CN107624189A - Method and apparatus for generating forecast model - Google Patents
Method and apparatus for generating forecast model Download PDFInfo
- Publication number
- CN107624189A CN107624189A CN201580080145.XA CN201580080145A CN107624189A CN 107624189 A CN107624189 A CN 107624189A CN 201580080145 A CN201580080145 A CN 201580080145A CN 107624189 A CN107624189 A CN 107624189A
- Authority
- CN
- China
- Prior art keywords
- image
- cnn
- training
- true value
- frame
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
Abstract
A kind of method for being used to generate that forecast model is distributed with the crowd density in prognostic chart picture frame and personnel count is disclosed, it includes:CNN is trained by inputting one or more crowd's image blocks of the frame in training set, in the crowd's image block inputted, there are each crowd's image block predetermined actual conditions Density Distribution and personnel to count;Frame is sampled from target scene image set and the training image with identified true value Density Distribution and counting/number is received from training set;Similar view data is retrieved from the training frames received for each target image frame sampled, to overcome the scene gap between target scene image set and training image;And by the way that similar view data is input into CNN to be finely adjusted to CNN, to determine the forecast model for the crowd density figure in prognostic chart picture frame and personnel's counting.
Description
Technical field
The application is related to be set for generating forecast model with what the crowd density distribution in prognostic chart picture frame and personnel counted
Standby and method.
Background technology
Pedestrian crowd in video is counted strong demand in terms of video monitoring, thus has attracted many
Notice.Due to serious shielding, scene perspective distortion and Crowds Distribute variation, it is one challenging that crowd, which counts,
Task.Due to pedestrian detection and tracking when in for crowd's scene it is difficult, so the most methods of state of the art
It is to be based on returning, and target is to learn the mapping between low-level feature and crowd's counting.However, these work are for specific
Scene, i.e., it may only be applied to Same Scene for crowd's counter model of special scenes study.In view of the field that can't see
Scape or the scene layout changed, it is necessary to new mark come re -training model.
There are many work to pass through detection or trajectory clustering (trajectory-clustering) to be counted to pedestrian
Number.But for crowd's enumeration problem, these methods are limited by serious block between people.Many kinds of methods attempt to lead to
The recurrence first (regressor) using being trained by low-level feature is crossed to predict global counting.These methods are more suitable for crowded
Environment, and it is computationally more effective.
Carry out counting the spatial information that have ignored pedestrian by global recurrence.Lempitsky et al.One kind is described to pass through
Pixel layer object densities figure returns the object count method carried out.After the work,Fiaschi et al.Use random forest
(random forest) returns object densities and improves training effectiveness.In addition to considering spatial information, returned based on density
Another advantage of method be:They can be estimated the counting of the object in any region of image.Utilize this
Advantage, interactive object number system is introduced, it visualizes area count to help user effectively to determine that correlation is anti-
Feedback.RodriguezeNumber of people testing result is improved using density map estimation.But these methods are for special scenes, and
It is not suitable for intersecting scene counting.
Many work employ studies in depth it is various monitoring application, such as personnel identify again, pedestrian detection, tracking, crowd
Behavioural analysis and crowd's segmentation.The differentiation power of depth model is benefited from their success.Sermanet et al.Show, for many
It is more effective from the hand-made feature of aspect ratio of depth model extraction for.However, not yet develop the depth counted for crowd
Spend model.
As many large-scale and good indication data sets are disclosed, a kind of non-parametric, data-driven side
Method is proposed out.Such method can be extended up easily, because they need not be trained.They are most like by obtaining
Label is transferred to test image by training image from training image, and these training images is matched with test image.Liu etc. PeoplePropose a kind of nonparametric method for analyzing image for the intensive deformation field sought between image.
The content of the invention
The disclosure solves the problems, such as crowd density and counts estimation, and its target is automatically to estimate given monitor video
Number/counting of density map and/or personnel on frame.
Intersect scene density present applicant proposes one kind and count estimating system.Even if any target is not present in training set
Scene, the system still are able to estimate that the density map of the scene counts with personnel.
In an aspect, a kind of equipment for being used to generate forecast model to predict crowd density figure and count is disclosed,
It includes density map creating unit, CNN generation units, set of metadata of similar data acquiring unit and model fine-adjusting unit.Density map creates single
Member, which is configured to the approximate training set that comes from, (to be had pedestrian's number of people mark, is indicated in everyone number of people position in area-of-interest (ROI)
Put) each Training scene perspective view (perspective map), it is true to be created based on label and perspective view on training set
It is worth (ground-truth) density map and counting.Density map represents the Crowds Distribute of each frame, and the integration of density map is equal to
Pedestrian's sum.CNN generation units are configured to construction and initialization crowd's convolutional neural networks, to be sampled by input from training set
To CNN crowd's image block and corresponding true value density map and count to train CNN.Set of metadata of similar data acquiring unit is configured to:
Sample frame is received from target scene and is received from training set with the true value density map created by CNN generation units and the sample of counting
This;And set of metadata of similar data is obtained from training set for each target scene, to overcome scene gap.Model fine-adjusting unit configures
Into the set of metadata of similar data being acquired and the CNN of construction the 2nd is received, wherein initializing second by using the first CNN trained
CNN, and model fine-adjusting unit is further configured to be finely adjusted the 2nd initialized CNN with set of metadata of similar data, so that the
Two CNN can predict density map and pedestrian counting and the region of interest in the area-of-interest of frame of video to be detected
Domain.
In present aspects, disclose a kind of for generating forecast model with the crowd density in prognostic chart picture frame point
The method that cloth and personnel count, methods described include:
CNN is trained by inputting one or more crowd's image blocks of the frame in training set, in the crowd inputted
There are each crowd's image block predetermined true value Density Distribution and personnel to count in image block;
Frame is sampled from target scene image set and received from training set have identified true value Density Distribution and
The training image of counting/number, and
Similar view data is obtained from the training frames received for each target image frame sampled, to overcome
Scene gap between target scene image set and training image;And
By the way that similar view data is input into CNN to be finely adjusted to CNN, to determine to be used in prognostic chart picture frame
The forecast model that crowd density figure and personnel count.
