CN109858389A - Vertical ladder demographic method and system based on deep learning - Google Patents
Vertical ladder demographic method and system based on deep learning Download PDFInfo
- Publication number
- CN109858389A CN109858389A CN201910023341.3A CN201910023341A CN109858389A CN 109858389 A CN109858389 A CN 109858389A CN 201910023341 A CN201910023341 A CN 201910023341A CN 109858389 A CN109858389 A CN 109858389A
- Authority
- CN
- China
- Prior art keywords
- target
- vertical ladder
- detection
- dbl
- feature
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of vertical ladder demographic method and system based on deep learning, the following steps are included: vertical ladder monitoring camera has unmanned algorithm to judge someone by camera, switch gate algorithm judges that vertical ladder door is closed, and carriage starts analysis request in operating status, triggers an image and vertical ladder pedestrian detection algorithm is called to do target detection；After receiving analysis request, an image is taken from main bit stream according to the timing node of triggering image, and start that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end；Specific vertical ladder number is obtained by detection block, and quantity is written in database, while being reported to aerial ladder platform.
Description
Technical field
The invention belongs to technical field of computer vision, and in particular to a kind of vertical ladder demographics side based on deep learning
Method and system.
Background technique
In vertical ladder use process, oppressive phenomenon caused by unavoidable partial fault or human factor, by detection side
Method can detect current oppressive number, and notify that rescue personnel can save in time to property or aerial ladder platform.Vertical ladder is equipped with weight
Force snesor has certain load range, is easy to cause failure of apparatus more than this range, from security consideration, passes through detection
Vertical ladder seating capacity can rationally control vertical ladder load range.The flow of the people of statistics seating in list portion vertical ladder one day, for cell scenario,
It counts cell number and is used for big data analysis, there is certain forewarning function in terms of safety management；For school, this stream of hospital
Big scene is measured, data on flows is statisticallyd analyze, can be run by Optimized Operation vertical ladder, improves operational efficiency；For store, purchase
The scenes such as object center count data on flows, help advertisement position rational deployment, increase the income etc. of operation assets.
The Chinese Patent Application No. of the prior art is CN201710157587.0, applies for a kind of entitled demographic method
And device and elevator scheduling method and system, the invention is by obtaining the image of stereoscopic vision video camera shooting, transformed space three
Information is tieed up, two-dimension projection is obtained according to the height-width information of three-dimensional information.For the number of people target picture in two-dimension projection
Plain block takes connection field mark.It needs using Target Max as the center of circle, pre-set radius obtains number of people target area, with UNICOM domain face
It is long-pending to obtain number of people quantity with preset area threshold value comparison.Image processing algorithm belongs to traditional mode recognizer in the invention, needs
Want some features of engineer.Such as pre-set radius circle covers candidate region, environment has been under different elevator scenes
Difference or human factor bring camera position deviation may bring the deviation of preset value, therefore not have pervasive
Property.
Summary of the invention
The technical problem to be solved in the present invention is to provide a kind of vertical ladder demographic method and system based on deep learning.
In order to solve the above technical problems, the present invention adopts the following technical scheme that:
The one side of the embodiment of the present invention provides a kind of vertical ladder demographic method based on deep learning, including following
Step:
Vertical ladder monitoring camera has unmanned algorithm to judge someone by camera, and switch gate algorithm judges that vertical ladder door is closed, and
Carriage starts analysis request in operating status, triggers an image and vertical ladder pedestrian detection algorithm is called to do target detection；
After receiving analysis request, an image is taken from main bit stream according to the timing node of triggering image, and start to adjust
The detection block for analyzing to the end is carried out with YOLOv3 algorithm；
Specific vertical ladder number is obtained by detection block, and quantity is written in database, while being reported to aerial ladder platform.
It is preferably, described to start that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end specifically:
Vertical ladder pedestrian image is done into normalized first, size change over, then will be image gridding to 416*416, if inspection
The target's center of survey falls in some grid, then the grid is just responsible for predicting the target；YOLOv3 obtains 9 using K-means
A priori frame takes the characteristic pattern of three scales to predict, when input is 416*416, characteristic pattern is respectively 13*13,26*26,
52*52；Each yolo layers is divided on the characteristic pattern of 3 scales, scale using three priori frames according to the size of priori frame
Big characteristic pattern uses small priori frame；Each grid can predict multiple bounding box, if the network is pre- in training
The bounding box of survey and the ground truth lap of label are maximum, that is, can determine whether target in grid, grid is special
The prediction of target is responsible in duty；When part, a target may cause target frame to repeat to detect by repeated detection, non-very big
Value inhibits to can detecte the bounding box of high superposed, removes all prediction blocks other than confidence level highest, guarantees same target only
Export a detection block；All prediction blocks obtain specific coordinate information and class categories by recurrence；Pass through setting one
Objective degrees of confidence threshold value, wherein confidence level is lap of the probability of target category multiplied by prediction target and calibration true value, height
In the prediction block of threshold value be last detection block.
