear end through parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model
Three parts, wherein front end is characterized mapping block, and middle-end is Analysis On Multi-scale Features aggregation module FME, and rear end is density map recurrence
Module.
In the present embodiment, Feature Mapping module using VGG-16 structure preceding 4 convolutional layers or VGG-16 structure preceding 6
A convolutional layer, wherein replacing maximum pond layer that can increase model under the premise of not introducing additional parameter using pond is stacked
Scale invariability.Convolutional layer in this module is stacked using the convolution kernel of 3*3, can be deepened network depth, be increased the non-of model
Linear property.The convolutional layer of multiple 3*3 filter filte more large-sized than one has less parameter, it is assumed that the input of convolutional layer
Characteristic pattern size with output is mutually all C, then the convolutional layer number of parameters of 3 3*3 is 27C2, the convolutional layer ginseng of a 7*7
Number is 49C2, it is possible to 3 3*3 convolution kernels are regarded as the decomposition of the convolution kernel of a 7*7, and (middle layer has nonlinear point
Solution, and play the role of implicit regularization), this replacement can make the parameter of model less.And network single layer mind
Enough through first number, the feature abstraction depth of generation is sufficiently high, improves the accuracy and the speed of service of network, possesses simultaneously
Less network parameter.The convolutional layer of Feature Mapping module in the Analysis On Multi-scale Features polymerization convolutional neural networks model of the present embodiment
It is the convolution kernel of 3*3, characteristics of image is remapped by it.In order to guarantee the non-linear of model, after each convolutional layer
A nonlinear activation function layer is all followed, the present embodiment uses amendment linear unit (Rectified linear
Units, ReLU), ReLU can accelerate the convergence of network.
In the present embodiment, Analysis On Multi-scale Features polymerization convolutional neural networks model preferably uses four layers of Analysis On Multi-scale Features polymerization mould
Block FME, Analysis On Multi-scale Features aggregation module FME use at least two scale branches, and the Analysis On Multi-scale Features in the present embodiment polymerize mould
Block preferably uses four branches, and Fig. 2 is Analysis On Multi-scale Features aggregation module in Analysis On Multi-scale Features polymerization convolutional neural networks model
Structural schematic diagram, as shown in Fig. 2, convolution kernel of each branch of Analysis On Multi-scale Features aggregation module only with 1*1 and 3*3, this mould
First branch of block is to cover Small object to retain upper one layer of characteristic dimension only with the convolution kernel of 1*1.Its excess-three
Branch uses the stacking of 3*3 convolution kernel all to imitate the receptive field of big volume core, and imitation is 3*3,5*5,7*7 convolution respectively
Core, but the convolution of one layer of 1*1 is increased before them, characteristic dimension is reduced into half.Simultaneously for simplicity, each
The port number of branch be both configured to it is equal, and in one ReLU function layer of each convolution kernel heel.Intuitively, more rulers
Degree characteristic aggregation module FME is exactly the integrated of a different size receptive field, this module can capture crowd in dense population
Multiple dimensioned appearance, be conducive to crowd counting.
In the present embodiment, density map regression block uses B layers of A column of empty convolution, and wherein A and B is positive integer.Through reality
Verifying, the density map regression block of the present embodiment is preferably the empty convolution of 5 layers of 3 column, and multilayer cavity convolution can be in not shadow
Increase receptive field in the case where ringing resolution ratio, passes through the higher crowd density figure of the available quality of empty convolution.Empty convolution
It is defined as follows:
Wherein, w (i, j) indicates filter;Y (m, n) indicates x (m, n) as input plus a filter w (i, j) point
Output when being not long and wide progress cavity convolution with M and N;R indicates spreading rate.
If spreading rate r=1, an empty convolution are exactly common convolution.Empty convolution be one of pond layer very
Good substitution, can make accuracy rate increase significantly in segmentation task.Although pond layer (such as maximum pond layer) is extensive
For maintaining the invariance and controlling over-fitting, but they also greatly reduce spatial resolution, it is meant that the space of characteristic pattern
Information is lost.Warp layer can reduce the loss of information, but additional complexity and execution delay may be not particularly suited for institute
There is something special.Empty convolution is better choice, it replaces pond and convolutional layer using sparse kernel.Characteristic pattern is being kept to differentiate
In terms of rate, compared with convolution+pond+deconvolution scheme, empty convolution has apparent advantage.Separately pass through positive research, when
The effect that spreading rate obtains when being 2 is best, therefore spreading rate r is preferably 2 in the present embodiment.
Table 1 is the framework table of Analysis On Multi-scale Features polymerization convolutional neural networks model, as shown in table 1:
Table 1
Wherein, since the value of density map is positive always, need to add ReLU activation letter after the convolutional layer of the last layer 1*1
Number reinforces the recurrence of density map.
Detailed parameter setting is listed in table 1, wherein all convolutional layers all keep original size using filling.Table
In lattice the parameter of convolutional layer be expressed as " conv (kernel size)-(filter number)-(dilation rate) " its
In, kernel size indicates convolution kernel size, and filter number indicates that port number, dilation rate indicate spreading rate;
Stacked-pooling is to stack pond, and the parameter for stacking pond is expressed as kernel collection.
The storehouse that pond is pond layer is stacked, in addition to a pond kernel, its pond operation is in the Feature Mapping to down-sampling
It calculates, wherein intermediate Feature Mapping Continuous plus formula are as follows:
Wherein, arrow s is to represent downward sample rate s;ρ indicates common maximum pond layer;k′iIndicate core size;s′iIt indicates
Step-length;Y′0=X is input feature vector mapping;Core size k 'iCorresponding to kiSome transformation;Step-length s 'I=1=s, s 'I > 1=1.
Stack the calculation formula of the output connection intermediate features mapping in pond are as follows:
Fig. 3 is the structural schematic diagram for stacking pond, as shown in figure 3, stacking the kernel collection K={ 2,2,3 }, step-length S=in pond
{ 2,1,1 } are shown by experiment, and good effect is played in the stacking pond configuration in Fig. 3 in the present embodiment model.Wherein, M, N
It respectively indicates long and wide;Map pool indicates maximum pond;2 X 2,3 X 3 are respectively the kernel size stacked in pond, stride
Indicate step-length;Channel average indicates channel mean.
The counter model construction method of the crowd of the present embodiment first carries out crowd's image in pre-stored data set
The processing of number of people mark, obtains crowd density figure, and crowd's image and crowd density figure corresponding with crowd's image are combined,
Obtain target data set;Target data is concentrated again and is carried out at data amplification according to the training set that the first preset ratio divides
Reason, obtains target training set;Finally using the training crowd image and training crowd's density map in target training set, to through parameter
Analysis On Multi-scale Features polymerization convolutional neural networks model after initialization process is trained, and obtains the counter model of crowd.Wherein,
The present embodiment is labeled to the number of people in crowd's image, does not influence blocking for body as a result, so not needing to instruct again
Practice and use perspective view on scene and test scene, improves the applicability of model.Analysis On Multi-scale Features polymerize convolutional neural networks mould
Type is the single-row network architecture with Analysis On Multi-scale Features polymerizable functional, is trained in model and in application, both can solve image
The dimensional variation problem of middle personage can also reduce operand, improve recognition efficiency.