The application it is other in terms of in, disclose it is a kind of be used for generate forecast model with the crowd in prognostic chart picture frame
The equipment that Density Distribution and personnel count, the equipment include:
CNN training units, CNN is trained by inputting one or more crowd's image blocks of the frame in training set,
In the crowd's image block inputted, there are each crowd's image block predetermined true value Density Distribution and personnel to count;
Set of metadata of similar data acquiring unit, frame is sampled from target scene image set and from training set receive have determined
True value Density Distribution and counting/number training image;And each target image frame for being sampled is from being received
Similar view data is obtained in training frames, to overcome the scene gap between target scene image set and training image;And
Model fine-adjusting unit, by the way that similar view data is input into CNN to be finely adjusted to CNN, to determine to use
The forecast model that crowd density figure and personnel in prognostic chart picture frame count.
The application it is other in terms of in, disclose it is a kind of be used for generate forecast model with the crowd in prognostic chart picture frame
The system that Density Distribution and personnel count, the system include:
Memory, it stores executable part;And
Processor, it is electrically coupled to memory to perform operation of the executable part with execution system, wherein, it is described to hold
Row part includes:
CNN training components, it is used to train by inputting one or more crowd's image blocks of the frame in training set
CNN, in the crowd's image block inputted, each crowd's image block has predetermined true value Density Distribution and personnel
Count;
Set of metadata of similar data obtaining widget, frame is sampled from target scene image set and from training set receive have determined
True value Density Distribution and counting/number training image;And each target image frame for being sampled is from being received
Similar view data is obtained in training frames, to overcome the scene gap between target scene image set and training image;And
Model trimming part, by the way that similar view data is input into CNN to be finely adjusted to CNN, to determine to use
The forecast model that crowd density figure and personnel in prognostic chart picture frame count.
According to required solution.There will be at least one of advantages below:
Multitask system-its can estimate together crowd density figure with count.It can be counted by the integration to density map
Calculate and keep count of.Two inter-related tasks can also help each other, so as to obtain more preferable solution for our training pattern.
Intersect scene ability-target scene and do not need extra pedestrian's label in the framework for intersecting scene counting.
Do not need crowd segmentation-its independent of crowd's foreground segmentation pre-process., all will be logical no matter whether crowd moves
Our model is crossed to capture crowd's texture, and system can obtain rational estimated result.
The following description and drawings elaborate some illustrative aspects of the disclosure.However, these aspect instructions can use this
Only a small number of modes in the various modes of principle disclosed.When considered in conjunction with the accompanying drawings, other aspects of the disclosure will be from this public affairs
That opens described in detail below becomes apparent.
Brief description of the drawings
The Exemplary, non-limiting embodiment of the present invention is described below with reference to accompanying drawing.Accompanying drawing is illustrative, and
Do not drawn in definite ratio typically.The same or like element on different figures is quoted with identical drawing reference numeral.
Fig. 1 is the schematic diagram of the block diagram for the equipment 1000 for illustrating an embodiment according to the application, and the equipment is used
In generation forecast model to predict crowd density figure and counting.
Fig. 2 is to illustrate to generate prediction module according to the equipment 1000 of the embodiment of the application with prognostic chart picture frame
Crowd density distribution and personnel count flow schematic diagram.
Fig. 3 is the signal of the flowchart process for the density map creating unit 10 for illustrating an embodiment according to the application
Figure.
Fig. 4 is the schematic diagram of the flowchart process for the CNN training units for illustrating an embodiment according to the application.
Fig. 5 is the schematic diagram of the general introduction for the crowd's CNN models for illustrating an embodiment according to the application, shown people
Group's CNN models have changeable target (switchable objective).
Fig. 6 is the schematic diagram for illustrating the flow according to the acquisition of the set of metadata of similar data of another embodiment herein.
Fig. 7 is the schematic diagram for being used to generate the system of forecast model for illustrating an embodiment according to the application, its
In by software come implement the present invention function.
Embodiment
It is contemplated for carrying out the present invention's with detailed reference to some particulars of the present invention, including by inventor
Optimal mode.The example of these particulars is illustrated in accompanying drawing.Although describe this with reference to these particulars
Invention, it is to be understood that the present invention is not restricted to described embodiment.On the contrary, it is intended to as may include such as
All alternatives, modification and equivalent in the spirit and scope of the present invention defined by the appended claims.It is described below
In, numerous specific details are elaborated to provide thorough understanding of the present invention.Can be in some in these no specific details
Or implement the present invention in the case of whole.In other examples, well-known process operation is not described in detail so as to unnecessary
Ground obscures the present invention.
Term used herein is only used for describing the purpose of particular and being not intended to be limiting the present invention.Such as this
Used in text, unless the context clearly indicates otherwise, otherwise singulative " one " and " described/to be somebody's turn to do " are also intended to comprising plural shape
Formula.It will be further understood that, when used in this specification, term includes providing stated feature, integer, step, operation, member
The presence of part and/or part, but be not precluded from other one or more features, integer, step, operation, element, part and/or
The presence or addition of its group.
Fig. 1 is the schematic diagram of the block diagram for the equipment 1000 for illustrating an embodiment according to the application, and the equipment is used
In generation forecast model to predict crowd density figure and counting.As indicated, equipment 1000 may include density map creating unit 10,
CNN generation units 20, set of metadata of similar data acquiring unit 30 and model fine-adjusting unit 40.
Fig. 2 is the general signal of the flowchart process 2000 for the equipment 1000 for illustrating an embodiment according to the application
Figure.In step s201, true value density map creating unit 10 is operated to be selected from one or more of training set training image frame
Image block is selected, and the true value pedestrian in the true value Crowds Distribute and selected image block in image block selected by determination amounts to
Number.In step s202, CNN training units 20 are operated with one or more crowd's image blocks by inputting the frame in training set
To train CNN, wherein in the crowd's image block inputted, each crowd's image block has predetermined true value Density Distribution
Counted with personnel.In step s203, set of metadata of similar data acquiring unit 30 is operated to be sampled from target scene image set to frame, simultaneously
The training image with identified true value Density Distribution and counting/number is received from training set, and it is each for being sampled
Target image frame obtains similar view data from the training frames received, to overcome target scene image set and training image
Between scene gap.In step s204, model fine-adjusting unit 40 operate with by by similar view data be input to CNN come
CNN is finely adjusted, to determine the forecast model for the crowd density figure in prognostic chart picture frame and personnel's counting.Hereafter will
The conjunction of density map creating unit 10, CNN generation units 20, set of metadata of similar data acquiring unit 30 and model fine-adjusting unit 40 is discussed in detail
Make.
1) density map creating unit 10
Initial input is training set to equipment 100 (that is, inputting to density map creating unit 10), and it is included from various prisons
Control a certain amount of frame of video of the camera capture with pedestrian's number of people label.Density map creating unit 10 is operated with based on being inputted
Training set export the density map of each frame of video and counting.
