CN109886085A - People counting method based on deep learning target detection - Google Patents
People counting method based on deep learning target detection Download PDFInfo
- Publication number
- CN109886085A CN109886085A CN201910004771.0A CN201910004771A CN109886085A CN 109886085 A CN109886085 A CN 109886085A CN 201910004771 A CN201910004771 A CN 201910004771A CN 109886085 A CN109886085 A CN 109886085A
- Authority
- CN
- China
- Prior art keywords
- network
- deep learning
- image
- crowd
- training
- 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 48
- 238000001514 detection method Methods 0.000 title claims abstract description 36
- 238000013135 deep learning Methods 0.000 title claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 32
- 230000006870 function Effects 0.000 claims abstract description 21
- 238000012545 processing Methods 0.000 claims abstract description 13
- 238000000605 extraction Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 7
- 230000006399 behavior Effects 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000002372 labelling Methods 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 238000005070 sampling Methods 0.000 claims 1
- 238000013528 artificial neural network Methods 0.000 abstract description 6
- 230000007812 deficiency Effects 0.000 abstract description 2
- 230000011218 segmentation Effects 0.000 description 7
- 238000011160 research Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 101150041570 TOP1 gene Proteins 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000013016 learning Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of people counting methods based on deep learning target detection, comprising the following steps: building deep learning network model: using the YOLO3 network with DarkNet network for basic network；The processing of training data: it by obtaining crowd's image data under more scenes, is handled using image mirrors, random interception way expands training set scale on scale；Training network: training parameter is declined to optimize network by loss function and gradient.The present invention is directed to the deficiency that existing crowd counts, counting statistics are carried out to crowd using based on deep neural network object detection method in certain circumstances, solve the problems, such as that the accuracy in traditional characteristic extracting method is low, it also solves the problems, such as simultaneously larger based on the error in the sparse situation of crowd in deep learning feature homing method, and detecting speed has very big promotion, detection speed is 4 times based on 101 layers of residual error network RetinaNet (view film network) speed, and precision is but suitable with it.
Description
Technical field
The present invention relates to a kind of people counting methods of computer vision field, more particularly to one kind to be based on deep learning mesh
Mark the people counting method of detection.
Background technique
With the growth of population, the acceleration of urbanization process, the behavior that crowd largely assembles is more and more, and scale is more next
Bigger, thus bring tread event also increases increasingly.Manager is mounted with a large amount of in order to facilitate the management in city in city
Camera.Realizing that crowd density estimation and accurate crowd count by monitor video at present is computer vision field
One of research hotspot.The technology is commonly utilized in:
1, the complicated occasion of height that crowd concentrates.Such as: gymnasium, dining room, this large-scale public place in square.Pass through crowd
Number system can estimate crowd density or the effective strength in specified region, can true grasp crowd's trend, prevent group different
The generation of ordinary affair part.
2, procedural style workplace.Such as: airport, railway station.These places true can be obtained using people counting method
To pedestrian's quantity and the accurate data of distribution, reliable basis is provided for science distribution service and management resource, rational management etc..
Traditional people counting method can be generally divided into three kinds:
1, pedestrian detection: judge to whether there is pedestrian in image or video sequence and give using computer vision technique
It is accurately positioned.This method is more direct, in the sparse scene of crowd, by detecting each of video pedestrian, into
And obtain the result of crowd's counting.
2, visual signature trajectory clustering: for video monitoring, generally be directed to sequence of video images, with KLT tracker and
The method of cluster, the number obtained by the trajectory clustering of continuous two width figure is come estimated number.This method has very stringent
It is required that brightness, crowd density etc. can all have a great impact to it.
3, based on the recurrence of feature:
Firstly, crowd is split from image convenient for subsequent feature extraction by foreground segmentation；
Then, a variety of different low-level image features of foreground extraction obtained from segmentation, common feature have: crowd's area and week
Length, marginal information, textural characteristics etc.；
Finally, the feature extracted to be revert to the number in image.Common homing method has: linear regression, Gauss
Process recurrence etc..
Due to using direct method to be easy the influence of difficulties such as being blocked under congested conditions, and indirect method is from crowd's
Global feature sets out, and with the ability that large-scale crowd counts, is suitble to than more crowded scene.