Preferably, the characteristic pattern of three scales is taken to predict, when input is 416*416, characteristic pattern is respectively 13*13,
26*26,52*52 are obtained using YOLOv3 detection network, specifically: YOLOv3 detection network has used 23 after first group of DBL
A residual unit, 6 groups of DBL, and increase convolutional layer and obtain detection of the feature output of 13*13 for big target；Wherein 23
A residual unit, Zuo Liao branch DBL and up-sampling obtain the feature of 26*26 and obtain by 19 residual units after 5 groups of DBL
26*26 feature has done tensor splicing, and wherein concat is used to splice the tensor of same scale, and spliced tensor passes through 5 groups of DBL
And it increases convolutional layer and obtains the feature output of 26*26；Wherein the 5 ZuDBLHou Zuo branch continues to be DBL obtains with up-sampling
The feature of 52*52, while tensor splicing is done with the 52*52 feature obtained by 11 residual units, the spliced tensor is again
By 6 groups of DBL and increase convolutional layer obtained 52*52 feature output be used for Small object detection；Final three layers of feature is defeated
It is provided commonly for the detection of vertical ladder pedestrian out.
The embodiment of the present invention further aspect is that provide a kind of vertical ladder passenger number statistical system based on deep learning, wrap
It includes:
Analysis request module has unmanned algorithm to judge someone, switch gate algorithm for vertical ladder monitoring camera by camera
Judge that vertical ladder door is closed, and carriage starts analysis request in operating status, triggers an image and calls vertical ladder pedestrian detection
Algorithm does target detection；
Vertical ladder demographics module, for after receiving analysis request, according to the timing node of triggering image from primary key
Stream takes an image, and starts that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end；
Uploading module is written, for obtaining specific vertical ladder number by detection block, and quantity is written in database, together
When be reported to aerial ladder platform.
It is preferably, described to start that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end specifically:
Vertical ladder pedestrian image is done into normalized first, size change over, then will be image gridding to 416*416, if inspection
The target's center of survey falls in some grid, then the grid is just responsible for predicting the target；YOLOv3 obtains 9 using K-means
A priori frame takes the characteristic pattern of three scales to predict, when input is 416*416, characteristic pattern is respectively 13*13,26*26,
52*52；Each yolo layers is divided on the characteristic pattern of 3 scales, scale using three priori frames according to the size of priori frame
Big characteristic pattern uses small priori frame；Each grid can predict multiple bounding box, if the network is pre- in training
The bounding box of survey and the ground truth lap of label are maximum, that is, can determine whether target in grid, grid is special
The prediction of target is responsible in duty；When part, a target may cause target frame to repeat to detect by repeated detection, non-very big
Value inhibits to can detecte the bounding box of high superposed, removes all prediction blocks other than confidence level highest, guarantees same target only
Export a detection block；All prediction blocks obtain specific coordinate information and class categories by recurrence；Pass through setting one
Objective degrees of confidence threshold value, wherein confidence level is lap of the probability of target category multiplied by prediction target and calibration true value, height
In the prediction block of threshold value be last detection block.
Preferably, the characteristic pattern of three scales is taken to predict, when input is 416*416, characteristic pattern is respectively 13*13,
26*26,52*52 are obtained using YOLOv3 detection network, specifically: YOLOv3 detection network has used 23 after first group of DBL
A residual unit, 6 groups of DBL, and increase convolutional layer and obtain detection of the feature output of 13*13 for big target；Wherein 23
A residual unit, Zuo Liao branch DBL and up-sampling obtain the feature of 26*26 and obtain by 19 residual units after 5 groups of DBL
26*26 feature has done tensor splicing, and wherein concat is used to splice the tensor of same scale, and spliced tensor passes through 5 groups of DBL
And it increases convolutional layer and obtains the feature output of 26*26；Wherein the 5 ZuDBLHou Zuo branch continues to be DBL obtains with up-sampling
The feature of 52*52, while tensor splicing is done with the 52*52 feature obtained by 11 residual units, the spliced tensor is again
By 6 groups of DBL and increase convolutional layer obtained 52*52 feature output be used for Small object detection；Final three layers of feature is defeated
It is provided commonly for the detection of vertical ladder pedestrian out.