Fig. 3 is the signal of the flowchart process for the density map creating unit 10 for illustrating an embodiment according to the application
Figure.In step s301, density map creating unit 10 operate with the perspective view of each Training scene/frame of the approximation from training set/
Distribution.The mark row people number of people is to indicate everyone number of people position in the area-of-interest of each training frames.In the sign number of people
In the case of position, the locus of pedestrian and human bodily form will be located in each frame.At step s302, the sky based on pedestrian
Between the perspective distortion of position, human bodily form and image create true value density map/distribution, to determine each frame middle row people/people
The true value density of group simultaneously estimates that the personnel in each frame of training set in crowd count.Specifically, true value density map/distribution table
Show the Crowds Distribute in each frame, and pedestrian's sum is equal to the integration of density map/distribution.
Specifically, the main target of crowd's CNN models to be discussed later is study mapping F:X-D, wherein X be from
One group of low-level feature of training image extraction, and D is crowd density figure/distribution of image.Assuming that denote the position of each pedestrian
Put, then the locus based on pedestrian, human bodily form and, the perspective distortion of image create density map/distribution.From training image
In the image block that is randomly chosen be considered as training sample, and density map/distribution of correspondence image piecemeal is considered as crowd CNN
The true value of model, crowd's CNN models will be further discussed later.As auxiliary mark, pass through the integration to density map/distribution
To calculate total crowd's number in selected training image piecemeal.It should be noted that sum will be decimal and non-integer.
In the prior art, density map recurrence true value is defined as to the sum of the Gaussian kernel centered on the position of object.
This density map/distribution is suitable for characterizing the Density Distribution of round shape object (such as, cell and bacterium).However, this hypothesis
It may fail when speaking of the general pedestrian crowd not in birds-eye view of camera wherein.Row in common monitoring camera
The example of people has three obvious characteristics:1) pedestrian image in monitor video has different chis due to perspective distortion
Degree;2) bodily form of pedestrian is more closely similar to ellipse compared with circle;3) due to serious shielding, people's head and shoulders is judged in each position
Put important implications of the place with the presence or absence of pedestrian.The body part of pedestrian is insecure for mark people.In view of these characteristics,
Crowd density is combined to create by the several distribution with perspective normalization (perspective normalization)
Figure/distribution.
Perspective normalization is necessary for estimation pedestrian's yardstick.For each scene, some adults will be randomly chosen
Pedestrian, then from the beginning they are indicated into pin.Assuming that the average height of adult be 175cm (such as), then can be by linear
Return and carry out approximate perspective view M.Pixel value in perspective view M (p) represents:Number of pixels in image is represented in actual scene
Some distance (for example, 1 meter) of the opening position.If a pedestrian is indicated with H pixels, on the center of the pedestrian
Perspective view is M (p)=H/1.75, and then, linearly interpolation perspective view is all saturating to obtain along the vertical and horizontal directions respectively for it
View.After perspective view/distribution of pedestrian's number of people Ph in obtaining area-of-interest (ROI) and center, according to
Lower rule creates crowd density figure/distribution:
Crowd density distribution core includes two:Normalization 2 as number of people part ties up Gaussian kernel Nh and as body part
Bivariate normal distribution Nb.Herein, Pb is the position of pedestrian body, and it is drawn by number of people position and perspective value estimation.For most
Pedestrian contour is represented goodly, and variance is set as (being directed to item Nh, and Nx=0.2M (p)),(being directed to item Nb).To ensure that the integration of all density values in density map/distribution is equal in original image
Total crowd's number, overall distribution is normalized by Z.
In short, for the number of people position with sign everyone, body build Density Distribution or core will be determined (hereafter
In be referred to as " core "), as described in formula (1).All body build cores of the personnel of (overlapping) all signs are combined to be formed
True value density map/distribution of each frame.The value that true value density map/distributed median is put is bigger, and crowd density is got in these positions
It is high.Further, since the normalized value of each body build core is equal to 1, so personnel's counting in crowd will be close equal in true value
Spend the sum of all values of body build core in figure/distribution.
2) CNN generation units 20
CNN generation units 20 are configured to first crowd's convolutional neural networks (CNN) of construction and initialization.Generation unit 20 is grasped
Make with acquisition/sampling crowd's image block from the frame in training set, and obtain the correspondence in sampled crowd's image block
True value density map and number (as determined by unit 10).Then, crowd's image that generation unit 20 will sample from training set
Piecemeal and its corresponding true value density map/distribution and number are input to as target objectivity (target objectiveness)
In CNN, to train CNN.
Fig. 4 is the flow for being used to generate and train CNN process 4000 for illustrating an embodiment according to the application
The schematic diagram of figure.
As indicated, in step s401, process 300 from frame sample one or more crowd's image block in training set,
And obtain the corresponding true value density map and number/crowd's number in sampled crowd's image block.Input is from training image
The image block of cutting.In order to obtain the pedestrian under similar scale, size of each image block at diverse location is basis
The perspective value of its center pixel carrys out selection.In this example, the 3X3 that each image block can be set in covering actual scene
Square metre.Then, as follows by image block distort (warp) to 72 (such as) pixel X72 (such as) pixel using as in step
The crowd's CNN models generated in 302.
In step s402, process 4000 is based on gaussian random and is distributed randomly to initialize crowd's convolutional neural networks.Fig. 5
In show the general introductions of crowd's CNN models with changeable target.
As indicated, crowd CNN models 500 include 3 convolutional layers (con1 to conv3) and three full articulamentum (fc4, fc5
With fc6 or fc7).Conv1 has 32 7X7X3 wave filters, and conv2 has 32 7X7X32 wave filters, and last convolution
Layer has 64 5X5X32 wave filters.After conv1 and conv2, the maximum pond layer with 2X2 core sizes is used.In Fig. 5
Unshowned amendment linear unit (ReLU) is the activation primitive applied after each convolutional layer and full articulamentum.It will be appreciated that
For purposes of illustration herein only by the number of wave filter and the number of layer description as an example, and the application be not limited to this
Some specific numbers and other numbers will be acceptable.
In step s403, process 400 learns from crowd image block to density map/mapping of distribution, such as by using
Small lot gradient declines and backpropagation is until that density map/convergence in distribution is created in such as by true value density map creating unit 10 is true
It is worth density/distribution.In step s404, process 400 switches target, learns the mapping from crowd image block to counting, Zhi Daosuo
The counting of study converges on the counting estimated by true value density map creating unit 10.In step 405, it is determined that estimated density
Whether figure/distribution and counting converge on true value, if it not, then repeat step s403 to s405.Hereinafter, it will be discussed in detail step
S403 to s405.