Fig. 4 is the flow chart of the counter model construction method embodiment two of crowd of the invention, as shown in figure 4, this implementation
On the basis of the counter model construction method of the crowd of example is the embodiment described in Fig. 1, further in further detail to the present invention
Technical solution be described.
As shown in figure 4, the counter model construction method of the crowd of the present embodiment can specifically include following steps:
S201, the processing of number of people mark is carried out to crowd's image in pre-stored data set, obtains crowd density figure;
The implementation procedure of the step is identical as the implementation procedure of S101 shown in FIG. 1, and details are not described herein again.
S202, crowd's image and crowd density figure corresponding with crowd's image are combined, obtain target data set;
The implementation procedure of the step is identical as the implementation procedure of S102 shown in FIG. 1, and details are not described herein again.
S203, by preset cutting mode, the instruction that divides according to the first preset ratio is concentrated to target data respectively
Practice the original trained crowd's image of each of concentration and the corresponding original trained crowd's density map of original trained crowd's image is cut out
It cuts, obtains at least two cutting images and at least two and cut density map;
Through the above steps, after obtaining target data set, target data set is divided according to the first preset ratio,
Specific division mode is in the above-described embodiments it is stated that no longer illustrate herein.Training set is obtained after division, is wrapped in training set
Original trained crowd's image and the corresponding original trained crowd's density map of original crowd's image are included, by preset cutting mode, is divided
It is other that each original trained crowd's image and original trained crowd's density map are cut, obtain at least two cutting images and extremely
Few two cuttings density maps, cut image with cutting density map be it is corresponding, how original trained crowd's image is cut out
It cuts, with regard to how to be cut to the corresponding original trained crowd's density map of original trained crowd's image.In the present embodiment, presets and cut out
The mode of cutting is preferably nine grids cutting method, i.e., original trained crowd's image and original trained crowd's density map is cut into nine palaces
The appearance of lattice, each original trained crowd's image cut out 9 cutting images, and each original trained crowd's density map is also cut out
Cut 9 cutting density maps.
S204, cutting image is subjected to overturning processing, obtains flipped image, and, density map will be cut and carried out at overturning
Reason obtains overturning density map;
Through the above steps, obtain after cutting image and cutting the corresponding cutting density map of image, will cutting image into
Row overturning processing, obtains flipped image;Density map will be cut and carry out overturning processing, obtain overturning density map.
S205, according to the second preset ratio of original trained crowd's image, image will be cut and flipped image expands, obtained
Enlarged image, and, density map will be cut and overturning density map expands, obtain expanding density map;
Through the above steps, it obtains after cutting image, cutting density map, flipped image and overturning density map, according to original
Second preset ratio of beginning training crowd's image, will cut image and flipped image expands, and obtain enlarged image;According to original instruction
The second preset ratio for practicing crowd's image, will cut density map and overturning density map expands, and obtain expanding density map.The present embodiment
In, the second preset ratio of original trained crowd's image is preferably the 90% of original trained crowd's image, i.e., will cut image, cut out
It cuts density map, flipped image and overturning density map and expands as the 90% of original trained crowd's image.
S206, it will be enlarged by image and expand density map storage into training set, obtain target training set;
Through the above steps, after obtaining enlarged image and the corresponding expansion density map of enlarged image, will be enlarged by image and
Expand density map storage into training set, obtains new training set, i.e. target training set.Wherein, the training in target training set
Crowd's image includes original trained crowd's image and enlarged image;Training crowd's density map in target training set includes original instruction
Practice crowd density figure and expands density map.
S207, by target training set training crowd image and training crowd's density map be input to through at parameter initialization
The training that first level is carried out in Analysis On Multi-scale Features polymerization convolutional neural networks model after reason, obtains the first training pattern;
Through the above steps, after obtaining target training set, by the training crowd image and training of human in target training set
Group's density map is input to through in parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model, to the model into
The training of row first level, to obtain the first training pattern.Wherein, training is that a Zhang Xunlian is selected from training set every time
Crowd's image trained crowd's density map corresponding with training crowd's image is input in model.
S208, the optimization for carrying out first level to the model parameter of the first training pattern using Euclidean distance loss formula,
Obtain the first Optimized model;
Through the above steps, after obtaining the first training pattern, using Euclidean distance loss formula to the first training pattern
Model parameter carries out the optimization of first level, to obtain the first Optimized model.
Euclidean distance loses formula are as follows:
Wherein, Θ expression parameter model;F(Xi;Θ) indicate output model;XiIndicate the training crowd figure of i-th of input
Picture;FiIndicate the corresponding trained crowd's density map of training crowd's image of i-th of input;The number of N expression input data.
S209, according to preset first the number of iterations, the process of the optimization of training and first level to first level into
Row iteration executes, and obtains first object model;
Through the above steps, after having carried out the training of first level and the optimization of first level, according to preset first
The process of the number of iterations, the optimization of training and first level to first level is iterated execution, to obtain first object
Model.Preset first the number of iterations is preferably 200,000 times in the present embodiment.
Iteration refers to the activity of repetition feedback procedure, and purpose is typically to approaching required target or result.Each time
Primary " iteration " is known as to the repetition of process, and the result that iteration obtains each time can be as the initial value of next iteration.It is right
The subprogram for needing to execute repeatedly in computer specific program, is once repeated, that is, repeats the circulation in program, until
Until meeting certain condition, also known as iteration.It is the excellent of the constantly training of progress first level and first level in the present embodiment
Change, until the number of iteration execution meets preset first the number of iterations.
S210, training crowd's image and training crowd's density map are input in first object model and carry out second level
Training, obtains the second training pattern;
Through the above steps, after obtaining first object model, training crowd's image and training crowd's density map are input to
In first object model, the training of second level is carried out to the model, to obtain the second training pattern.Wherein, training every time
It is that the trained crowd's image of selection one trained crowd's density map corresponding with training crowd's image is defeated from training set
Enter into model.
S211, lost using Euclidean distance formula and opposite number loss formula to the model parameter of the second training pattern into
The optimization of row second level obtains the second Optimized model;
Through the above steps, after obtaining the second training pattern, formula is lost using Euclidean distance and opposite number loses public affairs
Formula carries out the optimization of second level to the model parameter of the second training pattern jointly, obtains the second Optimized model.Utilize counterpart
Number loss formula model parameter is optimized, can focusing study predict the biggish sample of error, it is very dilute in absolute number
Under thin scene, by losing in a network using opposite number, accuracy rate can be made to be obviously improved.It is not only adopted in the present embodiment
Model parameter is optimized with Euclidean distance loss formula, also model parameter is carried out using opposite number loss formula excellent
Change, can be improved the precision of prediction under the sparse scene of absolute number.