Using the present invention with following the utility model has the advantages that the present invention passes through applied to aerial ladder platform and marks a large amount of vertical ladder rows
Personal data improves the detection of vertical ladder pedestrian target with improved YOLOv3 network training data set；And it is based on big data sample
The model that training obtains has good universality, adapts in various scenes, such as cell, school, hospital, hotel, market.
Detailed description of the invention
Fig. 1 is the step flow chart of the vertical ladder demographic method based on deep learning of the embodiment of the present invention；
Fig. 2 is that the YOLOv3 of the embodiment of the present invention detects network architecture schematic diagram；
Fig. 3 is the functional block diagram of the vertical ladder passenger number statistical system based on deep learning of the embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are some of the embodiments of the present invention, instead of all the embodiments.Based on this hair
Embodiment in bright, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, shall fall within the protection scope of the present invention.
Referring to Fig.1, it show a kind of vertical ladder demographic method based on deep learning disclosed by the embodiments of the present invention
Flow chart of steps, comprising the following steps:
S1, vertical ladder monitoring camera have unmanned algorithm to judge someone by camera, and switch gate algorithm judges that vertical ladder door is closed,
And carriage starts analysis request in operating status, triggers an image and vertical ladder pedestrian detection algorithm is called to do target detection；
S2 takes an image from main bit stream according to the timing node of triggering image, and start after receiving analysis request
YOLOv3 algorithm is called to carry out the detection block for analyzing to the end；
Wherein start that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end specifically:
Vertical ladder pedestrian image is done into normalized first, size change over, then will be image gridding to 416*416, if inspection
The target's center of survey falls in some grid, then the grid is just responsible for predicting the target；YOLOv3 obtains 9 using K-means
A priori frame takes the characteristic pattern of three scales to predict, when input is 416*416, characteristic pattern is respectively 13*13,26*26,
52*52；Each yolo layers is divided on the characteristic pattern of 3 scales, scale using three priori frames according to the size of priori frame
Big characteristic pattern uses small priori frame；Each grid can predict multiple bounding box, if the network is pre- in training
The bounding box of survey and the ground truth lap of label are maximum, that is, can determine whether target in grid, grid is special
The prediction of target is responsible in duty；When part, a target may cause target frame to repeat to detect by repeated detection, non-very big
Value inhibits to can detecte the bounding box of high superposed, removes all prediction blocks other than confidence level highest, guarantees same target only
Export a detection block；All prediction blocks obtain specific coordinate information and class categories by recurrence；Pass through setting one
Objective degrees of confidence threshold value, wherein confidence level is lap of the probability of target category multiplied by prediction target and calibration true value, height
In the prediction block of threshold value be last detection block.
The present invention is counted by detecting vertical ladder pedestrian head region, it is contemplated that big than more crowded feelings in flow of the people
Condition, vertical ladder pedestrian's serious shielding, therefore do not consider detection head and shoulder or whole trunk.Fig. 2 provides YOLOv3 detection network frame
Frame schematic diagram, network model size 246M.Input layer is the image for needing to detect, and DBL is convolutional layer+Batch
Normalization layers of+leaky excitation layer.Accelerate to calculate by three layer fusions.Wherein convolutional layer is rolled up for extracting feature
Product is a kind of mathematical operation to two real variable functions.Usual convolution input image data and a kernel function, output are then referred to as
Feature Mapping extracts image feature information by convolution algorithm in deep learning, if a two-dimensional image I as input,
Using a two-dimensional core K, formula is as follows:
Wherein Mr,McFor the row and column of I, Kr,KcFor the row and column of K, m, n are step-length, and i, j need to meet condition: 0≤i≤
Mr+Kr-1,0≤j≤Mc+Kc-1。
Normalization layers of Batch in order to solve the problems, such as depth network training when gradient disappearance, to hidden neuron
Nonlinear function mapping value does standardization:
Wherein x(k)The mapping obtained for neuron through overdriving transformation, k are constant, and E (x) is mean value, and Var (x) is side
Difference.
Leaky excitation layer does Nonlinear Mapping to input neuron:
Wherein xiFor neuron weight, yiFor mapping, aiFor coefficient, i is constant.