In the embodiment of the application, it introduces iteration handoff procedure in crowd CNN models 500, and it is used to replace
Ground optimizes density map/distribution estimation task and counts estimation task.The main task of crowd CNN models 400 is estimation input
Image block crowd density figure/distribution.In embodiment as shown in Figure 5, due to having two in CNN models 500
Individual pond layer, so density map/distribution of output is downsampled 18X18.Therefore, true value density map/distribution is also downward
It is sampled to 18X18.Because density map/distribution includes abundant partial detailed information, so CNN models 500 can benefit from study
Predicted density figure/distribution and the more preferable expression that crowd's image block can be obtained.The tale of the image block of input is returned
Return and be considered as the second task, it to density map image block by quadraturing to calculate.Two tasks are alternately mutually helped
Help and obtain more preferable solution.Two loss functions are defined according to following rule:
Wherein Θ is one group of parameter of CNN models, and N is the number of training sample.LDIt is estimated density map Fd (Xi;
Θ) the loss between (fc6 output) and true value density map Di.Similarly, LYIt is estimated crowd number Fy (Xi;Θ)(fc7
Output) with true value number YiBetween loss.Euclidean distance is used in both objective loss.Use a small amount of gradients
Decline and backpropagation is lost to minimize.
Changeable training program is outlined in algorithm 1.LDFirst object loss to be minimized is set to, because close
More spatial informations can be incorporated into CNN models by degree figure/distribution, so density map/distribution estimation needs model 500 to learn people
The general expression of group.After first object convergence, model 500 is switched over to minimize the global target for counting recurrence.
It is an easier task to count recurrence, and learns the task of its specific density figure/distribution recurrence faster.It should be noted that should be by
Photograph like or identical yardstick normalize the two target loss;Otherwise, the target with more large scale will be in training process
Middle dominance.Can be 10 by the yardstick weight setting of density loss, and can incite somebody to action in the embodiment of the application
The yardstick weight setting of counting loss is 1.Training is true in about 6 post-concentrations for switching iteration.Proposed switching study
Method can reach than widely used multi-task learning method better performance.
3) set of metadata of similar data acquiring unit 30
Set of metadata of similar data acquiring unit 30 is configured to:Sample frame is received from target scene, and is received from training set with by list
True value density map/the distribution and the sample of counting that member 10 creates;Then similarity number is obtained from training set for each target scene
According to overcome scene gap.
Crowd CNN models 500 are entered by proposed changeable learning process based on all Training scene data
Row pre-training.However, the crowd's scene each inquired about has its unique scene property, such as different visual angles, yardstick and not
Same Density Distribution.These properties significantly change the outward appearance of crowd's image block, and influence the performance of crowd CNN models 500.
In order to bridge the distribution gap between Training scene and test scene, nonparametric trimming scheme is designed to the CNN for crossing pre-training
Model 500 is suitable for the target scene that can't see.
The given target video from the scene that can't see, the sample with similar quality is obtained from training frames, and will
They are added to training data to be finely adjusted to crowd CNN models 500.Acquisition task is made up of two steps:Alternate scenes
Obtain and local image block obtains.
Alternate scenes obtain (step 601)The visual angle of scene and yardstick are the principal elements of influence crowd's outward appearance.Perspective view/
Distribution can indicate both visual angle and yardstick.To overcome the scale gap between different scenes, by the image block of each input
Be normalized to same yardstick, the yardstick in perspective view/distribution covering actual scene 3X3 square metres (such as).Therefore, it is non-
The first step of small parameter perturbations method concentrate on from all Training scenes obtain have the perspective view similar to target scene/point
The Training scene of cloth.The scene that those are acquired is referred to as candidate and finely tunes scene.Perspective descriptor is designed to represent each field
The visual angle of scape.Because perspective view/distribution is linearly fitted along y-axis, so its vertical gradient Δ My=M (y)-M (y-1) can be used as
Have an X-rayed descriptor.Based on the descriptor, for the scene that can't see, concentrated from whole training data and obtain top (for example, 20
It is individual) perspective view similar scene.Image in the scene being acquired is considered as to the alternate scenes obtained for topography's piecemeal.
Topography's piecemeal obtains (step 602)Second step is the selection similar image piecemeal from alternate scenes, these
Image block has the Density Distribution similar to the Density Distribution in test scene.In addition to visual angle and yardstick, crowd density point
Cloth has an effect on the skin mode of crowd.The higher crowd of density have it is even more serious block, and may only observe the number of people and
Shoulder.On the contrary, in sparse crowd, complete body build is presented in pedestrian.Therefore, set of metadata of similar data acquiring unit 30 is configured to
Attempt the Density Distribution of prediction target scene and obtain from alternate scenes to be matched with the similar of predicted target density distribution
Image block.For example, for the higher crowd's scene of density, the model that more dense image block should be obtained to be crossed to pre-training
It is finely adjusted and carrys out fit object scene.
The CNN models 500 crossed using the pre-training such as trained in the cell 20, we can calculate roughly target image
Each image block density and tale.Assuming that the image block with similar density map/distribution passes through pre-training mistake
Model 500 there is similar output.Based on prediction result, we calculate the histogram of the Density Distribution of target scene.Press
Each section (bin) is calculated according to following rule:
Wherein yiIt is the integral counting of sample i estimated density map/distribution.
Due to scene of the pedestrian station of wherein more than 20 in 3X3 square metres seldom be present, so working as yi>, should when 20
Image block is assigned to the 6th section (that is, ci=6).The Density Distribution of target scene can be obtained from equation (4).Then, from
It is acquired in Training scene and is randomly chosen image block, and controls the number of the different image block of density to be matched with
The Density Distribution of target scene.By this way, it is able to obtain with similar visual angle, chi using proposed method for trimming
The image block of degree and Density Distribution.
Model fine-adjusting unit 40
Model fine-adjusting unit 40 is configured to receive the set of metadata of similar data being acquired and utilizes the set of metadata of similar data to pre-training
The CNN 500 crossed is finely adjusted, so that CNN 500 can predict the density in the area-of-interest of frame of video to be detected
Figure/distribution and pedestrian counting and the area-of-interest.Trimmed crowd CNN models have reached more preferable for target scene
Performance.