Opposite number loses formula are as follows:
Wherein, FD(Xi;Θ) indicate the number of prediction;DiIndicate true value number;The number of N expression input data;Denominator is
Di+ 1 is that denominator is zero in order to prevent;The training weight of opposite number loss is L=1.0*L (Θ)+0.1*LD(Θ)。
S212, according to preset secondary iteration number, the process of the optimization of training and second level to second level into
Row iteration executes, and obtains the second object module;
Through the above steps, after having carried out the training of second level and the optimization of second level, according to preset second
The process of the number of iterations, the optimization of training and second level to second level is iterated execution, to obtain the second target
Model.Preset secondary iteration number is preferably 100,000 times in the present embodiment.
In the present embodiment, in addition to losing formula optimization iteration, retraining and use using first training and using Euclidean distance
Euclidean distance loses formula and opposite number loses formula optimization iteration, can also directly adopt training and be damaged using Euclidean distance
It loses formula and opposite number loses formula optimization iteration, but the number of iterations is preferably 300,000 times or more, can be only achieved mould in this way
The convergence of type.
S213, concentrate test crowd's image in the test set divided according to the first preset ratio defeated target data
Enter to the second object module, obtains target group's density map;
Through the above steps, after obtaining the second object module, target data is concentrated and is divided according to the first preset ratio
To test set in test crowd's image be input to the second object module, through the second object module to test crowd's image procossing
Afterwards, target group's density map is obtained.
S214, according to the corresponding test crowd density map of crowd's image is tested in target group's density map and test set, really
The test accuracy rate of cover half type;
Through the above steps, after obtaining target group's density map, crowd will be tested in target group's density map and test set
The corresponding test crowd density map of image is compared analysis, determines the test accuracy rate of the second object module.
S215, judge whether test accuracy rate is greater than default accuracy rate;If so, step S216 is executed, if it is not, executing step
S210;
Through the above steps, after obtaining test accuracy rate, test accuracy rate is compared with default accuracy rate, is judged
Whether test accuracy rate is greater than default accuracy rate, if test accuracy rate is greater than default accuracy rate, thens follow the steps S216;If
Test accuracy rate is less than or equal to default accuracy rate, then it represents that the accuracy rate of the second object module is not up to standard, then by this second
Object module reacquires data and is trained again to model as first object model, Lai Zhihang step S210.
S216, using the second object module as the counter model of crowd.
Through the above steps, if it is judged that test accuracy rate is greater than default accuracy rate, then illustrate the second object module
Accuracy rate is met the requirements, using the second object module as the counter model of crowd.
The counter model construction method of the crowd of the present embodiment first carries out crowd's image in pre-stored data set
The processing of number of people mark, obtains crowd density figure, is labeled in this way to the number of people in crowd's image, do not influence on blocking for body
As a result, so not needing to improve the applicability of model using perspective view on retraining scene and test scene.To target data
The original trained crowd's image and original trained crowd's density map collected in the training set divided cuts, overturns and expands, and obtains
To enlarged image and expand density map, and will be enlarged by image and expand density map storage into training set, obtains target training set,
To realize the data amplification of training set, guarantee that the training data of model is sufficient.Convolutional neural networks are polymerize to Analysis On Multi-scale Features
Model iteration executes the training of first level and the optimization of first level, obtains first object model, then to first object model
Iteration executes the training of second level and the optimization of second level, obtains the second object module, realized in optimization process Euclidean away from
From loss formula and opposite number loss formula while Optimized model, to improve the precision of prediction of number.By target data set
Test crowd's image in the test set of division is input in the second object module, target group's density map is obtained, according to target
Test crowd's density map in crowd density figure and test set, determines the test accuracy rate of model, if it is judged that test is accurate
Rate is greater than default accuracy rate and is capable of the accuracy rate of testing model so then using the second object module as the counter model of crowd,
To guarantee the accuracy of count results.The Analysis On Multi-scale Features polymerization convolutional neural networks model that the present embodiment uses is with more
The single-row network architecture of scale feature polymerizable functional is trained and in application, both can solve the ruler of personage in image in model
Variation issue is spent, operand can be also reduced, improves recognition efficiency.
Fig. 5 is the flow chart of the method for counting embodiment of crowd of the invention, as shown in figure 5, the crowd of the present embodiment
Method of counting can specifically include following steps:
S301, crowd's image to be counted is obtained;
The present embodiment, it is necessary first to obtain crowd's image to be counted.Wherein, crowd's image to be counted is to need to count the number of people
Image.
S302, the counter model based on the crowd constructed in advance input crowd's image to be counted, it is close to obtain crowd to be counted
Degree figure;
Through the above steps, after getting crowd's image to be counted, which is input to by above-mentioned
In embodiment in the counter model of the crowd of the counter model construction method building of crowd, pass through the place of the counter model of the crowd
Reason, obtains the corresponding crowd density figure to be counted of crowd's image to be counted.Wherein, which is by many points
Composition, each point represents a people, and the position of point is head position of the people in crowd's image to be counted.
S303, using accumulation calculating method, numerical value in the crowd density figure to be counted that adds up in each pixel obtains pixel
Total value, using pixel total value as the number in crowd's image to be counted.
Through the above steps, after obtaining crowd density figure to be counted, using accumulation calculating method, add up crowd density to be counted
Numerical value in figure in each pixel obtains pixel total value, using pixel total value as the people in crowd's image to be counted
Number.
The method of counting of the crowd of the present embodiment, obtains crowd's image to be counted first, and by crowd's image to be counted
It is input in the counter model of the crowd constructed in advance by the counter model construction method of crowd, obtains crowd density to be counted
Figure, finally, using accumulation calculating method, numerical value in the crowd density figure to be counted that adds up in each pixel obtains pixel sum
Value, using pixel total value as the number in crowd's image to be counted.The present embodiment is carried out to the number of people in crowd's image
Mark, on body block do not influence as a result, so building crowd counter model when do not need in Training scene and test
The Analysis On Multi-scale Features polymerization for using perspective view in scene, improving the applicability of model, and used in the counter model of crowd
Convolutional neural networks model is the single-row network architecture with Analysis On Multi-scale Features polymerizable functional, is trained and applies in model
When, it both can solve the dimensional variation problem of personage in image, and can also reduce operand, and improved recognition efficiency.
In order to which more comprehensively, corresponding to the counter model construction method of crowd provided in an embodiment of the present invention, the application is also mentioned
The counter model construction device of crowd is supplied.
Fig. 6 is the structural schematic diagram of the counter model construction device embodiment one of crowd of the invention, as shown in fig. 6, this
The counter model construction device of the crowd of embodiment includes first processing module 11, data combination module 12, data amplification module
13 and Second processing module 14.
First processing module 11 is obtained for carrying out the processing of number of people mark to crowd's image in pre-stored data set
Crowd density figure;
Data combination module 12 is obtained for crowd's image and crowd density figure corresponding with crowd's image to be combined
To target data set;
Data expand module 13, carry out for concentrating to target data according to the training set that the first preset ratio divides
Data amplification processing, obtains target training set;
Second processing module 14 is right for utilizing training crowd image and training crowd's density map in target training set
It is trained through parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model, obtains the count module of crowd
Type.