In addition it introduces residual unit and deepens network depth.ResN is residual unit, and wherein x is output to the nerve for upper one layer
The value of member；W is the weight that x passes to neuron process；Y is the output valve that x is acquired in certain neuron by activation primitive:
Y=F (x, ω)+x
YOLOv3 network has used 23 residual units, 6 groups of DBL after first group of DBL, and increases convolutional layer and obtain
The feature output of 13*13 is used for the detection of big target.Wherein after 23 residual units, 5 groups of DBL Zuo Liao branch DBL with above adopt
Sample obtains the feature of 26*26 and has done tensor splicing by the 26*26 feature that 19 residual units obtain, and wherein concat is used for
Splice the tensor of same scale, spliced tensor, which passes through 5 groups of DBL and increases convolutional layer, obtains the feature output of 26*26.Its
In the 5 ZuDBLHou Zuo branch continue to be DBL and up-sampling obtains the feature of 52*52, while being obtained with by 11 residual units
52*52 feature do tensor splicing, the spliced tensor is using 6 groups of DBL and increases convolutional layer and has obtained the spy of 52*52
Sign output is used for the detection of Small object.Final three layers of feature output is provided commonly for the detection of vertical ladder pedestrian.13* is used on the whole
Tri- scales of 13,26*26,52*52 detect different target sizes, since YOLOv3 network is for part Small object area
The detection effect in domain is not very well, by the way that cluster obtains small anchor again on vertical ladder pedestrian's data set, to improve
The detection performance of YOLOv3.
S3 obtains specific vertical ladder number by detection block, and quantity is written in database, while being reported to aerial ladder flat
Platform.
A kind of above vertical ladder demographic method based on deep learning provided in an embodiment of the present invention, it is different by mark
Vertical ladder pedestrian sample under scene amounts to 390,000 width, using YOLOv3 network training model, obtains for data set training
Model has good universality in application aspect, adapts to the pedestrian under various intensities of illumination, different vertical ladder types, vertical ladder scene
Identification has very high accuracy rate.
It is corresponding with present invention method, the embodiment of the invention also provides a kind of vertical ladder people based on deep learning
Number statistical system, functional block diagram is referring to fig. 2, comprising:
Analysis request module has unmanned algorithm to judge someone, switch gate algorithm for vertical ladder monitoring camera by camera
Judge that vertical ladder door is closed, and carriage starts analysis request in operating status, triggers an image and calls vertical ladder pedestrian detection
Algorithm does target detection；
Vertical ladder demographics module, for after receiving analysis request, according to the timing node of triggering image from primary key
Stream takes an image, and starts that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end；
Wherein start that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end specifically:
Vertical ladder pedestrian image is done into normalized first, size change over, then will be image gridding to 416*416, if inspection
The target's center of survey falls in some grid, then the grid is just responsible for predicting the target；YOLOv3 obtains 9 using K-means
A priori frame takes the characteristic pattern of three scales to predict, when input is 416*416, characteristic pattern is respectively 13*13,26*26,
52*52；Each yolo layers is divided on the characteristic pattern of 3 scales, scale using three priori frames according to the size of priori frame
Big characteristic pattern uses small priori frame；Each grid can predict multiple bounding box, if the network is pre- in training
The bounding box of survey and the ground truth lap of label are maximum, that is, can determine whether target in grid, grid is special
The prediction of target is responsible in duty；When part, a target may cause target frame to repeat to detect by repeated detection, non-very big
Value inhibits to can detecte the bounding box of high superposed, removes all prediction blocks other than confidence level highest, guarantees same target only
Export a detection block；All prediction blocks obtain specific coordinate information and class categories by recurrence；Pass through setting one
Objective degrees of confidence threshold value, wherein confidence level is lap of the probability of target category multiplied by prediction target and calibration true value, height
In the prediction block of threshold value be last detection block.
The present invention is counted by detecting vertical ladder pedestrian head region, it is contemplated that big than more crowded feelings in flow of the people
Condition, vertical ladder pedestrian's serious shielding, therefore do not consider detection head and shoulder or whole trunk.Fig. 2 provides YOLOv3 detection network frame
Frame schematic diagram, network model size 246M.Input layer is the image for needing to detect, and DBL is convolutional layer+Batch
Normalization layers of+leaky excitation layer.Accelerate to calculate by three layer fusions.Wherein convolutional layer is rolled up for extracting feature
Product is a kind of mathematical operation to two real variable functions.Usual convolution input image data and a kernel function, output are then referred to as
Feature Mapping extracts image feature information by convolution algorithm in deep learning, if a two-dimensional image I as input,
Using a two-dimensional core K, formula is as follows:
Wherein Mr,McFor the row and column of I, Kr,KcFor the row and column of K, m, n are step-length, and i, j need to meet condition: 0≤i≤
Mr+Kr-1,0≤j≤Mc+Kc-1。
Normalization layers of Batch in order to solve the problems, such as depth network training when gradient disappearance, to hidden neuron
Nonlinear function mapping value does standardization:
Wherein x(k)The mapping obtained for neuron through overdriving transformation, k are constant, and E (x) is mean value, and Var (x) is side
Difference.