In the embodiment of the application, fine-adjusting unit 40 samples the similar image piecemeal obtained from unit 30, and
The similar image piecemeal obtained is input to CNN that pre-training crosses to be finely adjusted to it (for example, by using small lot ladder
Degree decline and backpropagation until density map/convergence in distribution in the true value density such as created by true value density map creating unit 10/
Distribution).Then, fine-adjusting unit 40 switches objectivity and learns the mapping from crowd image block to counting, until what is learnt
Count the counting for converging on and being estimated by true value density map creating unit 10.Finally, it is determined that estimated density map/distribution and counting
Whether true value is converged on, if it not, then repeating above step.
The trimmed forecast model generated by model fine-adjusting unit 40 can receive frame of video to be detected and region of interest
Domain, then predict the estimated density map and pedestrian counting in area-of-interest.
Such as by it will be apparent to those skilled in the art that system, method or computer program product can the invention is embodied as.Cause
This, the present invention can use complete hardware embodiment and hardware aspect (it can be typically referred to as to " unit ", " electricity herein
Road ", " module " or " system ").The major part of invention sexual function and many invention principles are when realizing most preferably by integrated circuit
(IC) support, such as digital signal processor and therefore software or application-specific integrated circuit.While it may be possible to pay great efforts and many
Design alternative by (such as) available time, current technology and economic consideration drive, but still expect ordinary skill people
It will readily be able under the guiding of member concept and principle disclosed herein and produce IC with minimum experiment.Therefore, in order to succinct
Any risk of fuzzy principles and concepts according to the present invention is simultaneously preferably minimized by property, to such software and IC (if any)
Be discussed further be limited to regard to the principle as used in preferred embodiment and the key element for concept.
In addition, the present invention can use complete software embodiment (including firmware, resident software, microcode etc.) or with reference to software
Embodiment.In addition, the present invention can use the form for the computer program product being embodied in any tangible performance media,
The performance media have the computer usable program code being embodied in the media.Fig. 7 illustrates to be used to generate forecast model
The system 7000 counted with the crowd density distribution in prognostic chart picture frame and personnel.System 7000 includes:Memory 3001, it is deposited
The executable part of storage;And processor 3002, it is electrically coupled to memory 3001 to perform executable part, with execution system
3000 operation.These executable parts may include:True value density map creates part 701, CNN training components 702, set of metadata of similar data
Obtaining widget 703 and model trimming part 704.
True value density map creates part 701 and is disposed for:Selected from one or more of training set training image frame
Select image block;And the true value pedestrian in the true value Crowds Distribute and selected image block in image block selected by determining amounts to
Number.CNN training components 702 be disposed for by input one or more crowd's image blocks of frame in training set come
CNN is trained, in the crowd's image block inputted, each crowd's image block has predetermined true value Density Distribution and personnel
Count.
Set of metadata of similar data obtaining widget 703 is configured to sample frame from target scene image set and receive have institute really from training set
Fixed true value Density Distribution and the training image of counting/number;And each target image frame for being sampled is from being received
Training frames in obtain similar view data, to overcome the scene gap between target scene image set and training image.
Model trimming part 703 is disposed for by the way that similar view data is input into CNN 500 to be carried out to CNN
Fine setting, to determine the forecast model for the crowd density figure and personnel's counting being used in prognostic chart picture frame.
The function of part 701 to 704 is analogous respectively to the function of unit 10 to 40, and therefore omits it herein and retouch in detail
State.
Although having been described for the preferred exemplary of the present invention, those skilled in the art can know that basic invention is general
At once these examples are made with change or modification after thought.Appended claims are intended to be considered as including preferred exemplary and all changes
Change or modification is all fallen within the scope of the present invention.
Claims (22)
1. a kind of be used to generate the method that forecast model is counted with the crowd density distribution in prognostic chart picture frame and personnel, it is wrapped
Include:
CNN is trained by inputting one or more crowd's image blocks of the frame in training set, is schemed in the crowd of the input
As in piecemeal, there are each crowd's image block predetermined true value Density Distribution and personnel to count;
Frame, which is sampled, from target scene image set and received from the training set has the predetermined true value Density Distribution
With the training image of counting/number;
Similar view data is obtained from the training frames received for each target image frame sampled, it is described to overcome
Scene gap between target scene image set and the training image;And
By the way that the similar view data is input into the CNN to be finely adjusted to the CNN, to determine to be used for predict
The forecast model that crowd density figure and personnel in picture frame count.
2. according to the method for claim 1, further comprise:
Image block is selected from one or more of described training set training image frame;And in image block selected by determining
True value Crowds Distribute and selected image block in true value pedestrian's tale.
3. according to the method for claim 2, wherein the determination further comprises:
Each personnel of number of people position of the identification with sign in each picture frame;
Determine the body build core of each personnel identified;And
All identified body build cores are combined to form the true value of each frame density map/distribution, wherein in the crowd
The counting of the personnel is equal to the sum of all values of body build core described in the true value density map/distribution.
4. according to the method for claim 1, wherein the training further comprises:
The CNN is randomly initialized based on gaussian random distribution;
From the training image sampled picture piecemeal;
Estimate that the pedestrian in the Crowds Distribute in sampled image block and the image block sampled is total by the CNN
Number;
The parameter of the CNN is updated, until estimated convergence in distribution is distributed in the true value;And
Further update the parameter of the CNN, converged on until estimated number determined by true value number, so as to obtain warp
Cross the CNN of pre-training.
5. according to the method for claim 4, wherein the acquisition set of metadata of similar data further comprises:
The candidate with the perspective distribution similar to the target image frame is obtained from the training image frame and finely tunes frame data;
And
The selection similar image with the Density Distribution similar to the Density Distribution of the target image frame point from alternate scenes
Block.
6. according to the method for claim 5, wherein, the fine setting further comprises:
From the similar image piecemeal sampled picture piecemeal;
Divided by the CNN by pre-training come the image estimated the Crowds Distribute in sampled image block and sampled
Pedestrian's sum in block;
The parameter of the CNN is updated, until the estimated convergence in distribution is distributed in the true value;And
The parameter for the CNN that the pre-training is crossed further is updated, until estimated number converges on the identified true value
Number, to obtain trimmed CNN.
7. the method according to any one of claim 1 to 6, wherein, pass through the product to identified true value Density Distribution
Divide and counted to determine the tale of the personnel in described image frame.
8. the method according to any one of claim 1 to 6, wherein, the space bit based on the pedestrian in each picture frame
Put, the perspective distortion of the human bodily form in each picture frame and image creates the Crowds Distribute.