The counter model construction device of the crowd of the present embodiment, first by first processing module 11 to pre-stored data
Crowd's image of concentration carries out the processing of number of people mark, obtains crowd density figure, and passes through data combination module 12 for crowd's image
It is combined with crowd density figure corresponding with crowd's image, obtains target data set;Module 13 is expanded to mesh by data again
Data amplification processing is carried out according to the training set that the first preset ratio divides in mark data set, obtains target training set;Most
Afterwards by Second processing module 14 using the training crowd image and training crowd's density map in target training set, to through at the beginning of parameter
Beginningization treated Analysis On Multi-scale Features polymerization convolutional neural networks model is trained, and obtains the counter model of crowd.Wherein, originally
Embodiment is labeled to the number of people in crowd's image, does not influence blocking for body as a result, so not needing retraining
Perspective view is used in scene and test scene, improves the applicability of model.Analysis On Multi-scale Features polymerize convolutional neural networks model
For the single-row network architecture with Analysis On Multi-scale Features polymerizable functional, it is trained in model and in application, both can solve in image
The dimensional variation problem of personage can also reduce operand, improve recognition efficiency.
Fig. 7 is the structural schematic diagram of the counter model construction device embodiment two of crowd of the invention, as shown in fig. 7, this
On the basis of counter model construction device embodiment described in Fig. 6 of the crowd of embodiment, data amplification module 13 includes cutting out
Unit 131, roll-over unit 132, expanding unit 133 and storage unit 134.
Unit 131 is cut, is used for through preset cutting mode, respectively trained crowd's image original to each of training set
And the corresponding original trained crowd's density map of original trained crowd's image is cut, and obtains at least two cutting images and at least
Two cutting density maps;
Roll-over unit 132 carries out overturning processing for that will cut image, obtains flipped image, and, density map will be cut
Overturning processing is carried out, overturning density map is obtained;
Expanding unit 133 will cut image and flipchart for the second preset ratio according to original trained crowd's image
As expansion, enlarged image is obtained, and, density map will be cut and overturning density map expands, obtain expansion density map;
Storage unit 134 obtains target training set for will be enlarged by image and expand density map storage into training set.
Further, in the counter model construction device of the crowd of the present embodiment, Second processing module 14 includes the first instruction
Practice unit 141, first and optimizes unit 142, iterative processing unit 143, the second training unit 144, second optimization unit 145 and mould
Type determination unit 146;The counter model construction device of the crowd of the present embodiment further includes model measurement module 15, determining module 16
With judgment module 17.
First training unit 141, for inputting the training crowd image in target training set with training crowd's density map
The training that first level is carried out in through parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model, obtains
First training pattern;
Wherein, training crowd's image includes original trained crowd's image and enlarged image;Training crowd's density map includes original
Begin training crowd's density map and expansion density map.
First optimization unit 142, for being carried out using model parameter of the Euclidean distance loss formula to the first training pattern
The optimization of first level obtains the first Optimized model;
Iterative processing unit 143 is used for training and first level to first level according to preset first the number of iterations
The process of optimization be iterated execution, obtain first object model;
Second training unit 144, for crowd's image and training crowd's density map will to be trained to be input to first object model
The middle training for carrying out second level, obtains the second training pattern;
Second optimization unit 145, for losing formula and opposite number loss formula to the second training using Euclidean distance
The model parameter of model carries out the optimization of second level, obtains the second Optimized model;
Iterative processing unit 143 is also used to training and the second level according to preset secondary iteration number, to second level
The process of other optimization is iterated execution, obtains the second object module;
Model measurement module 15, for concentrating target data in the test set divided according to the first preset ratio
Test crowd's image is input to the second object module, obtains target group's density map;
Determining module 16, for according to the corresponding test crowd of test crowd's image in target group's density map and test set
Density map determines the test accuracy rate of model;
Judgment module 17, for judging whether test accuracy rate is greater than default accuracy rate;
Model determination unit 146, if being greater than default accuracy rate for test accuracy rate, using the second object module as crowd
Counter model.
The counter model construction device of the crowd of the present embodiment, first processing module 11 is in pre-stored data set first
Crowd's image carry out the processing of number of people mark, obtain crowd density figure, the number of people in crowd's image be labeled in this way, for
Blocking for body is indifferent, so not needing to improve model using perspective view on retraining scene and test scene
Applicability.The original instruction in training set that cutting unit 131, roll-over unit 132 and expanding unit 133 divide target data set
Practice crowd's image and original trained crowd's density map cuts, overturns and expands, obtain enlarged image and expand density map, deposits
Storage unit 134 will be enlarged by image and expand density map storage into training set, target training set be obtained, to realize training set
Data amplification guarantees that the training data of model is sufficient.Iterative processing unit 143 polymerize convolutional neural networks mould to Analysis On Multi-scale Features
Type iteration executes the training of first level and the optimization of first level, obtains first object model, then change to first object model
The training of substitute performance second level and the optimization of second level obtain the second object module, realize Euclidean distance in optimization process
Formula and opposite number loss formula Optimized model simultaneously are lost, to improve the precision of prediction of number.Model measurement module 15
By target data set divide test set in test crowd's image be input in the second object module, determining module 16 according to
To target group's density map and test set in test crowd's density map, the test accuracy rate of model is determined, if it is determined that mould
Block 17 judges that test accuracy rate is greater than default accuracy rate, then model determination unit 146 is using the second object module as crowd's
Counter model, is capable of the accuracy rate of testing model in this way, to guarantee the accuracy of count results.More rulers that the present embodiment uses
Degree characteristic aggregation convolutional neural networks model is the single-row network architecture with Analysis On Multi-scale Features polymerizable functional, is instructed in model
Practice and in application, both can solve the dimensional variation problem of personage in image, can also reduce operand, improves recognition efficiency.
In order to which more comprehensively, corresponding to the method for counting of crowd provided in an embodiment of the present invention, present invention also provides crowds
Counting device.
Fig. 8 is the structural schematic diagram of the counting device embodiment of crowd of the invention, as shown in figure 8, the people of the present embodiment
The counting device of group includes image collection module 21, density map determining module 22 and demographics module 23.
Image collection module 21, for obtaining crowd's image to be counted;
Density map determining module 22 inputs crowd's image to be counted for the counter model based on the crowd constructed in advance,
Obtain crowd density figure to be counted;The counter model of crowd through the foregoing embodiment in crowd counter model construction method structure
It builds;
Demographics module 23, for utilizing accumulation calculating method, in the crowd density figure to be counted that adds up in each pixel
Numerical value obtains pixel total value, using pixel total value as the number in crowd's image to be counted.