Leaky excitation layer does Nonlinear Mapping to input neuron:
Wherein xiFor neuron weight, yiFor mapping, aiFor coefficient, i is constant.
In addition it introduces residual unit and deepens network depth.ResN is residual unit, and wherein x is output to the nerve for upper one layer
The value of member；W is the weight that x passes to neuron process；Y is the output valve that x is acquired in certain neuron by activation primitive:
Y=F (x, ω)+x
YOLOv3 network has used 23 residual units, 6 groups of DBL after first group of DBL, and increases convolutional layer and obtain
The feature output of 13*13 is used for the detection of big target.Wherein after 23 residual units, 5 groups of DBL Zuo Liao branch DBL with above adopt
Sample obtains the feature of 26*26 and has done tensor splicing by the 26*26 feature that 19 residual units obtain, and wherein concat is used for
Splice the tensor of same scale, spliced tensor, which passes through 5 groups of DBL and increases convolutional layer, obtains the feature output of 26*26.Its
In the 5 ZuDBLHou Zuo branch continue to be DBL and up-sampling obtains the feature of 52*52, while being obtained with by 11 residual units
52*52 feature do tensor splicing, the spliced tensor is using 6 groups of DBL and increases convolutional layer and has obtained the spy of 52*52
Sign output is used for the detection of Small object.Final three layers of feature output is provided commonly for the detection of vertical ladder pedestrian.13* is used on the whole
Tri- scales of 13,26*26,52*52 detect different target sizes, since YOLOv3 network is for part Small object area
The detection effect in domain is not very well, by the way that cluster obtains small anchor again on vertical ladder pedestrian's data set, to improve
The detection performance of YOLOv3.
Uploading module is written, for obtaining specific vertical ladder number by detection block, and quantity is written in database, together
When be reported to aerial ladder platform.A kind of above vertical ladder passenger number statistical system based on deep learning provided in an embodiment of the present invention, leads to
The vertical ladder pedestrian sample crossed under mark different scenes amounts to 390,000 width, using YOLOv3 network training model, for the data set
The model that training obtains has good universality in application aspect, adapts to various intensities of illumination, different vertical ladder types, vertical ladder field
Pedestrian's identification under scape, has very high accuracy rate.
It should be appreciated that exemplary embodiment as described herein is illustrative and be not restrictive.Although being retouched in conjunction with attached drawing
One or more embodiments of the invention is stated, it should be understood by one skilled in the art that not departing from through appended right
In the case where the spirit and scope of the present invention defined by it is required that, the change of various forms and details can be made.
Claims (6)
1. a kind of vertical ladder demographic method based on deep learning, which comprises the following steps:
Vertical ladder monitoring camera has unmanned algorithm to judge someone by camera, and switch gate algorithm judges that vertical ladder door is closed, and carriage
Start analysis request in operating status, trigger an image and vertical ladder pedestrian detection algorithm is called to do target detection；
After receiving analysis request, an image is taken from main bit stream according to the timing node of triggering image, and start to call
YOLOv3 algorithm carries out the detection block for analyzing to the end；
Specific vertical ladder number is obtained by detection block, and quantity is written in database, while being reported to aerial ladder platform.
2. the vertical ladder demographic method based on deep learning as described in claim 1, which is characterized in that described to start to call
YOLOv3 algorithm carries out the detection block for analyzing to the end specifically:
Vertical ladder pedestrian image is done into normalized first, size change over, then will be image gridding to 416*416, if detection
Target's center falls in some grid, then the grid is just responsible for predicting the target；YOLOv3 obtains 9 elder generations using K-means
Frame is tested, the characteristic pattern of three scales is taken to predict, when input is 416*416, characteristic pattern is respectively 13*13,26*26,52*
52；Each yolo layers be divided on the characteristic pattern of 3 scales using three priori frames according to the size of priori frame, and scale is big
Characteristic pattern uses small priori frame；Each grid can predict multiple bounding box, if the neural network forecast in training
The ground truth lap of bounding box and label is maximum, that is, can determine whether target in grid, grid sole duty is negative
Blame the prediction of target；When part, a target may cause target frame to repeat to detect by repeated detection, non-maximum suppression
System can detecte the bounding box of high superposed, removes all prediction blocks other than confidence level highest, guarantees that same target only exports
One detection block；All prediction blocks obtain specific coordinate information and class categories by recurrence；By the way that a target is arranged
Confidence threshold value, wherein confidence level is the probability of target category multiplied by the lap of prediction target and calibration true value, is higher than threshold
The prediction block of value is last detection block.