9. a kind of be used to generate the equipment that forecast model is counted with the crowd density distribution in prognostic chart picture frame and personnel, it is wrapped
Include:
CNN training units (20), trained by inputting one or more crowd's image blocks of the frame in training set
CNN, in the crowd's image block inputted, each crowd's image block has predetermined true value Density Distribution and personnel
Count;
Set of metadata of similar data acquiring unit (30), is sampled from target scene image set to frame, and has from training set reception
The training image of identified true value Density Distribution and counting/number, and each target image frame for being sampled is from institute
Similar view data is obtained in the training frames of reception, to overcome between the target scene image set and the training image
Scene gap;And
Model fine-adjusting unit (40), it is micro- to be carried out to the CNN by the way that the similar view data is input into the CNN
Adjust, to determine the forecast model for the crowd density figure in prognostic chart picture frame and personnel's counting.
10. equipment according to claim 9, it further comprises:
True value density map creating unit (10), image point is selected from one or more of described training set training image frame
Block;And true value pedestrian's tale in the true value Crowds Distribute and selected image block in image block selected by determining.
11. equipment according to claim 10, wherein the true value density map creating unit (10) is configured to by following
Step is come true value pedestrian's tale in the true value Crowds Distribute and the selected image block in image block selected by determining:
Each personnel of number of people position of the identification with sign in each picture frame;
Determine the body build core of each personnel identified;And
All identified body build cores are combined to form the true value of each frame density map/distribution, wherein in the crowd
The counting of the personnel is equal to the sum of all values of body build core described in the true value density map/distribution.
12. equipment according to claim 9, wherein the CNN training units (20) trained by following steps it is described
CNN:
The CNN is randomly initialized based on gaussian random distribution;
From the training image sampled picture piecemeal;
Estimate that the pedestrian in the Crowds Distribute in sampled image block and the image block sampled is total by the CNN
Number;
The parameter of the CNN is updated, until estimated convergence in distribution is distributed in the true value;And
Further update the parameter of the CNN, converged on until estimated number determined by true value number, to obtain process
The CNN of pre-training.
13. equipment according to claim 12, wherein the set of metadata of similar data acquiring unit (30) is configured to:
The candidate with the perspective distribution similar to the target image frame is obtained from the training image frame and finely tunes frame data;
And
The selection similar image with the Density Distribution similar to the Density Distribution of the target image frame point from alternate scenes
Block.
14. equipment according to claim 13, wherein the fine-adjusting unit is further configured to be used for:
From the similar image piecemeal sampled picture piecemeal;
The CNN crossed by the pre-training is come the image block estimating the Crowds Distribute in sampled image block He sampled
In pedestrian sum;
The parameter of the CNN is updated, until estimated convergence in distribution is distributed in the true value;And
Further update the parameter for the CNN that the pre-training is crossed, converged on until estimated number determined by true value number,
To obtain trimmed CNN.
15. the equipment according to any one of claim 9 to 14, wherein, by identified true value Density Distribution
Integration is counted to determine the tale of the personnel in described image frame.
16. the equipment according to any one of claim 9 to 14, wherein, the space based on the pedestrian in each picture frame
The perspective distortion of position, the human bodily form in each picture frame and image creates the Crowds Distribute.
17. a kind of be used to generate the system that forecast model is counted with the crowd density distribution in prognostic chart picture frame and personnel, it is wrapped
Include:
Memory, it stores executable part;And
Processor, it is electrically coupled to the memory to perform the executable part to perform the operation of the system, wherein,
The executable part includes:
CNN training components, CNN is trained by inputting one or more crowd's image blocks of the frame in training set, it is defeated in institute
In the crowd's image block entered, there are each crowd's image block predetermined true value Density Distribution and personnel to count;
Set of metadata of similar data obtaining widget, frame is sampled from target scene image set, has from training set reception and determined
True value Density Distribution and counting/number training image, and each target image frame for being sampled is from being received
Similar view data is obtained in training frames, to overcome between the scene between the target scene image set and the training image
Gap;And
Model trimming part, by the way that the similar view data is input into the CNN to be finely adjusted to the CNN, with
Determine the forecast model for the crowd density figure in prognostic chart picture frame and personnel's counting.
18. system according to claim 17, it further comprises:
True value density map creates part, and image block is selected from one or more of described training set training image frame;With
And true value pedestrian's tale in the true value Crowds Distribute and selected image block in image block selected by determining.
19. system according to claim 17, it is configured to pass through following steps wherein the true value density map creates part
Come the true value pedestrian tale in the true value Crowds Distribute and selected image block in image block selected by determining:
Each personnel of number of people position of the identification with sign in each picture frame;
Determine the body build core of each personnel identified;And
All identified body build cores are combined to form the true value of each frame density map/distribution, wherein in the crowd
The counting of the personnel is equal to the sum of all values of body build core described in the true value density map/distribution.
20. system according to claim 17, wherein the CNN training components train the CNN by following steps:
The CNN is randomly initialized based on gaussian random distribution;
From the training image sampled picture piecemeal;
Estimate that the pedestrian in the Crowds Distribute in sampled image block and the image block sampled is total by the CNN
Number;
The parameter of the CNN is updated, until estimated convergence in distribution is distributed in the true value;And
Further update the parameter of the CNN, converged on until the number of the estimation determined by true value number, it is pre- to obtain
The CNN trained.
21. system according to claim 20, wherein the set of metadata of similar data obtaining widget is configured to:
The candidate with the perspective distribution similar to the target image frame is obtained from the training image frame and finely tunes frame data;
And
The selection similar image with the Density Distribution similar to the Density Distribution of the target image frame point from alternate scenes
Block.
22. system according to claim 21, wherein the trimming part is further configured to be used for:
From the similar image piecemeal sampled picture piecemeal;
The CNN crossed by the pre-training is come the image block estimating the Crowds Distribute in sampled image block He sampled
In pedestrian sum;
The parameter for the CNN that the pre-training is crossed is updated, until estimated convergence in distribution is distributed in the true value;And
Further update the parameter of the CNN, until the estimated number converge on it is described determined by true value number, with
Obtain trimmed CNN.