The counting device of the crowd of the present embodiment, image collection module 21 obtain crowd's image to be counted, and density map determines
Crowd's image to be counted is input to the counting of the crowd constructed in advance by the counter model construction method of crowd by module 22
In model, crowd density figure to be counted is obtained, finally, demographics module 23 utilizes accumulation calculating method, add up crowd to be counted
Numerical value in density map in each pixel obtains pixel total value, using pixel total value as in crowd's image to be counted
Number.The present embodiment is labeled to the number of people in crowd's image, does not influence blocking for body as a result, so building
It does not need to improve the applicability of model using perspective view on Training scene and test scene when the counter model of crowd, and
And the Analysis On Multi-scale Features polymerization convolutional neural networks model used in the counter model of crowd is with Analysis On Multi-scale Features polymerization function
Can the single-row network architecture, be trained in model and in application, both can solve the dimensional variation problem of personage in image, also can
Operand is reduced, recognition efficiency is improved.
About the device in above-described embodiment, wherein modules execute the concrete mode of operation in related this method
Embodiment in be described in detail, no detailed explanation will be given here.
It is understood that same or similar part can mutually refer in the various embodiments described above, in some embodiments
Unspecified content may refer to the same or similar content in other embodiments.
It should be noted that in the description of the present invention, term " first ", " second " etc. are used for description purposes only, without
It can be interpreted as indication or suggestion relative importance.In addition, in the description of the present invention, unless otherwise indicated, the meaning of " multiple "
Refer at least two.
Any process described otherwise above or method description are construed as in flow chart or herein, and expression includes
It is one or more for realizing specific logical function or process the step of executable instruction code module, segment or portion
Point, and the range of the preferred embodiment of the present invention includes other realization, wherein can not press shown or discussed suitable
Sequence, including according to related function by it is basic simultaneously in the way of or in the opposite order, Lai Zhihang function, this should be of the invention
Embodiment person of ordinary skill in the field understood.
It should be appreciated that each section of the invention can be realized with hardware, software, firmware or their combination.Above-mentioned
In embodiment, software that multiple steps or method can be executed in memory and by suitable instruction execution system with storage
Or firmware is realized.It, and in another embodiment, can be under well known in the art for example, if realized with hardware
Any one of column technology or their combination are realized: having a logic gates for realizing logic function to data-signal
Discrete logic, with suitable combinational logic gate circuit specific integrated circuit, programmable gate array (PGA), scene
Programmable gate array (FPGA) etc..
Those skilled in the art are understood that realize all or part of step that above-described embodiment method carries
It suddenly is that relevant hardware can be instructed to complete by program, described program can store in a kind of computer readable storage medium
In, which when being executed, includes the steps that one or a combination set of embodiment of the method.
It, can also be in addition, each functional unit in each embodiment of the present invention can integrate in a processing module
It is that each unit physically exists alone, can also be integrated in two or more units in a module.Above-mentioned integrated mould
Block both can take the form of hardware realization, can also be realized in the form of software function module.The integrated module is such as
Fruit is realized and when sold or used as an independent product in the form of software function module, also can store in a computer
In read/write memory medium.
Storage medium mentioned above can be read-only memory, disk or CD etc..
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
Centainly refer to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be any
One or more embodiment or examples in can be combined in any suitable manner.
Although the embodiments of the present invention has been shown and described above, it is to be understood that above-described embodiment is example
Property, it is not considered as limiting the invention, those skilled in the art within the scope of the invention can be to above-mentioned
Embodiment is changed, modifies, replacement and variant.
Claims (10)
1. the counter model construction method of crowd a kind of characterized by comprising
The processing of number of people mark is carried out to crowd's image in pre-stored data set, obtains crowd density figure;
Crowd's image and crowd density figure corresponding with crowd's image are combined, target data set is obtained;
The target data is concentrated and carries out data amplification processing according to the training set that the first preset ratio divides, obtains mesh
Mark training set;
Using the training crowd image and training crowd's density map in the target training set, to through parameter initialization, treated
Analysis On Multi-scale Features polymerization convolutional neural networks model is trained, and obtains the counter model of crowd.
2. the method according to claim 1, wherein described concentrate according to the first default ratio the target data
Example divides obtained training set and carries out data amplification processing, obtains target training set, comprising:
By preset cutting mode, original trained crowd's image described to each of the training set and the original instruction respectively
Practice the corresponding original trained crowd's density map of crowd's image to cut, obtains at least two cutting images and at least two
Cut density map;
The cutting image is subjected to overturning processing, obtains flipped image, and, the cutting density map is carried out at overturning
Reason obtains overturning density map;
According to the second preset ratio of original trained crowd's image, the cutting image and the flipped image are expanded,
Enlarged image is obtained, and, the cutting density map and the overturning density map are expanded, obtain expanding density map;
By the enlarged image and expansion density map storage into the training set, target training set is obtained;
Wherein, trained crowd's image includes original trained crowd's image and the enlarged image;The trained crowd
Density map includes original trained crowd's density map and the expansion density map.
3. the method according to claim 1, wherein the training crowd using in the target training set schemes
Picture and training crowd's density map are instructed to through parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model
Practice, obtain the counter model of crowd, comprising:
By in the target training set trained crowd's image and trained crowd's density map be input to it is initial through parameter
The training that first level is carried out in Analysis On Multi-scale Features polymerization convolutional neural networks model of changing that treated, obtains the first training
Model;
The optimization for carrying out first level to the model parameter of first training pattern using Euclidean distance loss formula obtains the
One Optimized model;
According to preset first the number of iterations, the process of the optimization of training and the first level to the first level is carried out
Iteration executes, and obtains first object model;
Trained crowd's image and trained crowd's density map are input in the first object model and carry out the second level
Other training obtains the second training pattern;
Using the Euclidean distance lose formula and opposite number loss formula to the model parameter of second training pattern into
The optimization of row second level obtains the second Optimized model;
According to preset secondary iteration number, the process of the optimization of training and the second level to the second level is carried out
Iteration executes, and obtains the second object module;
Using second object module as the counter model of the crowd.
4. according to the method described in claim 3, it is characterized in that, described using second object module as the crowd's
Before counter model, further includes:
Concentrate test crowd's image in the test set divided according to first preset ratio defeated the target data
Enter to second object module, obtains target group's density map;
According to the corresponding test crowd density of the test crowd image in target group's density map and the test set
Figure, determines the test accuracy rate of model;
Judge whether the test accuracy rate is greater than default accuracy rate;
It is accordingly, described using second object module as the counter model of the crowd, comprising:
If the test accuracy rate is greater than the default accuracy rate, using second object module as the count module of the crowd
Type.
5. according to the method described in claim 3, it is characterized in that, the Euclidean distance loses formula are as follows:
Wherein, Θ expression parameter model;F(Xi;Θ) indicate output model;XiIndicate trained crowd's figure of i-th of input
Picture;FiIndicate the corresponding trained crowd's density map of trained crowd's image of i-th of input;N indicates input data
Number;
The opposite number loses formula are as follows:
Wherein, FD(Xi;Θ) indicate the number of prediction;DiIndicate true value number;The number of N expression input data;Denominator is Di+1
It is that denominator is zero in order to prevent;The training weight of the opposite number loss is L=1.0*L (Θ)+0.1*LD(Θ)。
6. a kind of method of counting of crowd characterized by comprising
Obtain crowd's image to be counted;
Based on the counter model of the crowd constructed in advance, crowd's image to be counted is inputted, crowd density figure to be counted is obtained;
The counter model of the crowd is constructed by the counter model construction method of the described in any item crowds of claim 1-5;
Using accumulation calculating method, numerical value in the crowd density figure to be counted that adds up in each pixel obtains pixel sum
Value, using the pixel total value as the number in crowd's image to be counted.