3. the vertical ladder demographic method based on deep learning as claimed in claim 2, which is characterized in that take three scales
Characteristic pattern prediction, when input is 416*416, characteristic pattern is respectively 13*13, and 26*26,52*52 detect net using YOLOv3
Network obtains, specifically: YOLOv3 detection network has used 23 residual units, 6 groups of DBL after first group of DBL, and increases volume
The feature output that lamination obtains 13*13 is used for the detection of big target；The wherein Zuo Liao branch after 23 residual units, 5 groups of DBL
DBL and up-sampling obtain the feature of 26*26 and have done tensor splicing by the 26*26 feature that 19 residual units obtain, wherein
Concat is used to splice the tensor of same scale, and spliced tensor, which passes through 5 groups of DBL and increases convolutional layer, obtains 26*26's
Feature output；Wherein the 5 ZuDBLHou Zuo branch continues to be DBL and up-sampling obtains the feature of 52*52, at the same with pass through 11
The 52*52 feature that residual unit obtains does tensor splicing, and the spliced tensor is using 6 groups of DBL and increases convolutional layer and obtains
The feature output for having arrived 52*52 is used for the detection of Small object；Final three layers of feature output is provided commonly for the detection of vertical ladder pedestrian.
4. a kind of vertical ladder passenger number statistical system based on deep learning characterized by comprising
Analysis request module has unmanned algorithm to judge someone, the judgement of switch gate algorithm for vertical ladder monitoring camera by camera
Vertical ladder door is closed, and carriage starts analysis request in operating status, triggers an image and calls vertical ladder pedestrian detection algorithm
Do target detection；
Vertical ladder demographics module, for being taken according to the timing node of triggering image from main bit stream after receiving analysis request
One image, and start that YOLOv3 algorithm is called to carry out the detection block for analyzing to the end；
Uploading module is written, for obtaining specific vertical ladder number by detection block, and quantity is written in database, while on
Registration aerial ladder platform.
5. the vertical ladder passenger number statistical system based on deep learning as claimed in claim 4, which is characterized in that described to start to call
YOLOv3 algorithm carries out the detection block for analyzing to the end specifically:
Vertical ladder pedestrian image is done into normalized first, size change over, then will be image gridding to 416*416, if detection
Target's center falls in some grid, then the grid is just responsible for predicting the target；YOLOv3 obtains 9 elder generations using K-means
Frame is tested, the characteristic pattern of three scales is taken to predict, when input is 416*416, characteristic pattern is respectively 13*13,26*26,52*
52；Each yolo layers be divided on the characteristic pattern of 3 scales using three priori frames according to the size of priori frame, and scale is big
Characteristic pattern uses small priori frame；Each grid can predict multiple bounding box, if the neural network forecast in training
The ground truth lap of bounding box and label is maximum, that is, can determine whether target in grid, grid sole duty is negative
Blame the prediction of target；When part, a target may cause target frame to repeat to detect by repeated detection, non-maximum suppression
System can detecte the bounding box of high superposed, removes all prediction blocks other than confidence level highest, guarantees that same target only exports
One detection block；All prediction blocks obtain specific coordinate information and class categories by recurrence；By the way that a target is arranged
Confidence threshold value, wherein confidence level is the probability of target category multiplied by the lap of prediction target and calibration true value, is higher than threshold
The prediction block of value is last detection block.