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2015/079178 WO2016183766A1 (en) | 2015-05-18 | 2015-05-18 | Method and apparatus for generating predictive models |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107624189A true CN107624189A (en) | 2018-01-23 |
CN107624189B CN107624189B (en) | 2020-11-20 |
Family
ID=57319199
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201580080145.XA Active CN107624189B (en) | 2015-05-18 | 2015-05-18 | Method and apparatus for generating a predictive model |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN107624189B (en) |
WO (1) | WO2016183766A1 (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109034355A (en) * | 2018-07-02 | 2018-12-18 | 百度在线网络技术(北京)有限公司 | Number prediction technique, device, equipment and the storage medium of fine and close crowd |
CN109409318A (en) * | 2018-11-07 | 2019-03-01 | 四川大学 | Training method, statistical method, device and the storage medium of statistical model |
CN109447008A (en) * | 2018-11-02 | 2019-03-08 | 中山大学 | Population analysis method based on attention mechanism and deformable convolutional neural networks |
CN109815936A (en) * | 2019-02-21 | 2019-05-28 | 深圳市商汤科技有限公司 | A kind of target object analysis method and device, computer equipment and storage medium |
CN110197502A (en) * | 2019-06-06 | 2019-09-03 | 山东工商学院 | A kind of multi-object tracking method that identity-based identifies again and system |
CN111340801A (en) * | 2020-03-24 | 2020-06-26 | 新希望六和股份有限公司 | Livestock checking method, device, equipment and storage medium |
CN112990530A (en) * | 2020-12-23 | 2021-06-18 | 北京软通智慧城市科技有限公司 | Regional population number prediction method and device, electronic equipment and storage medium |
Families Citing this family (35)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10455363B2 (en) * | 2015-11-04 | 2019-10-22 | xAd, Inc. | Systems and methods for using geo-blocks and geo-fences to discover lookalike mobile devices |
US10547971B2 (en) | 2015-11-04 | 2020-01-28 | xAd, Inc. | Systems and methods for creating and using geo-blocks for location-based information service |
CN107566781B (en) * | 2016-06-30 | 2019-06-21 | 北京旷视科技有限公司 | Video monitoring method and video monitoring equipment |
CN106997459B (en) * | 2017-04-28 | 2020-06-26 | 成都艾联科创科技有限公司 | People counting method and system based on neural network and image superposition segmentation |
CN108875456B (en) * | 2017-05-12 | 2022-02-18 | 北京旷视科技有限公司 | Object detection method, object detection apparatus, and computer-readable storage medium |
CN107330364B (en) * | 2017-05-27 | 2019-12-03 | 上海交通大学 | A kind of people counting method and system based on cGAN network |
CN107563349A (en) * | 2017-09-21 | 2018-01-09 | 电子科技大学 | A kind of Population size estimation method based on VGGNet |
CN107657226B (en) * | 2017-09-22 | 2020-12-29 | 电子科技大学 | People number estimation method based on deep learning |
CN107609597B (en) * | 2017-09-26 | 2020-10-13 | 嘉世达电梯有限公司 | Elevator car number detection system and detection method thereof |
EP3704558A4 (en) | 2017-11-01 | 2021-07-07 | Nokia Technologies Oy | Depth-aware object counting |
CN107977025A (en) * | 2017-11-07 | 2018-05-01 | 中国农业大学 | A kind of regulator control system and method for industrialized aquiculture dissolved oxygen |
CN108154089B (en) * | 2017-12-11 | 2021-07-30 | 中山大学 | Size-adaptive-based crowd counting method for head detection and density map |
CN108615027B (en) * | 2018-05-11 | 2021-10-08 | 常州大学 | Method for counting video crowd based on long-term and short-term memory-weighted neural network |
CN109117791A (en) * | 2018-08-14 | 2019-01-01 | 中国电子科技集团公司第三十八研究所 | A kind of crowd density drawing generating method based on expansion convolution |
US11134359B2 (en) | 2018-08-17 | 2021-09-28 | xAd, Inc. | Systems and methods for calibrated location prediction |
US10349208B1 (en) | 2018-08-17 | 2019-07-09 | xAd, Inc. | Systems and methods for real-time prediction of mobile device locations |
US11172324B2 (en) | 2018-08-17 | 2021-11-09 | xAd, Inc. | Systems and methods for predicting targeted location events |
CN109635634B (en) * | 2018-10-29 | 2023-03-31 | 西北大学 | Pedestrian re-identification data enhancement method based on random linear interpolation |
CN111191667B (en) * | 2018-11-15 | 2023-08-18 | 天津大学青岛海洋技术研究院 | Crowd counting method based on multiscale generation countermeasure network |
CN111291587A (en) * | 2018-12-06 | 2020-06-16 | 深圳光启空间技术有限公司 | Pedestrian detection method based on dense crowd, storage medium and processor |
CN110826496B (en) * | 2019-11-07 | 2023-04-07 | 腾讯科技(深圳)有限公司 | Crowd density estimation method, device, equipment and storage medium |
US11106904B2 (en) * | 2019-11-20 | 2021-08-31 | Omron Corporation | Methods and systems for forecasting crowd dynamics |
CN110942015B (en) * | 2019-11-22 | 2023-04-07 | 上海应用技术大学 | Crowd density estimation method |
CN111062275A (en) * | 2019-12-02 | 2020-04-24 | 汇纳科技股份有限公司 | Multi-level supervision crowd counting method, device, medium and electronic equipment |
CN111178235A (en) * | 2019-12-27 | 2020-05-19 | 卓尔智联(武汉)研究院有限公司 | Target quantity determination method, device, equipment and storage medium |
CN111274900B (en) * | 2020-01-15 | 2021-01-01 | 北京航空航天大学 | Empty-base crowd counting method based on bottom layer feature extraction |
CN111626141B (en) * | 2020-04-30 | 2023-06-02 | 上海交通大学 | Crowd counting model building method, counting method and system based on generated image |
CN111652168B (en) * | 2020-06-09 | 2023-09-08 | 腾讯科技(深圳)有限公司 | Group detection method, device, equipment and storage medium based on artificial intelligence |
CN112001274B (en) * | 2020-08-06 | 2023-11-17 | 腾讯科技(深圳)有限公司 | Crowd density determining method, device, storage medium and processor |
CN111898578B (en) * | 2020-08-10 | 2023-09-19 | 腾讯科技(深圳)有限公司 | Crowd density acquisition method and device and electronic equipment |
CN113822111A (en) * | 2021-01-19 | 2021-12-21 | 北京京东振世信息技术有限公司 | Crowd detection model training method and device and crowd counting method and device |
CN112801018B (en) * | 2021-02-07 | 2023-07-07 | 广州大学 | Cross-scene target automatic identification and tracking method and application |
CN113033342A (en) * | 2021-03-10 | 2021-06-25 | 西北工业大学 | Crowd scene pedestrian target detection and counting method based on density estimation |
CN113269224B (en) * | 2021-03-24 | 2023-10-31 | 华南理工大学 | Scene image classification method, system and storage medium |