7. the counter model construction device of crowd a kind of characterized by comprising
It is close to obtain crowd for carrying out the processing of number of people mark to crowd's image in pre-stored data set for first processing module
Degree figure;
Data combination module, for crowd's image and crowd density figure corresponding with crowd's image to be combined,
Obtain target data set;
Data expand module, count for concentrating to the target data according to the training set that the first preset ratio divides
It is handled according to amplification, obtains target training set;
Second processing module, for utilizing training crowd image and training crowd's density map in the target training set, to warp
Parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model is trained, and obtains the counter model of crowd.
8. device according to claim 7, which is characterized in that the Second processing module includes: the first training unit,
One optimization unit, iterative processing unit, the second training unit, the second optimization unit and model determination unit;
First training unit, for by the target training set trained crowd's image and the trained crowd it is close
Degree figure is input to through carrying out the first order in parameter initialization treated Analysis On Multi-scale Features polymerization convolutional neural networks model
Other training obtains the first training pattern;
The first optimization unit, for being carried out using model parameter of the Euclidean distance loss formula to first training pattern
The optimization of first level obtains the first Optimized model;
The iterative processing unit, for according to preset first the number of iterations, training to the first level and described the
The process of the other optimization of level-one is iterated execution, obtains first object model;
Second training unit, for trained crowd's image and trained crowd's density map to be input to described first
The training that second level is carried out in object module, obtains the second training pattern;
The second optimization unit, for losing formula and opposite number loss formula to described second using the Euclidean distance
The model parameter of training pattern carries out the optimization of second level, obtains the second Optimized model;
The iterative processing unit is also used to training according to preset secondary iteration number, to the second level and described
The process of the optimization of second level is iterated execution, obtains the second object module;
Model determination unit, for using second object module as the counter model of the crowd.
9. device according to claim 8, which is characterized in that the Analysis On Multi-scale Features polymerize convolutional neural networks model packet
Include Feature Mapping module, Analysis On Multi-scale Features aggregation module and density map regression block;
The Feature Mapping module uses the preceding X convolutional layer of VGG-16 structure, and wherein X is 4 or 6;
The convolutional layer is the convolution kernel of 3*3;
The Analysis On Multi-scale Features aggregation module uses at least two scale branches;
The density map regression block uses B layers of A column of empty convolution, and wherein A and B is positive integer.
10. a kind of counting device of crowd characterized by comprising image collection module, density map determining module and number system
Count module;
Described image obtains module, for obtaining crowd's image to be counted;
The density map determining module inputs crowd's figure to be counted for the counter model based on the crowd constructed in advance
Picture obtains crowd density figure to be counted;The counter model of the crowd passes through the meter of the described in any item crowds of claim 1-5
The building of exponential model construction method;
The demographics module, for utilizing accumulation calculating method, in the crowd density figure to be counted that adds up in each pixel
Numerical value, pixel total value is obtained, using the pixel total value as the number in crowd's image to be counted.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910497050.8A CN110210603A (en) | 2019-06-10 | 2019-06-10 | Counter model construction method, method of counting and the device of crowd |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910497050.8A CN110210603A (en) | 2019-06-10 | 2019-06-10 | Counter model construction method, method of counting and the device of crowd |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110210603A true CN110210603A (en) | 2019-09-06 |
Family
ID=67791824
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910497050.8A Pending CN110210603A (en) | 2019-06-10 | 2019-06-10 | Counter model construction method, method of counting and the device of crowd |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110210603A (en) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110852172A (en) * | 2019-10-15 | 2020-02-28 | 华东师范大学 | Method for expanding crowd counting data set based on Cycle Gan picture collage and enhancement |
CN110879982A (en) * | 2019-11-15 | 2020-03-13 | 苏州大学 | Crowd counting system and method |
CN110991317A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Crowd counting method based on multi-scale perspective sensing type network |
CN111046747A (en) * | 2019-11-21 | 2020-04-21 | 北京金山云网络技术有限公司 | Crowd counting model training method, crowd counting method, device and server |
CN111091093A (en) * | 2019-12-17 | 2020-05-01 | 苏州大学 | Method, system and related device for estimating number of high-density crowds |
CN111429466A (en) * | 2020-03-19 | 2020-07-17 | 北京航空航天大学 | Space-based crowd counting and density estimation method based on multi-scale information fusion network |
CN111507183A (en) * | 2020-03-11 | 2020-08-07 | 杭州电子科技大学 | Crowd counting method based on multi-scale density map fusion cavity convolution |
CN111626184A (en) * | 2020-05-25 | 2020-09-04 | 齐鲁工业大学 | Crowd density estimation method and system |
JP2021047538A (en) * | 2019-09-17 | 2021-03-25 | キヤノン株式会社 | Image processing device, image processing method, and program |
CN112632601A (en) * | 2020-12-16 | 2021-04-09 | 苏州玖合智能科技有限公司 | Crowd counting method for subway carriage scene |
CN112633106A (en) * | 2020-12-16 | 2021-04-09 | 苏州玖合智能科技有限公司 | Crowd characteristic recognition network construction and training method suitable for large depth of field |
CN112651390A (en) * | 2021-03-11 | 2021-04-13 | 江苏金智教育信息股份有限公司 | Method and device for counting picture human heads based on convolutional neural network |
CN112734736A (en) * | 2021-01-15 | 2021-04-30 | 山东农业大学 | Accurate measuring and calculating method and device for swarm vigor based on computer vision |
CN113468995A (en) * | 2021-06-22 | 2021-10-01 | 之江实验室 | Crowd counting method based on density grade perception |
CN114463694A (en) * | 2022-01-06 | 2022-05-10 | 中山大学 | Semi-supervised crowd counting method and device based on pseudo label |
CN115797873A (en) * | 2023-02-06 | 2023-03-14 | 泰山学院 | Crowd density detection method, system, equipment, storage medium and robot |
CN116311083A (en) * | 2023-05-19 | 2023-06-23 | 华东交通大学 | Crowd counting model training method and system |
CN117953361A (en) * | 2024-03-27 | 2024-04-30 | 西北工业大学青岛研究院 | Underwater fish shoal small target steady counting method based on density map |
Citations (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080118106A1 (en) * | 2006-11-22 | 2008-05-22 | Regents Of The University Of Minnesota | Crowd counting and monitoring |
US20160055645A1 (en) * | 2013-03-26 | 2016-02-25 | Hitachi Kokusai Electric Inc. | People counting device and people counting method |
WO2016183770A1 (en) * | 2015-05-18 | 2016-11-24 | Xiaogang Wang | A system and a method for predicting crowd attributes |
CN106570564A (en) * | 2016-11-03 | 2017-04-19 | 天津大学 | Multi-scale pedestrian detection method based on depth network |
CN106599827A (en) * | 2016-12-09 | 2017-04-26 | 浙江工商大学 | Small target rapid detection method based on deep convolution neural network |
CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
CN107330364A (en) * | 2017-05-27 | 2017-11-07 | 上海交通大学 | A kind of people counting method and system based on cGAN networks |
US20180005071A1 (en) * | 2013-06-25 | 2018-01-04 | University Of Central Florida Research Foundation, Inc. | Multi-Source, Multi-Scale Counting in Dense Crowd Images |
US20180018524A1 (en) * | 2015-12-16 | 2018-01-18 | Intel Corporation | Fully convolutional pyramid networks for pedestrian detection |
US20180068198A1 (en) * | 2016-09-06 | 2018-03-08 | Carnegie Mellon University | Methods and Software for Detecting Objects in an Image Using Contextual Multiscale Fast Region-Based Convolutional Neural Network |
US9953236B1 (en) * | 2017-03-10 | 2018-04-24 | TuSimple | System and method for semantic segmentation using dense upsampling convolution (DUC) |
US20180231871A1 (en) * | 2016-06-27 | 2018-08-16 | Zhejiang Gongshang University | Depth estimation method for monocular image based on multi-scale CNN and continuous CRF |
CN108615027A (en) * | 2018-05-11 | 2018-10-02 | 常州大学 | A method of video crowd is counted based on shot and long term memory-Weighted Neural Network |
CN108876774A (en) * | 2018-06-07 | 2018-11-23 | 浙江大学 | A kind of people counting method based on convolutional neural networks |
CN109101930A (en) * | 2018-08-18 | 2018-12-28 | 华中科技大学 | A kind of people counting method and system |
CN109241858A (en) * | 2018-08-13 | 2019-01-18 | 湖南信达通信息技术有限公司 | A kind of passenger flow density detection method and device based on rail transit train |
CN109271960A (en) * | 2018-10-08 | 2019-01-25 | 燕山大学 | A kind of demographic method based on convolutional neural networks |
CN109359520A (en) * | 2018-09-04 | 2019-02-19 | 汇纳科技股份有限公司 | People counting method, system, computer readable storage medium and server |
CN109543695A (en) * | 2018-10-26 | 2019-03-29 | 复旦大学 | General density people counting method based on multiple dimensioned deep learning |
CN109636802A (en) * | 2019-01-18 | 2019-04-16 | 天津工业大学 | Pulmonary parenchyma based on depth convolutional neural networks is through CT image partition method |
CN109815867A (en) * | 2019-01-14 | 2019-05-28 | 东华大学 | A kind of crowd density estimation and people flow rate statistical method |
US20190228268A1 (en) * | 2016-09-14 | 2019-07-25 | Konica Minolta Laboratory U.S.A., Inc. | Method and system for cell image segmentation using multi-stage convolutional neural networks |
-
2019
- 2019-06-10 CN CN201910497050.8A patent/CN110210603A/en active Pending
Patent Citations (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080118106A1 (en) * | 2006-11-22 | 2008-05-22 | Regents Of The University Of Minnesota | Crowd counting and monitoring |
US20160055645A1 (en) * | 2013-03-26 | 2016-02-25 | Hitachi Kokusai Electric Inc. | People counting device and people counting method |
US20180005071A1 (en) * | 2013-06-25 | 2018-01-04 | University Of Central Florida Research Foundation, Inc. | Multi-Source, Multi-Scale Counting in Dense Crowd Images |
WO2016183770A1 (en) * | 2015-05-18 | 2016-11-24 | Xiaogang Wang | A system and a method for predicting crowd attributes |
US20180018524A1 (en) * | 2015-12-16 | 2018-01-18 | Intel Corporation | Fully convolutional pyramid networks for pedestrian detection |
US20180231871A1 (en) * | 2016-06-27 | 2018-08-16 | Zhejiang Gongshang University | Depth estimation method for monocular image based on multi-scale CNN and continuous CRF |
US20180068198A1 (en) * | 2016-09-06 | 2018-03-08 | Carnegie Mellon University | Methods and Software for Detecting Objects in an Image Using Contextual Multiscale Fast Region-Based Convolutional Neural Network |
US20190228268A1 (en) * | 2016-09-14 | 2019-07-25 | Konica Minolta Laboratory U.S.A., Inc. | Method and system for cell image segmentation using multi-stage convolutional neural networks |
CN106570564A (en) * | 2016-11-03 | 2017-04-19 | 天津大学 | Multi-scale pedestrian detection method based on depth network |
CN106599827A (en) * | 2016-12-09 | 2017-04-26 | 浙江工商大学 | Small target rapid detection method based on deep convolution neural network |
CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
US9953236B1 (en) * | 2017-03-10 | 2018-04-24 | TuSimple | System and method for semantic segmentation using dense upsampling convolution (DUC) |
CN107330364A (en) * | 2017-05-27 | 2017-11-07 | 上海交通大学 | A kind of people counting method and system based on cGAN networks |
CN108615027A (en) * | 2018-05-11 | 2018-10-02 | 常州大学 | A method of video crowd is counted based on shot and long term memory-Weighted Neural Network |
CN108876774A (en) * | 2018-06-07 | 2018-11-23 | 浙江大学 | A kind of people counting method based on convolutional neural networks |
CN109241858A (en) * | 2018-08-13 | 2019-01-18 | 湖南信达通信息技术有限公司 | A kind of passenger flow density detection method and device based on rail transit train |
CN109101930A (en) * | 2018-08-18 | 2018-12-28 | 华中科技大学 | A kind of people counting method and system |
CN109359520A (en) * | 2018-09-04 | 2019-02-19 | 汇纳科技股份有限公司 | People counting method, system, computer readable storage medium and server |
CN109271960A (en) * | 2018-10-08 | 2019-01-25 | 燕山大学 | A kind of demographic method based on convolutional neural networks |
CN109543695A (en) * | 2018-10-26 | 2019-03-29 | 复旦大学 | General density people counting method based on multiple dimensioned deep learning |
CN109815867A (en) * | 2019-01-14 | 2019-05-28 | 东华大学 | A kind of crowd density estimation and people flow rate statistical method |
CN109636802A (en) * | 2019-01-18 | 2019-04-16 | 天津工业大学 | Pulmonary parenchyma based on depth convolutional neural networks is through CT image partition method |
Non-Patent Citations (5)
Title |
---|
SIYU HUANG: "Stacked Pooling: Improving Crowd Counting by Boosting Scale Invariance", 《ARXIV:1808.07456》, pages 2 - 4 * |
XINKUN CAO: "Scale aggregation network for accurate and efficient crowd counting", 《PROCEEDINGS OF THE EUROPEAN CONFERENCE ON COMPUTER VISION(ECCV)》, pages 1 - 17 * |
吴青科: "结合神经网络与多列特征图聚合的人群计数", pages 214 - 216 * |
范绿源;仝明磊;李敏;南昊;: "静态图像中采用混合卷积结构进行人群密度估计", no. 03 * |
邓远志;胡钢;: "基于特征金字塔的人群密度估计方法", no. 06 * |
Cited By (30)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11721036B2 (en) | 2019-09-17 | 2023-08-08 | Canon Kabushiki Kaisha | Image processing apparatus, image processing method, and computer-readable medium |
JP2021047538A (en) * | 2019-09-17 | 2021-03-25 | キヤノン株式会社 | Image processing device, image processing method, and program |
CN112598725A (en) * | 2019-09-17 | 2021-04-02 | 佳能株式会社 | Image processing apparatus, image processing method, and computer readable medium |
JP7383435B2 (en) | 2019-09-17 | 2023-11-20 | キヤノン株式会社 | Image processing device, image processing method, and program |
CN110852172A (en) * | 2019-10-15 | 2020-02-28 | 华东师范大学 | Method for expanding crowd counting data set based on Cycle Gan picture collage and enhancement |
CN110879982A (en) * | 2019-11-15 | 2020-03-13 | 苏州大学 | Crowd counting system and method |
CN110879982B (en) * | 2019-11-15 | 2023-06-20 | 苏州大学 | Crowd counting system and method |
CN111046747A (en) * | 2019-11-21 | 2020-04-21 | 北京金山云网络技术有限公司 | Crowd counting model training method, crowd counting method, device and server |
CN111046747B (en) * | 2019-11-21 | 2023-04-18 | 北京金山云网络技术有限公司 | Crowd counting model training method, crowd counting method, device and server |
CN110991317A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Crowd counting method based on multi-scale perspective sensing type network |
CN110991317B (en) * | 2019-11-29 | 2023-05-16 | 中山大学 | Crowd counting method based on multi-scale perspective sensing network |
CN111091093A (en) * | 2019-12-17 | 2020-05-01 | 苏州大学 | Method, system and related device for estimating number of high-density crowds |
CN111507183B (en) * | 2020-03-11 | 2021-02-02 | 杭州电子科技大学 | Crowd counting method based on multi-scale density map fusion cavity convolution |
CN111507183A (en) * | 2020-03-11 | 2020-08-07 | 杭州电子科技大学 | Crowd counting method based on multi-scale density map fusion cavity convolution |
CN111429466A (en) * | 2020-03-19 | 2020-07-17 | 北京航空航天大学 | Space-based crowd counting and density estimation method based on multi-scale information fusion network |
CN111626184A (en) * | 2020-05-25 | 2020-09-04 | 齐鲁工业大学 | Crowd density estimation method and system |
CN112632601B (en) * | 2020-12-16 | 2024-03-12 | 苏州玖合智能科技有限公司 | Crowd counting method for subway carriage scene |
CN112633106A (en) * | 2020-12-16 | 2021-04-09 | 苏州玖合智能科技有限公司 | Crowd characteristic recognition network construction and training method suitable for large depth of field |
CN112632601A (en) * | 2020-12-16 | 2021-04-09 | 苏州玖合智能科技有限公司 | Crowd counting method for subway carriage scene |
CN112734736A (en) * | 2021-01-15 | 2021-04-30 | 山东农业大学 | Accurate measuring and calculating method and device for swarm vigor based on computer vision |
CN112734736B (en) * | 2021-01-15 | 2022-10-04 | 山东农业大学 | Accurate measuring and calculating method and device for swarm vigor based on computer vision |
CN112651390A (en) * | 2021-03-11 | 2021-04-13 | 江苏金智教育信息股份有限公司 | Method and device for counting picture human heads based on convolutional neural network |
CN113468995A (en) * | 2021-06-22 | 2021-10-01 | 之江实验室 | Crowd counting method based on density grade perception |
CN114463694A (en) * | 2022-01-06 | 2022-05-10 | 中山大学 | Semi-supervised crowd counting method and device based on pseudo label |
CN114463694B (en) * | 2022-01-06 | 2024-04-05 | 中山大学 | Pseudo-label-based semi-supervised crowd counting method and device |
CN115797873A (en) * | 2023-02-06 | 2023-03-14 | 泰山学院 | Crowd density detection method, system, equipment, storage medium and robot |
CN115797873B (en) * | 2023-02-06 | 2023-05-26 | 泰山学院 | Crowd density detection method, system, equipment, storage medium and robot |
CN116311083A (en) * | 2023-05-19 | 2023-06-23 | 华东交通大学 | Crowd counting model training method and system |
CN116311083B (en) * | 2023-05-19 | 2023-09-05 | 华东交通大学 | Crowd counting model training method and system |
CN117953361A (en) * | 2024-03-27 | 2024-04-30 | 西北工业大学青岛研究院 | Underwater fish shoal small target steady counting method based on density map |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110210603A (en) | Counter model construction method, method of counting and the device of crowd | |
CN110210551B (en) | Visual target tracking method based on adaptive subject sensitivity | |
CN112541483B (en) | Dense face detection method combining YOLO and blocking-fusion strategy | |
CN114937151B (en) | Lightweight target detection method based on multiple receptive fields and attention feature pyramid | |
CN105512289B (en) | Image search method based on deep learning and Hash | |
CN112101430B (en) | Anchor frame generation method for image target detection processing and lightweight target detection method | |
Jain et al. | Autoencoders for level generation, repair, and recognition | |
CN106845621B (en) | Dense population number method of counting and system based on depth convolutional neural networks | |
CN106874688B (en) | Intelligent lead compound based on convolutional neural networks finds method | |
CN107229904A (en) | A kind of object detection and recognition method based on deep learning | |
CN108681774A (en) | Based on the human body target tracking method for generating confrontation network negative sample enhancing | |
CN108470354A (en) | Video target tracking method, device and realization device | |
CN109002845A (en) | Fine granularity image classification method based on depth convolutional neural networks | |
CN108734151A (en) | Robust long-range method for tracking target based on correlation filtering and the twin network of depth | |
JP2018005520A (en) | Object detection device and object detection method | |
CN110309835B (en) | Image local feature extraction method and device | |
CN108038420A (en) | A kind of Human bodys' response method based on deep video | |
CN103942571B (en) | Graphic image sorting method based on genetic programming algorithm | |
CN107529650A (en) | Network model construction and closed loop detection method, corresponding device and computer equipment | |
CN110378348A (en) | Instance of video dividing method, equipment and computer readable storage medium | |
CN109377445A (en) | Model training method, the method, apparatus and electronic system for replacing image background | |
CN109492596B (en) | Pedestrian detection method and system based on K-means clustering and regional recommendation network | |
CN110059616A (en) | Pedestrian's weight identification model optimization method based on fusion loss function | |
CN108664838A (en) | Based on the monitoring scene pedestrian detection method end to end for improving RPN depth networks | |
CN106651915A (en) | Target tracking method of multi-scale expression based on convolutional neural network |
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 |
Application publication date: 20190906 |