6. the vertical ladder passenger number statistical system based on deep learning as claimed in claim 5, which is characterized in that take three scales
Characteristic pattern prediction, when input is 416*416, characteristic pattern is respectively 13*13, and 26*26,52*52 detect net using YOLOv3
Network obtains, specifically: YOLOv3 detection network has used 23 residual units, 6 groups of DBL after first group of DBL, and increases volume
The feature output that lamination obtains 13*13 is used for the detection of big target；The wherein Zuo Liao branch after 23 residual units, 5 groups of DBL
DBL and up-sampling obtain the feature of 26*26 and have done tensor splicing by the 26*26 feature that 19 residual units obtain, wherein
Concat is used to splice the tensor of same scale, and spliced tensor, which passes through 5 groups of DBL and increases convolutional layer, obtains 26*26's
Feature output；Wherein the 5 ZuDBLHou Zuo branch continues to be DBL and up-sampling obtains the feature of 52*52, at the same with pass through 11
The 52*52 feature that residual unit obtains does tensor splicing, and the spliced tensor is using 6 groups of DBL and increases convolutional layer and obtains
The feature output for having arrived 52*52 is used for the detection of Small object；Final three layers of feature output is provided commonly for the detection of vertical ladder pedestrian.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910023341.3A CN109858389B (en) | 2019-01-10 | 2019-01-10 | Vertical ladder people counting method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910023341.3A CN109858389B (en) | 2019-01-10 | 2019-01-10 | Vertical ladder people counting method and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109858389A true CN109858389A (en) | 2019-06-07 |
| CN109858389B CN109858389B (en) | 2021-06-04 |
Family
ID=66894403
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910023341.3A Active CN109858389B (en) | 2019-01-10 | 2019-01-10 | Vertical ladder people counting method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109858389B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222673A (en) * | 2019-06-21 | 2019-09-10 | 杭州宇泛智能科技有限公司 | A kind of passenger flow statistical method based on head detection |
| CN110414391A (en) * | 2019-07-15 | 2019-11-05 | 河北工业大学 | Active mobile vehicle bottom hazardous material detection device based on deep learning algorithm |
| CN110738148A (en) * | 2019-09-29 | 2020-01-31 | 浙江新再灵科技股份有限公司 | Cloud target detection algorithm based on heterogeneous platform |
| CN110765904A (en) * | 2019-10-11 | 2020-02-07 | 浙江新再灵科技股份有限公司 | Device and method for predicting crowd consumption index based on elevator scene |
| CN111353377A (en) * | 2019-12-24 | 2020-06-30 | 浙江工业大学 | Elevator passenger number detection method based on deep learning |
| CN111476600A (en) * | 2020-03-23 | 2020-07-31 | 浙江新再灵科技股份有限公司 | Method for counting and analyzing audience number of vertical ladder advertisement |
| CN111986253A (en) * | 2020-08-21 | 2020-11-24 | 日立楼宇技术（广州）有限公司 | Method, device, equipment and storage medium for detecting crowding degree of elevator |
| CN112926500A (en) * | 2021-03-22 | 2021-06-08 | 重庆邮电大学 | Pedestrian detection method combining head and overall information |
| CN113012335A (en) * | 2021-03-22 | 2021-06-22 | 上海工程技术大学 | Subway platform guide queuing system based on YOLOv3 face detection |
| CN114333120A (en) * | 2022-03-14 | 2022-04-12 | 南京理工大学 | Bus passenger flow detection method and system |
| CN114495003A (en) * | 2022-01-24 | 2022-05-13 | 上海申视信科技有限公司 | People number identification and statistics method and system based on improved YOLOv3 network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108069307A (en) * | 2016-11-14 | 2018-05-25 | 杭州海康威视数字技术股份有限公司 | The method and device that a kind of number in elevator is counted |
| CN108792853A (en) * | 2018-06-26 | 2018-11-13 | 潍坊学院 | elevator dispatching system and method |
-
2019
- 2019-01-10 CN CN201910023341.3A patent/CN109858389B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108069307A (en) * | 2016-11-14 | 2018-05-25 | 杭州海康威视数字技术股份有限公司 | The method and device that a kind of number in elevator is counted |
| CN108792853A (en) * | 2018-06-26 | 2018-11-13 | 潍坊学院 | elevator dispatching system and method |
Non-Patent Citations (1)
| Title |
|---|
| 木盏: "yolo系列之yolo v3", 《HTTPS://BLOG.CSDN.NET/LCCZZU/ARTICLE/DETAILS/85881672》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222673B (en) * | 2019-06-21 | 2021-04-06 | 杭州宇泛智能科技有限公司 | Passenger flow statistical method based on head detection |
| CN110222673A (en) * | 2019-06-21 | 2019-09-10 | 杭州宇泛智能科技有限公司 | A kind of passenger flow statistical method based on head detection |
