CN110059667A - Pedestrian counting method - Google Patents
Pedestrian counting method Download PDFInfo
- Publication number
- CN110059667A CN110059667A CN201910353398.XA CN201910353398A CN110059667A CN 110059667 A CN110059667 A CN 110059667A CN 201910353398 A CN201910353398 A CN 201910353398A CN 110059667 A CN110059667 A CN 110059667A
- Authority
- CN
- China
- Prior art keywords
- result
- pedestrian
- regression
- prediction
- classification
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000001514 detection method Methods 0.000 claims abstract description 42
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 33
- 230000011218 segmentation Effects 0.000 claims abstract description 32
- 238000011176 pooling Methods 0.000 claims abstract description 20
- 238000012549 training Methods 0.000 claims abstract description 14
- 238000012805 post-processing Methods 0.000 claims abstract description 11
- 239000011159 matrix material Substances 0.000 claims description 13
- 238000005457 optimization Methods 0.000 claims description 4
- 235000013399 edible fruits Nutrition 0.000 claims description 3
- 230000011731 head segmentation Effects 0.000 claims 1
- 239000000284 extract Substances 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 19
- 238000004364 calculation method Methods 0.000 description 4
- 238000012544 monitoring process Methods 0.000 description 4
- 230000004048 modification Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 241001674808 Biastes Species 0.000 description 1
- 101000657326 Homo sapiens Protein TANC2 Proteins 0.000 description 1
- 102100034784 Protein TANC2 Human genes 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of pedestrian counting methods, comprising: S1: obtaining monitor video image to be detected；S2: inputting improved depth convolutional neural networks for image to be detected, i.e., improved Cascade R-CNN network, extracts feature；S3: by RPN network training, Softmax classification is carried out to image and is determined with target area；S4: by Pooling layers of ROIs, the IoU threshold value changed in returning every time carries out multiple regression to frame coordinate, and multiple regression network structure is cascade structure；S5: after returning more each time as a result, the best cascade number of selection, exports optimal number of people prediction segmentation result and regression forecasting result；S6: carrying out pedestrian detection post-processing, and the prediction of crowd's foreground segmentation carries out Hadamard product with the prediction of crowd density figure；S7: exporting final pedestrian's amount detection and pedestrian density schemes.By the above-mentioned means, the present invention can be suitable for the pedestrian counting and Density Detection of different occasions, it is effective to improve testing result precision and speed.
Description
Technical field
The present invention relates to a kind of pedestrian counting methods.
Background technique
Accurately estimating real-time number in monitoring scene can help related personnel to carry out emergency event early warning and thing in advance
The safety of life and property of decision afterwards, people will be protected.
The current existing pedestrian counting method based on deep learning is broadly divided into two kinds: 1) based on network structure characteristic
Method；2) network training process is different.Pedestrian counting method based on deep learning has some limitations.Method
1) multiple row convolutional neural networks are used more, need great amount of samples, complexity is high；Method 2) training time is too long, human body target point
The problems such as resolution is lower, feature differentiation is more difficult.
Summary of the invention
The purpose of the present invention is to provide a kind of pedestrian counting methods.
To solve the above problems, the present invention provides a kind of pedestrian counting method, comprising:
Step S1 obtains monitor video image to be detected；
The monitor video image to be detected is inputted improved depth convolutional neural networks, that is, improved by step S2
Cascade R-CNN network is to extract feature；
The feature extracted is trained by step S3 by RPN network, to the monitor video image to be detected
It carries out Softmax classification to determine with target area, the judgement result B0 of the result C0 and target area to be classified；
Step S4, based on Softmax classification with the judgement of target area as a result, and by ROIs Pooling layers,
The IoU threshold value changed in returning every time carries out multiple regression to frame coordinate, wherein the network structure of the multiple regression is grade
It is coupled structure；
Step S5 compares after returning each time in multiple regression as a result, the best cascade number of selection, exports optimal pre-
Survey number of people segmentation result and regression result；
Step S6 is based on optimum prediction number of people segmentation result and regression result, and carrying out pedestrian detection post-processing includes: people
The prediction of group's foreground segmentation carries out Hadamard product with the prediction of crowd density figure；
Step S7 exports final pedestrian's predicted quantity and pedestrian density on the basis of pedestrian detection post-processing
Figure.
Further, in the above-mentioned methods, the monitor video image to be detected is inputted improved depth by step S2
Convolutional neural networks, that is, improved Cascade R-CNN network is spent to extract feature, comprising:
The monitor video image to be detected first passes through five convolution stages, the Conv for being 3*3 including 13 kernels, and 3
A kernel is the Maxpool of 2*2；
It is improved improved depth convolutional neural networks are inputted by the monitor video image in five convolution stages
Cascade R-CNN network to extract feature.
Further, in the above-mentioned methods, step S4, the judgement knot based on Softmax classification and target area
Fruit, and by ROIs Pooling layers, the IoU threshold value changed in returning every time carries out multiple regression to frame coordinate, wherein institute
The network structure for stating multiple regression is cascade structure, comprising:
S41: being a by the judgement result B0 input IoU threshold value of target area1ROIs Pooling, train detector
H1 predicts the result C1 of classification and the judgement result B1 of target area；
S42: it will determine that result B1 input IoU threshold value is a2=a1+x1The ROIs Pooling of (0 < x < 1), trains
Detector H2 predicts the result C2 of classification and the judgement result B2 of target area；
S43: it will determine that result B2 input IoU threshold value is a3=a2+x2The ROIs Pooling of (0 < x < 1), trains
Detector H3 predicts the result C3 of classification and the judgement result B3 of target area.
Further, in the above-mentioned methods, step S5 compares after returning each time in multiple regression as a result, selecting most
Good cascade number exports optimum prediction number of people segmentation result and regression result, comprising:
S51: using the index for including AP, the size of C0~C3 of result B0~B3 and classification is compared to determine, is selected optimal
The number of people divide prediction result Bn(n=1,2,3) and regression forecasting result Cn(n=1,2,3)；
S52: by BbWith CbCarry out training objective classification using multitask assembling loss function to return with detection block.
Further, in the above-mentioned methods, step S6 is based on optimum prediction number of people segmentation result and regression result, carries out
Pedestrian detection post-processing includes: the prediction of crowd's foreground segmentation and crowd density figure prediction progress Hadamard product, comprising:
S61: by optimum prediction number of people segmentation result and regression result Bn(n=1,2,3) and Cn(n=1,2,3) is damaged
Unwise calculation returns the loss function of loss item using detection, uses | | Lloc||1Loss function Optimization Prediction biases t=(tx, ty,
tw, th) and target biasWherein, t represents detection block, x, and y represents the upper left position of detection block,
W, h respectively represents the width and height of detection block；
S62:Lcls|objThe loss of each detection block classification is exported, totally 2 class, when providing target detection score pobj, network
Branch excludes score lower than threshold value O firstpRegion；With LobjEqually, Lcls|objEach position is exported by Softmax layers
Probabilityα=β=1/3 obtains following loss function:
S63: by optimum prediction number of people segmentation result and regression result Bn(n=1,2,3) and CnThe result of (n=1,2,3)
Matrix carries out Hadamard product, i.e. two matrix corresponding elements are multiplied:
Output=Freg⊙Fseg。
Compared with prior art, the beneficial effects of the present invention are:
1: the present invention uses full convolutional network, can input image to be detected of any size, improve traditional network
Input the problem that size must be consistent；
2: the present invention uses improved Cascade R-CNN network, and multipass changes the size of different IoU threshold values,
Can be with the quality of sample when limited guarantee training, and more accurate pedestrian counting detector is trained, it improves traditional
Threshold value is single or excessively high causes over-fitting；
3: the present invention adapts to detect with the pedestrian countings of different scenes, can efficiently and accurately predict pedestrian's quantity with
Density.
Detailed description of the invention
Fig. 1 is the flow chart of the pedestrian counting method based on concatenated convolutional neural network of one embodiment of the invention；
Fig. 2 is the convolutional neural networks structural schematic diagram of one embodiment of the invention；
Fig. 3 is that the cascade structure of one embodiment of the invention returns flowage structure schematic diagram；
Fig. 4 is the multitask training loss function calculation flow chart of one embodiment of the invention.
Specific embodiment
In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, with reference to the accompanying drawing and specific real
Applying mode, the present invention is described in further detail.
As shown in Figure 1, the present invention provides a kind of pedestrian counting method, comprising:
Step S1 obtains monitor video image to be detected；
The monitor video image to be detected is inputted improved depth convolutional neural networks, that is, improved by step S2
Cascade R-CNN network is to extract feature；
The feature extracted is trained by step S3 by RPN network, to the monitor video image to be detected
It carries out Softmax classification to determine with target area, the judgement result B0 of the result C0 and target area to be classified；
Step S4, based on Softmax classification with the judgement of target area as a result, and by ROIs Pooling layers,
The IoU threshold value changed in returning every time carries out multiple regression to frame coordinate, wherein the network structure of the multiple regression is grade
It is coupled structure；
Step S5 compares after returning each time in multiple regression as a result, the best cascade number of selection, exports optimal pre-
Survey number of people segmentation result and regression result；
Step S6 is based on optimum prediction number of people segmentation result and regression result, and carrying out pedestrian detection post-processing includes: people
The prediction of group's foreground segmentation carries out Hadamard product with the prediction of crowd density figure；
Step S7 exports final pedestrian's predicted quantity and pedestrian density on the basis of pedestrian detection post-processing
Figure.
Here, the present invention is not subtracted using the IoU threshold value being continuously improved in concatenated convolutional neural network in guarantee sample number
The detector of high quality is trained in the case where few, is shortened the training time, is effectively improved the accuracy rate of result.
In one preferred embodiment of pedestrian counting method of the invention, step S2, by the monitor video figure to be detected
As inputting improved depth convolutional neural networks, that is, improved Cascade R-CNN network to extract feature, comprising:
The monitor video image to be detected first passes through five convolution stages, the Conv for being 3*3 including 13 kernels, and 3
A kernel is the Maxpool of 2*2；
It is improved improved depth convolutional neural networks are inputted by the monitor video image in five convolution stages
Cascade R-CNN network to extract feature.
In one preferred embodiment of pedestrian counting method of the invention, step S4, based on Softmax classification and mesh
Mark the judgement in region as a result, and by ROIs Pooling layer, the IoU threshold value changed in returning every time is more to the progress of frame coordinate
Secondary recurrence, wherein the network structure of the multiple regression is cascade structure, comprising:
S41: being a by the judgement result B0 input IoU threshold value of target area1ROIs Pooling, train detector
H1 predicts the result C1 of classification and the judgement result B1 of target area；
S42: it will determine that result B1 input IoU threshold value is a2=a1+x1The ROIs Pooling of (0 < x < 1), trains
Detector H2 predicts the result C2 of classification and the judgement result B2 of target area；
S43: it will determine that result B2 input IoU threshold value is a3=a2+x2The ROIs Pooling of (0 < x < 1), trains
Detector H3 predicts the result C3 of classification and the judgement result B3 of target area.
In one preferred embodiment of pedestrian counting method of the invention, step S5 compares in multiple regression and returns each time
Afterwards as a result, the best cascade number of selection, exports optimum prediction number of people segmentation result and regression result, comprising:
S51: using the index for including AP, the size of C0~C3 of result B0~B3 and classification is compared to determine, is selected optimal
The number of people divide prediction result Bn(n=1,2,3) and regression forecasting result Cn(n=1,2,3)；
S52: by BnWith CnCarry out training objective classification using multitask assembling loss function to return with detection block.
In one preferred embodiment of pedestrian counting method of the invention, step S6 is based on optimum prediction number of people segmentation result
With regression result, carrying out pedestrian detection post-processing includes: that the prediction of crowd's foreground segmentation is carried out with the prediction of crowd density figure
Hadamard product, comprising:
S61: by optimum prediction number of people segmentation result and regression result Bn(n=1,2,3) and Cn(n=1,2,3) is damaged
Unwise calculation returns the loss function of loss item using detection, uses | | Lloc||1Loss function Optimization Prediction biases t=(tx, ty,
tw, th) and target biasWherein, t represents detection block, x, and y represents the upper left position of detection block,
W, h respectively represents the width and height of detection block；
S62:Lcls|objThe loss of each detection block classification is exported, totally 2 class, when providing target detection score pobj, network
Branch excludes score lower than threshold value O firstpRegion；With LobjEqually, Lcls|objEach position is exported by Softmax layers
Probabilityα=β=1/3 obtains following loss function:
S63: by optimum prediction number of people segmentation result and regression result Bn(n=1,2,3) and CnThe result of (n=1,2,3)
Matrix carries out Hadamard product, i.e. two matrix corresponding elements are multiplied:
Output=Freg⊙Fseg。
It is carried out here, returning two tasks to number of people foreground segmentation and number of people density simultaneously using multitask loss function
Multitask training, to obtain finer number of people density map prediction result.
Specifically, referring to Fig. 1, in one embodiment, a kind of pedestrian counting side based on concatenated convolutional neural network
Method the following steps are included:
S1: any scene monitoring region real time monitoring video is obtained, and handles framing image；
S2: inputting improved depth convolutional neural networks for image to be detected, i.e., improved Cascade R-CNN network,
Extract feature；
S3: by RPN network training, Softmax classification is carried out to image and is determined with target area；
S4: by Pooling layers of ROIs, the IoU threshold value changed in returning every time carries out multiple regression to frame coordinate, more
Secondary Recurrent networks structure is cascade structure；
S5: after returning more each time as a result, the best cascade number of selection, exports optimum prediction number of people segmentation result
With regression result；
S6: carrying out pedestrian detection post-processing, and the prediction of crowd's foreground segmentation predicts that carrying out Hadamard multiplies with crowd density figure
Product；
S7: exporting final pedestrian's predicted quantity and pedestrian density schemes.
Further details of elaboration is done to a kind of pedestrian counting method based on concatenated convolutional neural network below, but not
It should be as limit.
In step 1, it obtains any scene monitoring region and monitors video in real time, and be processed into (3,224,224) figure
Picture, to carry out subsequent step.
In step 2, in input step 1 then the image data of (3,224,224) passes through into five convolution stages
Liang Ge branch convolutional layer, output category result and coordinate frame regression result.
With reference to Fig. 2, a kind of pedestrian counting method based on concatenated convolutional neural network the following steps are included:
S21: the image data of (3,224,224) in input step 1, into first stage convolutional layer, this layer has 64
The convolution kernel of (3,3), activation primitive are Relu, and a convolution kernel has swept picture and generated a new matrix, and 64 convolution kernels are raw
At 64 layer matrixes.Then data input convolutional layer, and image data is 64*224*224 at this time.Then data input pond layer, step
Long (2,2) refer to laterally mobile 2 lattice every time, longitudinal 2 lattice mobile every time.After such pond, data become
The wide height of (64,112,112), matrix is halved by original 224, becomes 112.
S22: similarly, second and third, four, five convolution kernel numbers successively become 128,256,512,1024.Per stage pond
Later, matrix is reduced into the 1/2 of original matrix.
After 3 layers of pond of S23:13 layers of convolution sum, initial input image data becomes (512,7,7) and carries out Flatten
It calculates, data is evened up into vector, become one-dimensional 512*7*7=25088.
Further, step S3 obtains the knot of foreground segmentation the following steps are included: enter data into concatenated convolutional network
The matrix of consequence of fruit matrix and recurrence.
With reference to Fig. 3, the cascade structure in a kind of pedestrian counting method based on concatenated convolutional neural network is returned, including
Following steps:
S41: target area is determined that result B0 input IoU threshold value is a1ROIs Pooling, train detector
H1, predicts classification results C1 and target area determines result B1；
S42: being a by B1 input IoU threshold value2=a1+x1The ROls Pooling of (0 < x < 1), trains detector H2
It predicts classification results C2 and target area determines result B2；
S43: being a by B2 input IoU threshold value3=a2+x2The ROIs Pooling of (0 < x < 1), trains detector H3,
It predicts classification results C3 and target area determines result B3.
A kind of multitask training loss function with reference to Fig. 4, in the pedestrian counting method based on concatenated convolutional neural network
Calculation flow chart, comprising the following steps:
S51: using indexs such as AP, comparing the size of B0~B3 and C0~C3, selects optimal number of people segmentation prediction result
Bn(n=1,2,3) and regression forecasting result Cn(n=1,2,3)；
B52: by BnCarry out training objective classification using multitask assembling loss function with Cn to return with detection block.
Further, S6 the following steps are included:
S61: first by optimal number of people segmentation result and regression result Bn(n=1,2,3) and Cn(n=1,2,3) is carried out
Costing bio disturbance returns the loss function of loss item using detection, uses | | Lloc||1Loss function Optimization Prediction biases t=(tx,
ty, tw, th) and target biasT represents detection block, x, and y represents the upper left position of detection block, w, h
Represent the width and height of detection block.
S62:Lcls|objThe loss of each detection block classification is exported, totally 2 class.When providing target detection score pobj, network
Branch excludes score lower than threshold value O firstpRegion.With LobjEqually, Lcls|objEach position is exported by Softmax layers
Probabilityα=β=1/3.
S63: matrix of consequence is subjected to Hadamard product, i.e.,
S64: the total losses function of the effect predicted by loss function detection model, this method is smaller, and model is preferable.
The beneficial effects of the present invention are:
1: the present invention uses full convolutional network, can input image to be detected of any size, improve traditional network
Input the problem that size must be consistent；
2: the present invention uses improved Cascade R-CNN network, and multipass changes the size of different IoU threshold values,
Can be with the quality of sample when limited guarantee training, and more accurate pedestrian counting detector is trained, it improves traditional
Threshold value is single or excessively high causes over-fitting；
3: the present invention adapts to detect with the pedestrian countings of different scenes, can efficiently and accurately predict pedestrian's quantity with
Density.
1 each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with its
The difference of his embodiment, the same or similar parts in each embodiment may refer to each other.
Professional further appreciates that, list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, hard in order to clearly demonstrate
The interchangeability of part and software generally describes each exemplary composition and step according to function in the above description.
These functions are implemented in hardware or software actually, the specific application and design constraint depending on technical solution.
Professional technician can use different methods to achieve the described function each specific application, but this reality
Now it should not be considered as beyond the scope of the present invention.
Obviously, those skilled in the art can carry out various modification and variations without departing from essence of the invention to invention
Mind and range.In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies
Within, then the invention is also intended to include including these modification and variations.
Claims (5)
1. a kind of pedestrian counting method characterized by comprising
Step S1 obtains monitor video image to be detected；
The monitor video image to be detected is inputted improved depth convolutional neural networks, that is, improved by step S2
Cascade R-CNN network is to extract feature；
The feature extracted is trained by step S3 by RPN network, is carried out to the monitor video image to be detected
Softmax classification and target area judgement, the judgement result B0 of the result C0 and target area to be classified；
Step S4, based on Softmax classification with the judgement of target area as a result, and by ROIs Pooling layer, change
IoU threshold value in returning every time carries out multiple regression to frame coordinate, wherein the network structure of the multiple regression is level link
Structure；
Step S5 compares after returning each time in multiple regression as a result, the best cascade number of selection, exports the optimum prediction number of people
Segmentation result and regression result；
Step S6 is based on optimum prediction number of people segmentation result and regression result, and carrying out pedestrian detection post-processing includes: crowd's prospect
Segmentation prediction carries out Hadamard product with the prediction of crowd density figure；
Step S7 exports final pedestrian's predicted quantity and pedestrian density schemes on the basis of pedestrian detection post-processing.
2. pedestrian counting method as described in claim 1, which is characterized in that step S2, by the monitor video to be detected
Image inputs improved depth convolutional neural networks, that is, improved Cascade R-CNN network to extract feature, comprising:
The monitor video image to be detected first passes through five convolution stages, the Conv for being 3*3 including 13 kernels, in 3
Core is the Maxpool of 2*2；
Improved depth convolutional neural networks, that is, improved will be inputted by the monitor video image in five convolution stages
Cascade R-CNN network is to extract feature.
3. pedestrian counting method as claimed in claim 2, which is characterized in that step S4, based on Softmax classification and mesh
Mark the judgement in region as a result, and by ROIs Pooling layer, the IoU threshold value changed in returning every time carries out repeatedly frame coordinate
It returns, wherein the network structure of the multiple regression is cascade structure, comprising:
S41: being a by the judgement result B0 input IoU threshold value of target area1ROIs Pooling, train detector H1, in advance
Measure the result C1 of classification and the judgement result B1 of target area；
S42: it will determine that result B1 input IoU threshold value is a2=a1+x1The ROIs Pooling of (0 < x < 1), trains detector
H2 predicts the result C2 of classification and the judgement result B2 of target area；
S43: it will determine that result B2 input IoU threshold value is a3=a2+x2The ROIs Pooling of (0 < x < 1), trains detector
H3 predicts the result C3 of classification and the judgement result B3 of target area.
4. pedestrian counting method as claimed in claim 3, which is characterized in that step S5 compares in multiple regression and returns each time
It is after returning as a result, the best cascade number of selection, exports optimum prediction number of people segmentation result and regression result, comprising:
S51: using the index for including AP, the size of C0~C3 of result B0~B3 and classification is compared to determine, optimal people is selected
Head segmentation prediction result Bn(n=1,2,3) and regression forecasting result Cn(n=1,2,3)；
S52: by BnWith CnCarry out training objective classification using multitask assembling loss function to return with detection block.
5. pedestrian counting method as described in claim 1, which is characterized in that step S6, based on optimum prediction number of people segmentation knot
Fruit and regression result, carrying out pedestrian detection post-processing includes: that the prediction of crowd's foreground segmentation is carried out with the prediction of crowd density figure
Hadamard product, comprising:
S61: by optimum prediction number of people segmentation result and regression result Bn(n=1,2,3) and Cn(n=1,2,3) carries out loss meter
Calculate, the loss function of loss item returned using detection, use | | Lloc||1Loss function Optimization Prediction biases t=(tx, ty, tw, th)
With target biasWherein, t represents detection block, x, and y represents the upper left position of detection block, w, h difference
Represent the width and height of detection block；
S62:Lcls|objThe loss of each detection block classification is exported, totally 2 class, when providing target detection score pobj, network branches head
It first excludes score and is lower than threshold value OpRegion；With LobjEqually, Lcls|objThe probability of each position is exported by Softmax layersα=β=1/3 obtains following loss function:
S63: by optimum prediction number of people segmentation result and regression result Bn(n=1,2,3) and CnThe matrix of consequence of (n=1,2,3) into
Row Hadamard product, i.e. two matrix corresponding elements are multiplied:
Output=Freg⊙Fseg。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910353398.XA CN110059667A (en) | 2019-04-28 | 2019-04-28 | Pedestrian counting method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910353398.XA CN110059667A (en) | 2019-04-28 | 2019-04-28 | Pedestrian counting method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110059667A true CN110059667A (en) | 2019-07-26 |
Family
ID=67321438
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910353398.XA Pending CN110059667A (en) | 2019-04-28 | 2019-04-28 | Pedestrian counting method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110059667A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598672A (en) * | 2019-09-23 | 2019-12-20 | 天津天地伟业机器人技术有限公司 | Multi-region people counting method based on single camera |
| CN111368634A (en) * | 2020-02-05 | 2020-07-03 | 中国人民解放军国防科技大学 | Human head detection method, system and storage medium based on neural network |
| CN111523452A (en) * | 2020-04-22 | 2020-08-11 | 北京百度网讯科技有限公司 | Method and device for detecting human body position in image |
| CN111860331A (en) * | 2020-07-21 | 2020-10-30 | 北京北斗天巡科技有限公司 | Unmanned aerial vehicle is at face identification system in unknown territory of security protection |
| CN112001274A (en) * | 2020-08-06 | 2020-11-27 | 腾讯科技（深圳）有限公司 | Crowd density determination method, device, storage medium and processor |
| CN112464769A (en) * | 2020-11-18 | 2021-03-09 | 西北工业大学 | High-resolution remote sensing image target detection method based on consistent multi-stage detection |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107316058A (en) * | 2017-06-15 | 2017-11-03 | 国家新闻出版广电总局广播科学研究院 | Improve the method for target detection performance by improving target classification and positional accuracy |
| CN108062531A (en) * | 2017-12-25 | 2018-05-22 | 南京信息工程大学 | A kind of video object detection method that convolutional neural networks are returned based on cascade |
| CN108664838A (en) * | 2017-03-27 | 2018-10-16 | 北京中科视维文化科技有限公司 | Based on the monitoring scene pedestrian detection method end to end for improving RPN depth networks |
-
2019
- 2019-04-28 CN CN201910353398.XA patent/CN110059667A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108664838A (en) * | 2017-03-27 | 2018-10-16 | 北京中科视维文化科技有限公司 | Based on the monitoring scene pedestrian detection method end to end for improving RPN depth networks |
| CN107316058A (en) * | 2017-06-15 | 2017-11-03 | 国家新闻出版广电总局广播科学研究院 | Improve the method for target detection performance by improving target classification and positional accuracy |
| CN108062531A (en) * | 2017-12-25 | 2018-05-22 | 南京信息工程大学 | A kind of video object detection method that convolutional neural networks are returned based on cascade |
Non-Patent Citations (2)
| Title |
|---|
| ZHAOWEI CAI等: "Cascade R-CNN: Delving into High Quality Object Detection", 《IEEE》 * |
| 范峻铭: "智能视频监控中的多人体目标检测与跟踪", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598672A (en) * | 2019-09-23 | 2019-12-20 | 天津天地伟业机器人技术有限公司 | Multi-region people counting method based on single camera |
| CN110598672B (en) * | 2019-09-23 | 2023-07-04 | 天津天地伟业机器人技术有限公司 | Multi-region people counting method based on single camera |
| CN111368634A (en) * | 2020-02-05 | 2020-07-03 | 中国人民解放军国防科技大学 | Human head detection method, system and storage medium based on neural network |
| CN111523452A (en) * | 2020-04-22 | 2020-08-11 | 北京百度网讯科技有限公司 | Method and device for detecting human body position in image |
| CN111523452B (en) * | 2020-04-22 | 2023-08-25 | 北京百度网讯科技有限公司 | Method and device for detecting human body position in image |
| CN111860331A (en) * | 2020-07-21 | 2020-10-30 | 北京北斗天巡科技有限公司 | Unmanned aerial vehicle is at face identification system in unknown territory of security protection |
| CN112001274A (en) * | 2020-08-06 | 2020-11-27 | 腾讯科技（深圳）有限公司 | Crowd density determination method, device, storage medium and processor |
| 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 |
| CN112464769A (en) * | 2020-11-18 | 2021-03-09 | 西北工业大学 | High-resolution remote sensing image target detection method based on consistent multi-stage detection |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110059667A (en) | Pedestrian counting method | |
| CN109657547B (en) | Accompanying model-based abnormal trajectory analysis method | |
| Ranjan et al. | Iterative crowd counting | |
| CN109886307A (en) | A kind of image detecting method and system based on convolutional neural networks | |
| Ucat et al. | Postharvest grading classification of cavendish banana using deep learning and tensorflow | |
| CN107529650A (en) | Network model construction and closed loop detection method, corresponding device and computer equipment | |
| CN107480730A (en) | Power equipment identification model construction method and system, the recognition methods of power equipment | |
| CN109117864A (en) | Coronary heart disease risk prediction technique, model and system based on heterogeneous characteristic fusion | |
| CN109978918A (en) | A kind of trajectory track method, apparatus and storage medium | |
| CN107832802A (en) | Quality of human face image evaluation method and device based on face alignment | |
| CN107330372A (en) | A kind of crowd density based on video and the analysis method of unusual checking system | |
| CN107025420A (en) | The method and apparatus of Human bodys' response in video | |
| CN107506692A (en) | A kind of dense population based on deep learning counts and personnel's distribution estimation method | |
| CN107480642A (en) | A kind of video actions recognition methods based on Time Domain Piecewise network | |
| CN109544537A (en) | The fast automatic analysis method of hip joint x-ray image | |
| CN109635812B (en) | The example dividing method and device of image | |
| CN108491766A (en) | A kind of people counting method end to end based on depth decision forest | |
| Rittscher et al. | Towards the automatic analysis of complex human body motions | |
| CN113436174A (en) | Construction method and application of human face quality evaluation model | |
| CN101533475A (en) | Method for extracting feature of shape-adaptive neighborhood based remote sensing image | |
| Monigari et al. | Plant leaf disease prediction | |
| CN108459997A (en) | High skewness data value probability forecasting method based on deep learning and neural network | |
| CN108573197A (en) | Video actions detection method and device | |
| JP4299508B2 (en) | Operation and quality related analysis device in manufacturing process, related analysis method, and computer-readable storage medium | |
| CN107967455A (en) | A kind of transparent learning method of intelligent human-body multidimensional physical feature big data and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190726 |