The shortcomings that above-mentioned tradition method of counting, is as follows:
1, pedestrian detection is usually to be modeled with the boosting based on background and motion feature, using background modeling method,
The target for extracting foreground moving is carried out feature extraction in target area, is then classified using classifier, judged whether
Include pedestrian；The problem of background modeling is primarily present at present:
(1) it must adapt to the variation (such as the variation of illumination cause image chroma variation) of environment；
(2) other objects (such as leaf or trunk etc., correctly detected) intensively occurred in image；
(3) it is less applicable in the slightly intensive place of crowd.
The statistical learning pedestrian detection of background subtraction can be used in order to solve speed issue, on condition that background modeling
Method is effective enough, i.e. the good speed of effect is fast, but there are above-mentioned more defects for background modeling.
2, visual signature method of trajectory clustering carries out crowd's counting, also will receive background in similar pedestrian detection method and builds
The influence of mould.And this method is only suitable for using using under the fewer scene of number, for example uses on bus doorway, such as
There are a large amount of circumstance of occlusion in fruit scene, using the method obtain the result is that undesirable.
3, the direct method in the homing method based on feature needs foreground segmentation, and the quality of segmentation performance directly affects
To final calculated result, however foreground segmentation is originally a relatively difficult task, the performance of algorithm largely by
Its influence, compare the place of aggregation in crowd in this approach, performance and precision will greatly reduce, therefore here it is limit this
One key factor of method performance.Indirect method in homing method based on feature is that original image is first converted to crowd
Density map establishes model further according to density map.The method works well in the case that the crowd is dense in large size, but medium and small
Effect is declined under the less intensive place of the place of type, crowd.
The deep learning research tide of recent years is awfully hot, achieves breakthrough in various traditional fields.Convolution mind
Through the training that network implementations is end-to-end, without carrying out foreground segmentation and extracting feature, by obtaining high level after multilayer convolution
Abstract characteristics.Deep learning, which forms more abstract high level by combination low-level image feature, indicates attribute classification or feature, with hair
The distributed nature of existing data indicates.The deep learning of early stage is mainly with BP neural network, from coding dimensionality reduction and sparse self-editing
Based on the research of code device etc..By taking the model in ImageNet challenge match as an example, the deep learning breakthrough of AlexNet in 2012,
The model of the various deep learnings such as the appearance of the deep neural network of VGGNet in 2014, GoogLeNet, ResNet in 2015
Occur, causes comprehensive outburst of deep learning research.
Target detection network mainstream currently based on deep learning is to pass through training with depth residual error network for basic network
The coordinate and object category of bounding box realize the purpose of target detection, are with the shortcoming that depth residual error network is basic network
Speed is slow, and network end is several layers of redundancy phenomena, is not best to feature extraction efficiency.
Summary of the invention
The object of the invention is that solve the above-mentioned problems and provide it is a kind of by deep learning applied to crowd count
The people counting method based on deep learning target detection.
The present invention through the following technical solutions to achieve the above objectives:
A kind of people counting method based on deep learning target detection, comprising the following steps:
Step 1: building deep learning network model: using the YOLO3 network with DarkNet network for basic network；
Step 2: the processing of training data: by obtaining crowd's image data under more scenes, being handled using image mirrors, ruler
Random interception way expands training set scale on degree；
Step 3: training parameter training network: being declined to optimize network by loss function and gradient.
Preferably, on this basic network of DarkNet network, adding three scales in the step 1 and extracting spy
Sign, respectively Scale1, Scal e2, Scale3, wherein Scale1 adds some convolutional layers after basic network and exports again
Box information；Scale2 is up-sampled from the convolutional layer of the layer second from the bottom in Scale1, then the spy with the last one 16x16 size
Sign figure is added, and again by box information is exported after multiple convolution, scale is compared to Scale1 big twice: Scale3 and Scale2 class
Seemingly, the characteristic pattern that last output size is 32 × 32.
Preferably, the step 2 specifically includes the following steps:
Step 2.1: obtaining the image including multiple portraits as input using camera shooting crowd under multiple scenes
Image；
Step 2.2: existing sample data being subjected to sample equilibrium treatment, first ensures that the training sample under different scenes
Quantity is close, secondly by image size, resolution processes at same size；
Step 2.3: using interception, brightness and the expansion of contrast processing mode are existing at random on image mirrors processing, scale
Data；
Step 2.4: image labeling: needing to mark pedestrian place using deep learning object detection method detection crowd's quantity
The rectangle frame of picture position, each frame are determined by two coordinates of rectangle frame diagonal line；Annotation tool use labelimg, every
Picture can all generate an XML file after having marked, and image name in XML file and its pedestrian's coordinate information extraction are arrived
TXT file, for being extracted when network training.
Further, it in the step 2.1, further by downloading data with existing collection on the net, and is downloaded using crawlers
Crowd's image, with the image using camera shooting together as input picture.
Preferably, the step 3 specifically includes the following steps:
Step 3.1: input picture is divided into S × S grid, if the center of people is fallen in a grid, this
A grid is just responsible for detecting this people；
Step 3.2: each grid predicts B rectangle frame and the score of these rectangle frames, this score is for reacting
A possibility that model in this grid for predicting whether containing someone and being this people is how many；
Step 3.3: if people is not present in this grid, being scored at 0；Otherwise it is scored at 1, each rectangle frame packet
Containing 5 predicted values, respectively bx、by、bw、bhAnd confidence, coordinate (bx, by) represent the center of rectangle frame, bwAnd bhRespectively
Indicate that the width and height of rectangle frame, confidence indicate the probability value comprising people；
The target of e-learning is t, including tx、ty、twAnd th, calculation formula is as follows:
bx=σ (tx)+cx, by=σ (ty)+cy,
Wherein cxAnd cyIt is the offset of coordinate, p respectivelywAnd phIt is the side length of preset rectangle frame respectively；
The calculating of step 3.4:LOSS function, that is, loss function:
The specific formula of LOSS function is as follows:
In above-mentioned formula, front two row is coordinate prediction loss, and third behavior includes that the probability of detection object box loses, the
Four behaviors are lost not comprising the probability of detection object box, and last line is the loss of class prediction probability；Wherein LOSS is whole damage
Lose function, λcoordFor loss function coefficient,Expression judges whether j-th of box in i-th of grid is responsible for the object, xiFor
The center abscissa of actual frames,For the center abscissa of prediction block, yiFor the center ordinate of actual frames,For prediction block
Center ordinate, ωiFor the width of actual frames,For the width of prediction block, CiFor concrete class,To predict classification, pi(c)
For actual classification probability,To predict class probability；
Step 3.5: training parameter being declined by loss function and gradient, realizes optimization network purpose.
The beneficial effects of the present invention are:
The present invention is directed to the deficiency that existing crowd counts, in certain circumstances using based on deep neural network target detection
Method carries out counting statistics to crowd, solves the problems, such as that the accuracy in traditional characteristic extracting method is low, while also solving
Based on the larger problem of error in the sparse situation of crowd in deep learning feature homing method, and detect speed have it is very big
Promoted, detection speed be based on 4 times of 101 layers of residual error network RetinaNet (view film network) speed, precision but with its phase
When；More specific advantage is as follows:
1, it completes to be input to object space and class from original image using end-to-end deep neural network, that is, YOLO3 network
Other output extracts features from three different convolutional layers on basic network, then by the feature integration of extraction export frame and
Classification information, detection crowd are higher than traditional characteristic extracting mode precision；
2, it in terms of data processing, is handled using image mirrors, interception, brightness and contrast processing mode expand at random on scale
Training set scale is filled, accuracy is improved；
3, it is detected using this patent method ratio based on deep neural network feature homing method under the sparse scene of crowd
Crowd's quantity is more acurrate, speed faster, detection speed is based on 101 layers of residual error network RetinaNet (view film network) speed
4 times, precision but with its quite.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of YOLO3 network of the present invention；
Fig. 2 is the schematic diagram that trained network of the present invention carries out grid segmentation to input picture in the process.
Specific embodiment
The present invention will be further explained below with reference to the attached drawings:
People counting method of the present invention based on deep learning target detection, comprising the following steps:
Step 1: building deep learning network model: using the YOLO3 net with existing DarkNet network for basic network
Network；As shown in Figure 1, YOLO3 network is to add three scales on this basic network of DarkNet network and mention in this step
Take feature, respectively Scale1, Scal e2, Scale3, wherein Scale1 adds some convolutional layers again after basic network
Export box information；Scale2 from the convolutional layer of the layer second from the bottom in Scale1 up-sample, then with the last one 16x16 size
Characteristic pattern be added, again by after multiple convolution export box information, scale compared to Scale1 big twice: Scale3 with
Scale2 is similar, the characteristic pattern that last output size is 32 × 32.
DarkNet network its speed and accuracy rate compared with other networks is as shown in table 1:
Table 1
Using Backbone as backbone network in table 1, Top1, Top5 represent accuracy rate；Bn Ops/s is billion
The abbreviation of ofoperations per second, i.e., number of processes per second, unit are 1,000,000,000 times per second；BEFLP/s is
Billion floating point operations per second abbreviation, and the number of U.S. second processing floating point number operations,
FPS is the abbreviation of Frames Per Second, i.e., transmission frame number per second.
Step 2: the processing of training data: the training and optimization of network need a large amount of training sample, the matter of training sample
Amount determines that the quality of final mask, the present invention replace single field by obtaining crowd's image data under more scenes to a certain extent
Data are acquired under scape can enhance the generalization ability of final mask, be handled using image mirrors, random interception way expands on scale
Fill training set scale.
This step 2 specifically includes the following steps:
Step 2.1: obtained under multiple scenes using camera shooting crowd include multiple portraits image, in order into one
Step expands data volume, further by downloading data with existing collection on the net, and crowd's image is downloaded using crawlers, by all figures
As being used as input picture；
Step 2.2: existing sample data being subjected to sample equilibrium treatment, first ensures that the training sample under different scenes
Quantity is close, secondly because the approach different images resolution sizes for obtaining sample are different, will cause final training result in this way
Ideal effect cannot be reached, so by image size, resolution processes at same size；
Step 2.3: using interception, brightness and the expansion of contrast processing mode are existing at random on image mirrors processing, scale
Data；
Step 2.4: image labeling: needing to mark pedestrian place using deep learning object detection method detection crowd's quantity
The rectangle frame of picture position, each frame are determined by two coordinates of rectangle frame diagonal line；Annotation tool use labelimg, every
Picture can all generate an XML file after having marked, and image name in XML file and its pedestrian's coordinate information extraction are arrived
TXT file, for being extracted when network training.
Step 3: training parameter training network: being declined to optimize network by loss function and gradient.
This step 3 specifically includes the following steps:
Step 3.1: as shown in Fig. 2, input picture is divided into S × S grid, if the center of people is fallen in a grid
It is interior, then this grid is just responsible for detecting this people；
Step 3.2: each grid predicts B rectangle frame and the score of these rectangle frames, this score is for reacting
A possibility that model in this grid for predicting whether containing someone and being this people is how many；
Step 3.3: if people is not present in this grid, being scored at 0；Otherwise it is scored at 1, each rectangle frame packet
Containing 5 predicted values, respectively bx、by、bw、bhAnd confidence, coordinate (bx, by) represent the center of rectangle frame, bwAnd bhRespectively
Indicate that the width and height of rectangle frame, confidence indicate the probability value comprising people；
The target of e-learning is t, including tx、ty、twAnd th, calculation formula is as follows:
bx=σ (tx)+cx, by=σ (ty)+cy,
Wherein cxAnd cyIt is the offset of coordinate, p respectivelywAnd phIt is the side length of preset rectangle frame respectively；
The calculating of step 3.4:LOSS function, that is, loss function:
LOSS=LOSS1+LOSS2+LOSS3+LOSS4,
The specific formula of LOSS function is as follows:
In above-mentioned formula, front two row is coordinate prediction loss, and third behavior includes that the probability of detection object box loses, the
Four behaviors are lost not comprising the probability of detection object box, and last line, that is, fifth line is the loss of class prediction probability；Wherein LOSS
For whole loss function, λcoordFor loss function coefficient,Indicate to judge whether j-th of box in i-th of grid is responsible for being somebody's turn to do
Object, xiFor the center abscissa of actual frames,For the center abscissa of prediction block, yiFor the center ordinate of actual frames,For
The center ordinate of prediction block, ωiFor the width of actual frames,For the width of prediction block, CiFor concrete class,To predict class
Not, piIt (c) is actual classification probability,To predict class probability；
Step 3.5: training parameter being declined by loss function and gradient, realizes optimization network purpose.
Above-described embodiment is presently preferred embodiments of the present invention, is not a limitation on the technical scheme of the present invention, as long as
Without the technical solution that creative work can be realized on the basis of the above embodiments, it is regarded as falling into the invention patent
Rights protection scope in.
Claims (5)
1. a kind of people counting method based on deep learning target detection, it is characterised in that: the following steps are included:
Step 1: building deep learning network model: using the YOLO3 network with DarkNet network for basic network；
Step 2: the processing of training data: by obtaining crowd's image data under more scenes, being handled using image mirrors, on scale
Random interception way expands training set scale；
Step 3: training parameter training network: being declined to optimize network by loss function and gradient.
2. the people counting method according to claim 1 based on deep learning target detection, it is characterised in that: the step
In rapid 1, on this basic network of DarkNet network, adds three scales and extract feature, respectively Scale1, Scal
E2, Scale3, wherein Scale1 adds some convolutional layers after basic network and exports box information again；Scale2 is from Scale1
In layer second from the bottom convolutional layer up-sampling, then be added with the characteristic pattern of the last one 16x16 size, again by multiple volumes
Box information is exported after product, scale is similar with Scale2 compared to Scale1 big twice: Scale3, and last output size is 32 × 32
Characteristic pattern.
3. the people counting method according to claim 1 based on deep learning target detection, it is characterised in that: the step
Rapid 2 specifically includes the following steps:
Step 2.1: obtaining the image including multiple portraits as input picture using camera shooting crowd under multiple scenes；
Step 2.2: existing sample data being subjected to sample equilibrium treatment, first ensures that the training samples number under different scenes
It is close, secondly by image size, resolution processes at same size；
Step 2.3: using on image mirrors processing, scale, interception, brightness and contrast processing mode expand existing number at random
According to；
Step 2.4: image labeling: image where needing to mark pedestrian using deep learning object detection method detection crowd's quantity
The rectangle frame of position, each frame are determined by two coordinates of rectangle frame diagonal line；Annotation tool uses labelimg, every picture
An XML file can be all generated after having marked, image name in XML file and its pedestrian's coordinate information extraction is literary to TXT
Part, for being extracted when network training.
4. the people counting method according to claim 3 based on deep learning target detection, it is characterised in that: the step
In rapid 2.1, further by the net download data with existing collection, and using crawlers download crowd's image, with utilize camera
The image of shooting is together as input picture.
5. the people counting method according to claim 1 based on deep learning target detection, it is characterised in that: the step
Rapid 3 specifically includes the following steps:
Step 3.1: input picture is divided into S × S grid, if the center of people is fallen in a grid, this lattice
Son is just responsible for detecting this people；
Step 3.2: each grid predicts that B rectangle frame and the score of these rectangle frames, this score are used for reaction model
A possibility that for being predicted whether in this grid containing someone and being this people is how many；
Step 3.3: if people is not present in this grid, being scored at 0；Otherwise being scored at 1, each rectangle frame includes 5
A predicted value, respectively bx、by、bw、bhAnd confidence, coordinate (bx, by) represent the center of rectangle frame, bwAnd bhTable respectively
Show that the width and height of rectangle frame, confidence indicate the probability value comprising people；
The target of e-learning is t, including tx、ty、twAnd th, calculation formula is as follows:
Wherein cxAnd cyIt is the offset of coordinate, p respectivelywAnd phIt is the side length of preset rectangle frame respectively；
The calculating of step 3.4:LOSS function, that is, loss function:
The specific formula of LOSS function is as follows:
In above-mentioned formula, front two row is coordinate prediction loss, and third behavior includes that the probability of detection object box loses, fourth line
To lose not comprising the probability of detection object box, last line is the loss of class prediction probability；Wherein LOSS is whole loss letter
Number, λcoordFor loss function coefficient,Expression judges whether j-th of box in i-th of grid is responsible for the object, xiFor reality
The center abscissa of frame,For the center abscissa of prediction block, yiFor the center ordinate of actual frames,For the center of prediction block
Ordinate, ωiFor the width of actual frames,For the width of prediction block, CiFor concrete class,To predict classification, piIt (c) is real
Border class probability,To predict class probability；
Step 3.5: training parameter being declined by loss function and gradient, realizes optimization network purpose.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910004771.0A CN109886085A (en) | 2019-01-03 | 2019-01-03 | People counting method based on deep learning target detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910004771.0A CN109886085A (en) | 2019-01-03 | 2019-01-03 | People counting method based on deep learning target detection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109886085A true CN109886085A (en) | 2019-06-14 |
Family
ID=66925559
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910004771.0A Pending CN109886085A (en) | 2019-01-03 | 2019-01-03 | People counting method based on deep learning target detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109886085A (en) |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110390283A (en) * | 2019-07-12 | 2019-10-29 | 易诚高科（大连）科技有限公司 | Across camera shooting head's weight search method under a kind of business scenario |
| CN110490252A (en) * | 2019-08-19 | 2019-11-22 | 西安工业大学 | A kind of occupancy detection method and system based on deep learning |
| CN110490099A (en) * | 2019-07-31 | 2019-11-22 | 武汉大学 | A kind of subway common location stream of people's analysis method based on machine vision |
| CN110686669A (en) * | 2019-09-23 | 2020-01-14 | 中国海洋大学 | ROV sea cucumber distribution statistical method and device based on positioning compensation and visual perception |
| CN110852172A (en) * | 2019-10-15 | 2020-02-28 | 华东师范大学 | Method for expanding crowd counting data set based on Cycle Gan picture collage and enhancement |
| CN111079660A (en) * | 2019-12-19 | 2020-04-28 | 点睛数据科技(杭州)有限责任公司 | Cinema online people counting method based on thermal infrared imaging picture |
| CN111460924A (en) * | 2020-03-16 | 2020-07-28 | 上海师范大学 | Gate ticket-passing behavior detection method based on target detection |
| CN111461071A (en) * | 2020-04-30 | 2020-07-28 | 同济大学 | Floor live load statistical method and system based on deep learning algorithm |
| CN111639740A (en) * | 2020-05-09 | 2020-09-08 | 武汉工程大学 | Steel bar counting method based on multi-scale convolution neural network |
| CN111652168A (en) * | 2020-06-09 | 2020-09-11 | 腾讯科技（深圳）有限公司 | Group detection method, device and equipment based on artificial intelligence and storage medium |
| CN111832489A (en) * | 2020-07-15 | 2020-10-27 | 中国电子科技集团公司第三十八研究所 | Subway crowd density estimation method and system based on target detection |
| CN112004061A (en) * | 2020-09-03 | 2020-11-27 | 四川弘和通讯有限公司 | Oil discharge flow normative intelligent monitoring method based on computer vision |
| CN112004063A (en) * | 2020-09-03 | 2020-11-27 | 四川弘和通讯有限公司 | Method for monitoring connection correctness of oil discharge pipe based on multi-camera linkage |
| CN112288084A (en) * | 2020-10-23 | 2021-01-29 | 武汉大学 | Deep learning target detection network compression method based on feature map channel importance degree |
| CN112347841A (en) * | 2020-08-26 | 2021-02-09 | 南京邮电大学 | Deep learning-based anti-collision application method for gangue filling tamping mechanism |
| CN112989952A (en) * | 2021-02-20 | 2021-06-18 | 复旦大学 | Crowd density estimation method and device based on mask guidance |
| CN112990517A (en) * | 2019-12-12 | 2021-06-18 | 中移雄安信息通信科技有限公司 | Crowd distribution prediction method and system |
| CN113505247A (en) * | 2021-07-02 | 2021-10-15 | 兰州理工大学 | Content-based high-duration video pornographic content detection method |
| CN113673326A (en) * | 2021-07-14 | 2021-11-19 | 南京邮电大学 | Unmanned aerial vehicle platform crowd counting method and system based on image deep learning |
| CN113971667A (en) * | 2021-11-02 | 2022-01-25 | 上海可明科技有限公司 | Training and optimizing method for target detection model of surgical instrument in storage environment |
| WO2022022368A1 (en) * | 2020-07-28 | 2022-02-03 | 宁波环视信息科技有限公司 | Deep-learning-based apparatus and method for monitoring behavioral norms in jail |
| WO2023231290A1 (en) * | 2022-05-30 | 2023-12-07 | 哈尔滨工业大学（深圳） | Casualty recognition method and system based on deep learning in casualty gathering place scene |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778852A (en) * | 2016-12-07 | 2017-05-31 | 中国科学院信息工程研究所 | A kind of picture material recognition methods for correcting erroneous judgement |
| CN108805070A (en) * | 2018-06-05 | 2018-11-13 | 合肥湛达智能科技有限公司 | A kind of deep learning pedestrian detection method based on built-in terminal |
| CN108830145A (en) * | 2018-05-04 | 2018-11-16 | 深圳技术大学（筹） | A kind of demographic method and storage medium based on deep neural network |
| CN109117794A (en) * | 2018-08-16 | 2019-01-01 | 广东工业大学 | A kind of moving target behavior tracking method, apparatus, equipment and readable storage medium storing program for executing |
-
2019
- 2019-01-03 CN CN201910004771.0A patent/CN109886085A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106778852A (en) * | 2016-12-07 | 2017-05-31 | 中国科学院信息工程研究所 | A kind of picture material recognition methods for correcting erroneous judgement |
| CN108830145A (en) * | 2018-05-04 | 2018-11-16 | 深圳技术大学（筹） | A kind of demographic method and storage medium based on deep neural network |
| CN108805070A (en) * | 2018-06-05 | 2018-11-13 | 合肥湛达智能科技有限公司 | A kind of deep learning pedestrian detection method based on built-in terminal |
| CN109117794A (en) * | 2018-08-16 | 2019-01-01 | 广东工业大学 | A kind of moving target behavior tracking method, apparatus, equipment and readable storage medium storing program for executing |
Non-Patent Citations (1)
| Title |
|---|
| 邹阿金等: "基于深度残差网络和YOLO 的人物识别***", 《现代计算机》 * |
Cited By (31)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110390283B (en) * | 2019-07-12 | 2023-01-03 | 易诚高科（大连）科技有限公司 | Cross-camera pedestrian re-retrieval method in commercial scene |
| CN110390283A (en) * | 2019-07-12 | 2019-10-29 | 易诚高科（大连）科技有限公司 | Across camera shooting head's weight search method under a kind of business scenario |
| CN110490099A (en) * | 2019-07-31 | 2019-11-22 | 武汉大学 | A kind of subway common location stream of people's analysis method based on machine vision |
| CN110490099B (en) * | 2019-07-31 | 2022-10-21 | 武汉大学 | Subway public place pedestrian flow analysis method based on machine vision |
| CN110490252A (en) * | 2019-08-19 | 2019-11-22 | 西安工业大学 | A kind of occupancy detection method and system based on deep learning |
| CN110490252B (en) * | 2019-08-19 | 2022-11-15 | 西安工业大学 | Indoor people number detection method and system based on deep learning |
| CN110686669A (en) * | 2019-09-23 | 2020-01-14 | 中国海洋大学 | ROV sea cucumber distribution statistical method and device based on positioning compensation and visual perception |
| CN110852172A (en) * | 2019-10-15 | 2020-02-28 | 华东师范大学 | Method for expanding crowd counting data set based on Cycle Gan picture collage and enhancement |
| CN112990517A (en) * | 2019-12-12 | 2021-06-18 | 中移雄安信息通信科技有限公司 | Crowd distribution prediction method and system |
| CN111079660A (en) * | 2019-12-19 | 2020-04-28 | 点睛数据科技(杭州)有限责任公司 | Cinema online people counting method based on thermal infrared imaging picture |
| CN111460924B (en) * | 2020-03-16 | 2023-04-07 | 上海师范大学 | Gate ticket-evading behavior detection method based on target detection |
| CN111460924A (en) * | 2020-03-16 | 2020-07-28 | 上海师范大学 | Gate ticket-passing behavior detection method based on target detection |
| CN111461071A (en) * | 2020-04-30 | 2020-07-28 | 同济大学 | Floor live load statistical method and system based on deep learning algorithm |
| CN111639740A (en) * | 2020-05-09 | 2020-09-08 | 武汉工程大学 | Steel bar counting method based on multi-scale convolution neural network |
| 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 |
| CN111652168A (en) * | 2020-06-09 | 2020-09-11 | 腾讯科技（深圳）有限公司 | Group detection method, device and equipment based on artificial intelligence and storage medium |
| CN111832489A (en) * | 2020-07-15 | 2020-10-27 | 中国电子科技集团公司第三十八研究所 | Subway crowd density estimation method and system based on target detection |
| WO2022022368A1 (en) * | 2020-07-28 | 2022-02-03 | 宁波环视信息科技有限公司 | Deep-learning-based apparatus and method for monitoring behavioral norms in jail |
| CN112347841A (en) * | 2020-08-26 | 2021-02-09 | 南京邮电大学 | Deep learning-based anti-collision application method for gangue filling tamping mechanism |
| CN112347841B (en) * | 2020-08-26 | 2022-07-26 | 南京邮电大学 | Anti-collision application method of gangue filling tamping mechanism based on deep learning |
| CN112004063A (en) * | 2020-09-03 | 2020-11-27 | 四川弘和通讯有限公司 | Method for monitoring connection correctness of oil discharge pipe based on multi-camera linkage |
| CN112004061A (en) * | 2020-09-03 | 2020-11-27 | 四川弘和通讯有限公司 | Oil discharge flow normative intelligent monitoring method based on computer vision |
| CN112288084B (en) * | 2020-10-23 | 2022-09-02 | 武汉大学 | Deep learning target detection network compression method based on feature map channel importance |
| CN112288084A (en) * | 2020-10-23 | 2021-01-29 | 武汉大学 | Deep learning target detection network compression method based on feature map channel importance degree |
| CN112989952B (en) * | 2021-02-20 | 2022-10-18 | 复旦大学 | Crowd density estimation method and device based on mask guidance |
| CN112989952A (en) * | 2021-02-20 | 2021-06-18 | 复旦大学 | Crowd density estimation method and device based on mask guidance |
| CN113505247A (en) * | 2021-07-02 | 2021-10-15 | 兰州理工大学 | Content-based high-duration video pornographic content detection method |
| CN113673326A (en) * | 2021-07-14 | 2021-11-19 | 南京邮电大学 | Unmanned aerial vehicle platform crowd counting method and system based on image deep learning |
| CN113673326B (en) * | 2021-07-14 | 2023-08-15 | 南京邮电大学 | Unmanned plane platform crowd counting method and system based on image deep learning |
| CN113971667A (en) * | 2021-11-02 | 2022-01-25 | 上海可明科技有限公司 | Training and optimizing method for target detection model of surgical instrument in storage environment |
| WO2023231290A1 (en) * | 2022-05-30 | 2023-12-07 | 哈尔滨工业大学（深圳） | Casualty recognition method and system based on deep learning in casualty gathering place scene |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109886085A (en) | People counting method based on deep learning target detection | |
| CN111784685B (en) | Power transmission line defect image identification method based on cloud edge cooperative detection | |
| CN109815886B (en) | Pedestrian and vehicle detection method and system based on improved YOLOv3 | |
| CN106845621B (en) | Dense population number method of counting and system based on depth convolutional neural networks | |
| Zhang et al. | Deep convolutional neural networks for forest fire detection | |
| Changzhen et al. | A traffic sign detection algorithm based on deep convolutional neural network | |
| CN108647742A (en) | Fast target detection method based on lightweight neural network | |
| CN108520516A (en) | A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation | |
| CN110188807A (en) | Tunnel pedestrian target detection method based on cascade super-resolution network and improvement Faster R-CNN | |
| CN103871077B (en) | A kind of extraction method of key frame in road vehicles monitoring video | |
| CN114973317B (en) | Pedestrian re-recognition method based on multi-scale adjacent interaction characteristics | |
| CN109801297B (en) | Image panorama segmentation prediction optimization method based on convolution | |
| CN112560675B (en) | Bird visual target detection method combining YOLO and rotation-fusion strategy | |
| CN107659754B (en) | Effective concentration method for monitoring video under condition of tree leaf disturbance | |
| CN113379771B (en) | Hierarchical human body analysis semantic segmentation method with edge constraint | |
| CN109919073B (en) | Pedestrian re-identification method with illumination robustness | |
| CN112270331A (en) | Improved billboard detection method based on YOLOV5 | |
| CN108198202A (en) | A kind of video content detection method based on light stream and neural network | |
| CN115512251A (en) | Unmanned aerial vehicle low-illumination target tracking method based on double-branch progressive feature enhancement | |
| CN113158891B (en) | Cross-camera pedestrian re-identification method based on global feature matching | |
| CN109934170B (en) | Mine resource statistical method based on computer vision | |
| CN109840904B (en) | Detection method for large-scale difference parts of high-speed rail contact network | |
| CN108734109A (en) | A kind of visual target tracking method and system towards image sequence | |
| CN111191610A (en) | People flow detection and processing method in video monitoring | |
| Sun et al. | Moving vehicle detection and tracking based on optical flow method and immune particle filter under complex transportation environments |
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: 20190614 |