| CN110414391A (en) * | 2019-07-15 | 2019-11-05 | 河北工业大学 | Active mobile vehicle bottom hazardous material detection device based on deep learning algorithm |
| CN110738148A (en) * | 2019-09-29 | 2020-01-31 | 浙江新再灵科技股份有限公司 | Cloud target detection algorithm based on heterogeneous platform |
| CN110765904A (en) * | 2019-10-11 | 2020-02-07 | 浙江新再灵科技股份有限公司 | Device and method for predicting crowd consumption index based on elevator scene |
| CN111353377A (en) * | 2019-12-24 | 2020-06-30 | 浙江工业大学 | Elevator passenger number detection method based on deep learning |
| CN111476600A (en) * | 2020-03-23 | 2020-07-31 | 浙江新再灵科技股份有限公司 | Method for counting and analyzing audience number of vertical ladder advertisement |
| CN111476600B (en) * | 2020-03-23 | 2023-09-19 | 浙江新再灵科技股份有限公司 | Statistical analysis method for audience numbers of direct ladder advertisement |
| CN111986253B (en) * | 2020-08-21 | 2023-09-15 | 日立楼宇技术（广州）有限公司 | Method, device, equipment and storage medium for detecting elevator crowding degree |
| CN111986253A (en) * | 2020-08-21 | 2020-11-24 | 日立楼宇技术（广州）有限公司 | Method, device, equipment and storage medium for detecting crowding degree of elevator |
| CN112926500A (en) * | 2021-03-22 | 2021-06-08 | 重庆邮电大学 | Pedestrian detection method combining head and overall information |
| CN112926500B (en) * | 2021-03-22 | 2022-09-20 | 重庆邮电大学 | Pedestrian detection method combining head and overall information |
| CN113012335A (en) * | 2021-03-22 | 2021-06-22 | 上海工程技术大学 | Subway platform guide queuing system based on YOLOv3 face detection |
| CN114495003A (en) * | 2022-01-24 | 2022-05-13 | 上海申视信科技有限公司 | People number identification and statistics method and system based on improved YOLOv3 network |
| CN114333120A (en) * | 2022-03-14 | 2022-04-12 | 南京理工大学 | Bus passenger flow detection method and system |
| WO2023174240A1 (en) * | 2022-03-14 | 2023-09-21 | 南京理工大学 | Bus passenger flow detection method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109858389B (en) | 2021-06-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109858389A (en) | Vertical ladder demographic method and system based on deep learning | |
| CN111259850B (en) | Pedestrian re-identification method integrating random batch mask and multi-scale representation learning | |
| CN111241959B (en) | Method for detecting personnel not wearing safety helmet through construction site video stream | |
| CN107808139B (en) | Real-time monitoring threat analysis method and system based on deep learning | |
| CN109344736B (en) | Static image crowd counting method based on joint learning | |
| CN104992223B (en) | Intensive population estimation method based on deep learning | |
| CN112183471A (en) | Automatic detection method and system for standard wearing of epidemic prevention mask of field personnel | |
| CN112949572A (en) | Slim-YOLOv 3-based mask wearing condition detection method | |
| JP2004523840A (en) | Classification of Objects by Model Set | |
| CN108470138A (en) | Method for target detection and device | |
| CN112287827A (en) | Complex environment pedestrian mask wearing detection method and system based on intelligent lamp pole | |
| CN114220143B (en) | Face recognition method for wearing mask | |
| CN104134364B (en) | Real-time traffic sign identification method and system with self-learning capacity | |
| CN110033040A (en) | A kind of flame identification method, system, medium and equipment | |
| CN113869162A (en) | Violation identification method and system based on artificial intelligence | |
| CN107766823A (en) | Anomaly detection method in video based on key area feature learning | |
| CN116229560B (en) | Abnormal behavior recognition method and system based on human body posture | |
| CN109886242A (en) | A kind of method and system that pedestrian identifies again | |
| CN109919246A (en) | Pedestrian's recognition methods again based on self-adaptive features cluster and multiple risks fusion | |
| CN116824335A (en) | YOLOv5 improved algorithm-based fire disaster early warning method and system | |
| CN113378668A (en) | Method, device and equipment for determining accumulated water category and storage medium | |
| Su et al. | Semantic segmentation of high resolution remote sensing image based on batch-attention mechanism | |
| CN113095127B (en) | Building post-earthquake positioning and damage state evaluation method based on satellite images | |
| CN115578693A (en) | Construction safety early warning method and device based on significance neural network model | |
| CN117079351B (en) | Method and system for analyzing personnel behaviors in key areas |
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 |