CN115293465B (en) * | 2022-10-09 | 2023-02-14 | 枫树谷(成都)科技有限责任公司 | Crowd density prediction method and system |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7991193B2 (en) * | 2007-07-30 | 2011-08-02 | International Business Machines Corporation | Automated learning for people counting systems |
CN103971100A (en) * | 2014-05-21 | 2014-08-06 | 国家电网公司 | Video-based camouflage and peeping behavior detection method for automated teller machine |
CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
CN104573744A (en) * | 2015-01-19 | 2015-04-29 | 上海交通大学 | Fine granularity classification recognition method and object part location and feature extraction method thereof |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8195598B2 (en) * | 2007-11-16 | 2012-06-05 | Agilence, Inc. | Method of and system for hierarchical human/crowd behavior detection |
CN104268524A (en) * | 2014-09-24 | 2015-01-07 | 朱毅 | Convolutional neural network image recognition method based on dynamic adjustment of training targets |
-
2015
- 2015-05-18 WO PCT/CN2015/079178 patent/WO2016183766A1/en active Application Filing
- 2015-05-18 CN CN201580080145.XA patent/CN107624189B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7991193B2 (en) * | 2007-07-30 | 2011-08-02 | International Business Machines Corporation | Automated learning for people counting systems |
CN103971100A (en) * | 2014-05-21 | 2014-08-06 | 国家电网公司 | Video-based camouflage and peeping behavior detection method for automated teller machine |
CN104077613A (en) * | 2014-07-16 | 2014-10-01 | 电子科技大学 | Crowd density estimation method based on cascaded multilevel convolution neural network |
CN104573744A (en) * | 2015-01-19 | 2015-04-29 | 上海交通大学 | Fine granularity classification recognition method and object part location and feature extraction method thereof |
Non-Patent Citations (2)
Title |
---|
M.SZARVAS 等: "Pedestrian detection with convolutional neural networks", 《IEEE PROCEEDINGS. INTELLIGENT VEHICLES SYMPOSIUM,2005》 * |
覃勋辉: "多种人群密度场景下的人群计数", 《中国图象图形学报》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109034355A (en) * | 2018-07-02 | 2018-12-18 | 百度在线网络技术(北京)有限公司 | Number prediction technique, device, equipment and the storage medium of fine and close crowd |
US11302104B2 (en) | 2018-07-02 | 2022-04-12 | Baidu Online Network Technology (Beijing) Co., Ltd. | Method, apparatus, device, and storage medium for predicting the number of people of dense crowd |
CN109034355B (en) * | 2018-07-02 | 2022-08-02 | 百度在线网络技术(北京)有限公司 | Method, device and equipment for predicting number of people in dense crowd and storage medium |
CN109447008A (en) * | 2018-11-02 | 2019-03-08 | 中山大学 | Population analysis method based on attention mechanism and deformable convolutional neural networks |
CN109409318A (en) * | 2018-11-07 | 2019-03-01 | 四川大学 | Training method, statistical method, device and the storage medium of statistical model |
CN109815936A (en) * | 2019-02-21 | 2019-05-28 | 深圳市商汤科技有限公司 | A kind of target object analysis method and device, computer equipment and storage medium |
CN109815936B (en) * | 2019-02-21 | 2023-08-22 | 深圳市商汤科技有限公司 | Target object analysis method and device, computer equipment and storage medium |
CN110197502A (en) * | 2019-06-06 | 2019-09-03 | 山东工商学院 | A kind of multi-object tracking method that identity-based identifies again and system |
CN111340801A (en) * | 2020-03-24 | 2020-06-26 | 新希望六和股份有限公司 | Livestock checking method, device, equipment and storage medium |
CN112990530A (en) * | 2020-12-23 | 2021-06-18 | 北京软通智慧城市科技有限公司 | Regional population number prediction method and device, electronic equipment and storage medium |
CN112990530B (en) * | 2020-12-23 | 2023-12-26 | 北京软通智慧科技有限公司 | Regional population quantity prediction method, regional population quantity prediction device, electronic equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN107624189B (en) | 2020-11-20 |
WO2016183766A1 (en) | 2016-11-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107624189A (en) | Method and apparatus for generating forecast model | |
Mou et al. | IM2HEIGHT: Height estimation from single monocular imagery via fully residual convolutional-deconvolutional network | |
Mukhoti et al. | Evaluating bayesian deep learning methods for semantic segmentation | |
Rafi et al. | An Efficient Convolutional Network for Human Pose Estimation. | |
US20180114071A1 (en) | Method for analysing media content | |
CN103020606B (en) | Pedestrian detection method based on spatio-temporal context information | |
CN109766830A (en) | A kind of ship seakeeping system and method based on artificial intelligence image procossing | |
EP2131328A2 (en) | Method for automatic detection and tracking of multiple objects | |
JP6397379B2 (en) | CHANGE AREA DETECTION DEVICE, METHOD, AND PROGRAM | |
CN110555481A (en) | Portrait style identification method and device and computer readable storage medium | |
JP2017191501A (en) | Information processing apparatus, information processing method, and program | |
CN110765833A (en) | Crowd density estimation method based on deep learning | |
CN109727270A (en) | The movement mechanism and analysis of texture method and system of Cardiac Magnetic Resonance Images | |
CN110399882A (en) | A kind of character detecting method based on deformable convolutional neural networks | |
Sun et al. | Global Mask R-CNN for marine ship instance segmentation | |
CN109063549A (en) | High-resolution based on deep neural network is taken photo by plane video moving object detection method | |
Han et al. | Dr. vic: Decomposition and reasoning for video individual counting | |
CN116229560A (en) | Abnormal behavior recognition method and system based on human body posture | |
Gao et al. | Road extraction using a dual attention dilated-linknet based on satellite images and floating vehicle trajectory data | |
CN113780145A (en) | Sperm morphology detection method, sperm morphology detection device, computer equipment and storage medium | |
CN111652181B (en) | Target tracking method and device and electronic equipment | |
CN115605914A (en) | Object detection method, object detection device and object detection system | |
CN110490170A (en) | A kind of face candidate frame extracting method | |
Li et al. | Fuzzy classification of high resolution remote sensing scenes using visual attention features | |
CN114972335A (en) | Image classification method and device for industrial detection and computer equipment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |