CN109002752A - A kind of complicated common scene rapid pedestrian detection method based on deep learning - Google Patents
A kind of complicated common scene rapid pedestrian detection method based on deep learning Download PDFInfo
- Publication number
- CN109002752A CN109002752A CN201810021283.6A CN201810021283A CN109002752A CN 109002752 A CN109002752 A CN 109002752A CN 201810021283 A CN201810021283 A CN 201810021283A CN 109002752 A CN109002752 A CN 109002752A
- Authority
- CN
- China
- Prior art keywords
- training
- pedestrian
- neural networks
- convolutional neural
- box
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformation in the plane of the image
- G06T3/40—Scaling the whole image or part thereof
Abstract
The complicated common scene rapid pedestrian detection method based on deep learning that the present invention relates to a kind of, it include: that pixel size pretreatment is carried out to training image and test image, pre-training is carried out to convolutional neural networks based on classification task, pedestrian detection training is carried out to convolutional neural networks based on pedestrian detection task, the lower prediction box of confidence level is eliminated using threshold filtering, eliminates the multiforecasting to same a group traveling together using non-maximum suppression.In pre-training, using cross entropy as loss function.The regression result of network output prediction pedestrian position box is finally made as loss function using improved mean square error.It is filtered using threshold filtering and non-maximum suppression output prediction results all to convolutional neural networks to get to the location information of detection pedestrian using image as the input of convolutional neural networks in test phase, is achieved in the intelligent monitoring of pedestrian.
Description
Technical field
The present invention relates to image processing techniques, more particularly to a kind of quick row of the common scene based on convolutional neural networks
People's detection method.
Background technique
In recent years, monitoring camera was used in each public place, and the common scenes such as airport, station, hospital, road cover
Thousands of monitoring camera has been covered, has detected the pedestrian in common scene for the exception of analysis flow of the people, the discovery stream of people
Behavior track to specific crowd significant.It is past by manual analysis since the video data volume is huge and pedestrian is more
It is past to be difficult to rapidly and accurately analyze target pedestrian.And often speed is slower for existing some automatic pedestrian detection methods, it cannot
Complete the real time monitoring to pedestrian target.In order to realize the automatic real-time detection of pedestrian under common scene, one kind is studied public
Scene rapid pedestrian detection method is significant.
Summary of the invention
In view of this, the main purpose of the present invention is to provide a kind of detection accuracy height, detection are fireballing based on convolution
The common scene rapid pedestrian detection method of neural network, substantially increases detection accuracy, meanwhile, the reality to pedestrian may be implemented
When detect.
In order to achieve the above object, a kind of technical solution proposed by the present invention are as follows: public field based on convolutional neural networks
Scape rapid pedestrian detection method realizes that steps are as follows:
Step 1 reads database picture used in the training of tranining database picture, will using bilinear interpolation algorithm
Its pixel size stretches or boil down to fixed size A × B.
Step 2, using tranining database, the pre-training based on classification task is carried out to convolutional neural networks.Picture will be adjusted
The tranining database picture of plain size is compared by network output category result with input picture corresponding label as input
Compared with calculating loss function.Loss function is minimized, pre-training is carried out to convolutional neural networks.
Step 3, the picture of database for reading special scenes are stretched its pixel size using bilinear interpolation algorithm
Or boil down to fixed size.
Step 4 inherits weight obtained by pre-training network, changes convolutional neural networks end structure, uses special scenes
Database, is directed to the task of pedestrian detection, is adjusted training to neural network.In training using the image of special scenes as
The output of convolutional neural networks is carried out operation with the label of corresponding picture, calculates loss function by input.Minimize loss letter
Number, training convolutional neural networks.
Video is resolved into single frame, then is calculated using bilinear interpolation by pedestrian's video under step 5, reading common scene
Method stretches its pixel size or boil down to fixed size.
Step 6, using trained network, target detection is carried out to the pedestrian in picture.Pixel size will be adjusted
Picture to be measured is input in existing network, the feature of target object in image is extracted by convolutional neural networks, finally by two
The full articulamentum of layer exports the tensor of one 7 × 7 × (2 × 5) dimension.Tensor representation convolutional neural networks are pedestrian to be measured
98 prediction boxes out.
Step 7, set confidence level C threshold value C_threshold, to convolutional neural networks generate 98 prediction boxes into
Row filtering.Give up the prediction box that confidence level C is less than given threshold C_threshold.
Step 8, it is filtered using non-maximum suppression prediction box higher to degree of overlapping.When different prediction box intersections
When the ratio of area and union part area is more than defined threshold value IOU_threshold, then it is maximum only to retain confidence level C
Prediction box, and other boxes are inhibited.Prediction the block data x, y, w retained, h, C are the target detected
The spatial position coordinate and forecast confidence of pedestrian.
In the step 2, the process of pre-training convolutional neural networks is as follows:
The preceding 20 layers of convolutional layer and corresponding pond layer of network shown in step i) training Web vector graphic Fig. 1, then add later
Upper one layer of mean value pond layer and full articulamentum.
Step ii) rgb space corresponding to tranining database picture fixed size pixel by converted magnitude A × B
Input of the tensor data of × 3 dimensions as convolutional neural networks, exports the probability y for each classification resultsi。
Step iii) calculate network output probability yi' the cross entropy between label probabilityMake
For loss function, loss function loss is minimized, pre-training is carried out to network.
In the step 4, the method finally trained to convolutional neural networks based on pedestrian detection progress is as follows:
Step i) retains the structure of preceding 20 layers of convolutional layer and corresponding pond layer in pre-training network, and inherits its corresponding power
Value is increasing by 4 layers of convolutional layer and 2 layers of full articulamentum below, and is being randomly provided initial weight, keeps its network structure as shown in Figure 1.
Step ii) the full articulamentum of the last layer of network uses linear activation primitive: f (x)=x, and other full connection
Layer and convolutional layer use the line rectification activation primitive (Leaky ReLu) with leakage: f (x)=max (x, 0.1x).
Step iii) training sample be convert size special scenes database picture and its corresponding label.It will figure
Input of the tensor data that A × B × 3 of rgb space corresponding to piece A × B pixel is tieed up as convolutional neural networks.Neural network
Output be 7 × 7 × (2 × 5) tie up tensor.Indicate the 98 prediction boxes done to measured target.Each box has x, y,
This 5 data of w, h, C.
Step iv) database that reads special scenes correspond to the label of picture, and wherein pedestrian target is corresponding really for search
Block data calculates prediction block data x, y, w, h, C corresponding label data x ', y ', w ', h ' with network output,
C '=P × IOU.In calculating process, whole uniform picture is divided into 7 × 7 grid by imagination, if the true box of pedestrian target
Centre coordinate is fallen in some grid, then generates a group of labels x ', y ', w ', h ', C '=P × IOU.X ', y ' are true box
The coordinate of central point, value between 0~1, if the coordinate of true box central point where pedestrian in the lower left corner of corresponding grid,
Then its value is (0,0), if it in the upper right corner of grid, value is (1,1).W ', h ' are the length and width of true box, and value exists
Between 0~1, if the length of box or wide corresponding pixel size are 0, value 0, if the length of box or the corresponding pixel of width are big
Small is 448, then its value is 1.C '=P × IOU, wherein P=1, IOU are to predict box x, y, w, h and true box x ', y ', w ',
The intersection and union area ratio of h ' expression range.
Step v) calculates predicted value x, y, w, and h, C and the corresponding improvement of label value x ', y ', w ', h ', C '=P × IOU are square
Error loss function:
Wherein λcoord=5, λnoord=0.5, i indicate i-th in 7 × 7 grid, and j indicates that 2 of each grid are pre-
Survey j-th in box.If the centre coordinate of pedestrian target falls in i-th of grid and corresponding j-th of the prediction side of the grid
Frame and true pedestrian place box have maximum IOU, thenAndOtherwiseAndIt is minimum
Change loss function loss, network is trained.
In conclusion a kind of common scene rapid pedestrian detection method based on convolutional neural networks of the present invention,
Include: that single frames decomposition is carried out to the video under common scene, in decompositing the video frame come, uses the method for bilinear interpolation
Picture is converted into fixed pixel size.The training of pedestrian detection network is divided into pre-training and finally two processes of training,
In pre-training, use tranining database as training sample, be trained based on classification task, definition intersects entropy function as damage
Function is lost, whole network is trained by loss function, in final training process, inherits most of structure of pre-training network
And weight, network is improved, recurrence task is based on using the database of special scenes and is trained, is defined improved square
Error function trains whole network by minimizing loss function as loss function.During the test, with big after conversion
Small video frame exports all target pedestrian's prediction results by neural network as input, and output result is used threshold value
Filtering and non-maximum suppression are filtered, and are finally obtained the box for outlining pedestrian position, are achieved in the quick detection of pedestrian.
The advantages of the present invention over the prior art are that:
(1) present invention uses single convolutional neural networks, extracts to video frame images feature, in the detection of pedestrian
In the process, using video frame images as input, target line people position is outlined by the processing directly output of convolutional neural networks
Box.In trained and test process, it is all made of method end to end, therefore it is fast to detect speed.This method can be widely applied
In the complex scenes such as community, hospital, airport, station, school, to target, pedestrian makes real-time detection.
(2) present invention extracts the feature of video frame images using the convolutional neural networks in deep learning, uses training number
It is trained according to the database of library and special scenes.In the training process, convolutional neural networks have learnt various pedestrian's mesh
Target posture, and the high dimensional feature of pedestrian target is therefrom constantly extracted and learns, therefore, this method has generalization ability strong, Shandong
The strong feature of stick, can be applied to a variety of different scenes, and the target pedestrian different to macroscopic features detects.This
Detailed description of the invention
For implementation flow chart of the present invention.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, right below in conjunction with the accompanying drawings and the specific embodiments
The present invention is described in further detail.
The common scene intelligent video monitoring method that a kind of view-based access control model conspicuousness and depth of the present invention encode certainly,
Include: that single frames decomposition is carried out to the video under common scene, in decompositing the video frame come, uses the method for bilinear interpolation
Picture is converted into fixed pixel size.The training of pedestrian detection network is divided into pre-training and finally two processes of training,
In pre-training, use tranining database as training sample, be trained based on classification task, definition intersects entropy function as damage
Function is lost, whole network is trained by loss function, in final training process, inherits most of structure of pre-training network
And weight, network is improved, recurrence task is based on using the database of special scenes and is trained, is defined improved square
Error function trains whole network by minimizing loss function as loss function.During the test, with big after conversion
Small video frame exports all target pedestrian's prediction results by neural network as input, and output result is used threshold value
Filtering and non-maximum suppression are filtered, and are finally obtained the box for outlining pedestrian position, are achieved in the quick detection of pedestrian.
As shown in Figure 1, the present invention is implemented as follows step:
Step 1 reads database picture used in the training of tranining database picture, will using bilinear interpolation algorithm
Its pixel size stretches or boil down to A × B.
Step 2, using ImageNet database, the pre-training based on classification task is carried out to convolutional neural networks.It will adjust
The ImageNet database picture of whole pixel size is as input, by network output category result, mark corresponding with input picture
Label are compared, and calculate loss function.Loss function is minimized, pre-training is carried out to convolutional neural networks.
Step 3, the picture of database for reading special scenes are stretched its pixel size using bilinear interpolation algorithm
Or boil down to A × B.
Step 4 inherits weight obtained by pre-training network, changes convolutional neural networks end structure, uses special scenes
Database, is directed to the task of pedestrian detection, is adjusted training to neural network.In training using the image of special scenes as
The output of convolutional neural networks is carried out operation with the label of corresponding picture, calculates loss function by input.Minimize loss letter
Number, training convolutional neural networks.
Video is resolved into single frame, then is calculated using bilinear interpolation by pedestrian's video under step 5, reading common scene
Method stretches its pixel size or boil down to A × B.
Step 6, using trained network, target detection is carried out to the pedestrian in picture.Pixel size will be adjusted
Picture to be measured is input in existing network, the feature of target object in image is extracted by convolutional neural networks, finally by two
The full articulamentum of layer exports the tensor of one 7 × 7 × (2 × 5) dimension.Tensor representation convolutional neural networks are pedestrian to be measured
98 prediction boxes out.
Step 7, set confidence level C threshold value C_threshold, to convolutional neural networks generate 98 prediction boxes into
Row filtering.Give up the prediction box that confidence level C is less than given threshold C_threshold.
Step 8, it is filtered using non-maximum suppression prediction box higher to degree of overlapping.When different prediction box intersections
When the ratio of area and union part area is more than defined threshold value IOU_threshold, then it is maximum only to retain confidence level C
Prediction box, and other boxes are inhibited.Prediction the block data x, y, w retained, h, C are the target detected
The spatial position coordinate and forecast confidence of pedestrian.
In the step 2, the process of pre-training convolutional neural networks is as follows:
The preceding 20 layers of convolutional layer and corresponding pond layer of network shown in step i) training Web vector graphic Fig. 1, then add later
Upper one layer of mean value pond layer and full articulamentum.
Step ii) A × B × 3 of rgb space corresponding to A × B pixel of tranining database picture by converted magnitude ties up
Input of the tensor data as convolutional neural networks, export the probability y for each classification resultsi。
Step iii) calculate network output probability yi' the cross entropy between label probabilityMake
For loss function, loss function loss is minimized, pre-training is carried out to network.
In the step 4, the method finally trained to convolutional neural networks based on pedestrian detection progress is as follows:
Step i) retains the structure of preceding 20 layers of convolutional layer and corresponding pond layer in pre-training network, and inherits its corresponding power
Value is increasing by 4 layers of convolutional layer and 2 layers of full articulamentum below, and is being randomly provided initial weight, keeps its network structure as shown in Figure 1.
Step ii) the full articulamentum of the last layer of network uses linear activation primitive: f (x)=x, and other full connection
Layer and convolutional layer use the line rectification activation primitive (Leaky ReLu) with leakage: f (x)=max (x, 0.1x).
Step iii) training sample be convert size special scenes database picture and its corresponding label.It will figure
Input of the tensor data of 448 × 448 × 3 dimensions of rgb space corresponding to 448 × 448 pixel of piece as convolutional neural networks.
The output of neural network is that tensor is tieed up in 7 × 7 × (2 × 5).Indicate the 98 prediction boxes done to measured target.Each side
Frame has x, y, w, this 5 data of h, C.
Step iv) database that reads special scenes correspond to the label of picture, and wherein pedestrian target is corresponding really for search
Block data calculates prediction block data x, y, w, h, C corresponding label data x ', y ', w ', h ' with network output,
C '=P × IOU.In calculating process, whole uniform picture is divided into 7 × 7 grid by imagination, if the true box of pedestrian target
Centre coordinate is fallen in some grid, then generates a group of labels x ', y ', w ', h ', C '=P × IOU.X ', y ' are true box
The coordinate of central point, value between 0~1, if the coordinate of true box central point where pedestrian in the lower left corner of corresponding grid,
Then its value is (0,0), if it in the upper right corner of grid, value is (1,1).W ', h ' are the length and width of true box, and value exists
Between 0~1, if the length of box or wide corresponding pixel size are 0, value 0, if the length of box or the corresponding pixel of width are big
Small is 448, then its value is 1.C '=P × IOU, wherein P=1, IOU are to predict box x, y, w, h and true box x ', y ', w ',
The intersection and union area ratio of h ' expression range.
Step v) calculates predicted value x, y, w, and h, C and the corresponding improvement of label value x ', y ', w ', h ', C '=P × IOU are square
Error loss function:
Wherein λcoord=5, λnoord=0.5, i indicate i-th in 7 × 7 grid, and j indicates that 2 of each grid are pre-
Survey j-th in box.If the centre coordinate of pedestrian target falls in i-th of grid and corresponding j-th of the prediction side of the grid
Frame and true pedestrian place box have maximum IOU, thenAndOtherwiseAndIt is minimum
Change loss function loss, network is trained.
In conclusion the above is merely preferred embodiments of the present invention, being not intended to limit the scope of the present invention.
All within the spirits and principles of the present invention, any modification, equivalent replacement, improvement and so on should be included in of the invention
Within protection scope.
Claims (3)
1. a kind of complicated common scene rapid pedestrian detection method based on deep learning, it is characterised in that realize that steps are as follows:
Step 1 reads database picture used in the training of ImageNet database picture, will using bilinear interpolation algorithm
Its pixel size stretches or boil down to A × B.
Step 2, using tranining database, the pre-training based on classification task is carried out to convolutional neural networks.It is big pixel will to be adjusted
Small tranining database picture is compared with input picture corresponding label, is counted by network output category result as input
Calculate loss function.Loss function is minimized, pre-training is carried out to convolutional neural networks.
Its pixel size is stretched or is pressed using bilinear interpolation algorithm by step 3, the picture of database for reading special scenes
It is condensed to A × B.
Step 4 inherits weight obtained by pre-training network, changes convolutional neural networks end structure, uses the data of special scenes
Library, is directed to the task of pedestrian detection, is adjusted training to neural network.Using the image of special scenes as defeated in training
Enter, the output of convolutional neural networks is subjected to operation with the label of corresponding picture, calculates loss function.Loss function is minimized,
Training convolutional neural networks.
Video is resolved into single frame, then uses bilinear interpolation algorithm by pedestrian's video under step 5, reading common scene, will
Its pixel size stretches or boil down to A × B.
Step 6, using trained network, target detection is carried out to the pedestrian in picture.The to be measured of pixel size will be adjusted
Picture is input in existing network, and the feature of target object in image is extracted by convolutional neural networks, complete finally by two layers
Articulamentum exports the tensor of one 7 × 7 × (2 × 5) dimension.What tensor representation convolutional neural networks made pedestrian to be measured
98 prediction boxes.
Step 7, the threshold value C_threshold for setting confidence level C, the 98 prediction boxes generated to convolutional neural networks carried out
Filter.Give up the prediction box that confidence level C is less than given threshold C_threshold.
Step 8, it is filtered using non-maximum suppression prediction box higher to degree of overlapping.When different prediction box intersections part
When the ratio of area and union part area is more than defined threshold value IOU_threshold, then it is maximum pre- only to retain confidence level C
Box is surveyed, and other boxes are inhibited.Prediction the block data x, y, w retained, h, C are the target pedestrian detected
Spatial position coordinate and forecast confidence.
2. a kind of pedestrian detection method based on single convolutional neural networks according to claim 1, it is characterised in that: described
In step 2, the process of pre-training convolutional neural networks is as follows:
The preceding 20 layers of convolutional layer and corresponding pond layer of network shown in step i) training Web vector graphic Fig. 1, then one is added later
Layer mean value pond layer and full articulamentum.
Step ii) rgb space corresponding to 224 × 224 pixels of tranining database picture by converted magnitude 224 × 224
Input of the tensor data of × 3 dimensions as convolutional neural networks, exports the probability y for each classification resultsi。
Step iii) calculate network output probability yi' the cross entropy between label probabilityAs damage
Function is lost, loss function loss is minimized, pre-training is carried out to network.
3. a kind of pedestrian detection method based on single convolutional neural networks according to claim 1, it is characterised in that: described
In step 4
Step i) retains the structure of preceding 20 layers of convolutional layer and corresponding pond layer in pre-training network, and inherits its corresponding weight,
Increasing by 4 layers of convolutional layer and 2 layers of full articulamentum below, and be randomly provided initial weight, is keeping its network structure as shown in Figure 1.
Step ii) the full articulamentum of the last layer of network uses linear activation primitive: f (x)=x, and other full articulamentum and
Convolutional layer uses the line rectification activation primitive (Leaky ReLu) with leakage: f (x)=max (x, 0.1x).
Step iii) training sample be convert size special scenes database picture and its corresponding label.By picture A
Input of the tensor data of 448 × 448 × 3 dimensions of rgb space corresponding to × B pixel as convolutional neural networks.Nerve net
The output of network is that tensor is tieed up in 7 × 7 × (2 × 5).Indicate the 98 prediction boxes done to measured target.Each box has x,
This 5 data of y, w, h, C.
Step iv) database that reads special scenes correspond to the label of picture, search for the wherein corresponding true box of pedestrian target
Data calculate prediction block data x, y, w, h, C corresponding label data x ', y ', w ', h ', C '=P with network output
×IOU.In calculating process, whole uniform picture is divided into 7 × 7 grid by imagination, if the center of the true box of pedestrian target
Coordinate is fallen in some grid, then generates a group of labels x ', y ', w ', h ', C '=P × IOU.X ', y ' are true box center
The coordinate of point, value between 0~1, if the coordinate of true box central point where pedestrian in the lower left corner of corresponding grid,
Value is (0,0), if it in the upper right corner of grid, value is (1,1).W ', h ' are the length and width of true box, and value is 0~1
Between, if the length of box or wide corresponding pixel size are 0, value 0, if the length of box or the corresponding pixel size of width are
448, then its value is 1.C '=P × IOU, wherein P=1, IOU are prediction box x, y, w, h and true box x ', y ', w ', h ' table
The intersection and union area ratio that demonstration is enclosed.
Step v) calculates predicted value x, y, w, h, C and the corresponding improvement mean square error of label value x ', y ', w ', h ', C '=P × IOU
Loss function:
Wherein λcoord=5, λnoord=0.5, i indicate i-th in 7 × 7 grid, and j indicates 2 prediction boxes of each grid
In j-th.If the centre coordinate of pedestrian target falls in i-th of grid and corresponding j-th of the prediction box of the grid and true
Box where carrying out people has maximum IOU, thenAndOtherwiseAndMinimize loss
Function loss, is trained network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810021283.6A CN109002752A (en) | 2018-01-08 | 2018-01-08 | A kind of complicated common scene rapid pedestrian detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810021283.6A CN109002752A (en) | 2018-01-08 | 2018-01-08 | A kind of complicated common scene rapid pedestrian detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109002752A true CN109002752A (en) | 2018-12-14 |
Family
ID=64573523
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810021283.6A Pending CN109002752A (en) | 2018-01-08 | 2018-01-08 | A kind of complicated common scene rapid pedestrian detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109002752A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109766790A (en) * | 2018-12-24 | 2019-05-17 | 重庆邮电大学 | A kind of pedestrian detection method based on self-adaptive features channel |
| CN109829429A (en) * | 2019-01-31 | 2019-05-31 | 福州大学 | Security protection sensitive articles detection method under monitoring scene based on YOLOv3 |
| CN110110844A (en) * | 2019-04-24 | 2019-08-09 | 西安电子科技大学 | Convolutional neural networks method for parallel processing based on OpenCL |
| CN110321811A (en) * | 2019-06-17 | 2019-10-11 | 中国工程物理研究院电子工程研究所 | Depth is against the object detection method in the unmanned plane video of intensified learning |
| CN110322509A (en) * | 2019-06-26 | 2019-10-11 | 重庆邮电大学 | Object localization method, system and computer equipment based on level Class Activation figure |
| CN111160263A (en) * | 2019-12-30 | 2020-05-15 | 中国电子科技集团公司信息科学研究院 | Method and system for obtaining face recognition threshold |
| CN112115862A (en) * | 2020-09-18 | 2020-12-22 | 广东机场白云信息科技有限公司 | Crowded scene pedestrian detection method combined with density estimation |
| CN112580778A (en) * | 2020-11-25 | 2021-03-30 | 江苏集萃未来城市应用技术研究所有限公司 | Job worker mobile phone use detection method based on YOLOv5 and Pose-animation |
| CN114245140A (en) * | 2021-11-30 | 2022-03-25 | 慧之安信息技术股份有限公司 | Code stream prediction method and device based on deep learning |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106022237A (en) * | 2016-05-13 | 2016-10-12 | 电子科技大学 | Pedestrian detection method based on end-to-end convolutional neural network |
| CN106127815A (en) * | 2016-07-21 | 2016-11-16 | 广东工业大学 | A kind of tracking merging convolutional neural networks and system |
| CN106228575A (en) * | 2016-07-21 | 2016-12-14 | 广东工业大学 | Merge convolutional neural networks and the tracking of Bayesian filter and system |
| CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
| CN106650919A (en) * | 2016-12-23 | 2017-05-10 | 国家电网公司信息通信分公司 | Information system fault diagnosis method and device based on convolutional neural network |
| CN106845549A (en) * | 2017-01-22 | 2017-06-13 | 珠海习悦信息技术有限公司 | A kind of method and device of the scene based on multi-task learning and target identification |
| CN107203740A (en) * | 2017-04-24 | 2017-09-26 | 华侨大学 | A kind of face age estimation method based on deep learning |
-
2018
- 2018-01-08 CN CN201810021283.6A patent/CN109002752A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106022237A (en) * | 2016-05-13 | 2016-10-12 | 电子科技大学 | Pedestrian detection method based on end-to-end convolutional neural network |
| CN106127815A (en) * | 2016-07-21 | 2016-11-16 | 广东工业大学 | A kind of tracking merging convolutional neural networks and system |
| CN106228575A (en) * | 2016-07-21 | 2016-12-14 | 广东工业大学 | Merge convolutional neural networks and the tracking of Bayesian filter and system |
| CN106650919A (en) * | 2016-12-23 | 2017-05-10 | 国家电网公司信息通信分公司 | Information system fault diagnosis method and device based on convolutional neural network |
| CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
| CN106845549A (en) * | 2017-01-22 | 2017-06-13 | 珠海习悦信息技术有限公司 | A kind of method and device of the scene based on multi-task learning and target identification |
| CN107203740A (en) * | 2017-04-24 | 2017-09-26 | 华侨大学 | A kind of face age estimation method based on deep learning |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109766790B (en) * | 2018-12-24 | 2022-08-23 | 重庆邮电大学 | Pedestrian detection method based on self-adaptive characteristic channel |
| CN109766790A (en) * | 2018-12-24 | 2019-05-17 | 重庆邮电大学 | A kind of pedestrian detection method based on self-adaptive features channel |
| CN109829429A (en) * | 2019-01-31 | 2019-05-31 | 福州大学 | Security protection sensitive articles detection method under monitoring scene based on YOLOv3 |
| CN110110844A (en) * | 2019-04-24 | 2019-08-09 | 西安电子科技大学 | Convolutional neural networks method for parallel processing based on OpenCL |
| CN110321811A (en) * | 2019-06-17 | 2019-10-11 | 中国工程物理研究院电子工程研究所 | Depth is against the object detection method in the unmanned plane video of intensified learning |
| CN110321811B (en) * | 2019-06-17 | 2023-05-02 | 中国工程物理研究院电子工程研究所 | Target detection method in unmanned aerial vehicle aerial video for deep reverse reinforcement learning |
| CN110322509A (en) * | 2019-06-26 | 2019-10-11 | 重庆邮电大学 | Object localization method, system and computer equipment based on level Class Activation figure |
| CN111160263A (en) * | 2019-12-30 | 2020-05-15 | 中国电子科技集团公司信息科学研究院 | Method and system for obtaining face recognition threshold |
| CN111160263B (en) * | 2019-12-30 | 2023-09-05 | 中国电子科技集团公司信息科学研究院 | Method and system for acquiring face recognition threshold |
| CN112115862A (en) * | 2020-09-18 | 2020-12-22 | 广东机场白云信息科技有限公司 | Crowded scene pedestrian detection method combined with density estimation |
| CN112115862B (en) * | 2020-09-18 | 2023-08-29 | 广东机场白云信息科技有限公司 | Congestion scene pedestrian detection method combined with density estimation |
| CN112580778A (en) * | 2020-11-25 | 2021-03-30 | 江苏集萃未来城市应用技术研究所有限公司 | Job worker mobile phone use detection method based on YOLOv5 and Pose-animation |
| CN114245140A (en) * | 2021-11-30 | 2022-03-25 | 慧之安信息技术股份有限公司 | Code stream prediction method and device based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109002752A (en) | A kind of complicated common scene rapid pedestrian detection method based on deep learning | |
| Tao et al. | Smoke detection based on deep convolutional neural networks | |
| CN104077613B (en) | Crowd density estimation method based on cascaded multilevel convolution neural network | |
| CN104978580B (en) | A kind of insulator recognition methods for unmanned plane inspection transmission line of electricity | |
| CN107358257B (en) | Under a kind of big data scene can incremental learning image classification training method | |
| CN110135319A (en) | A kind of anomaly detection method and its system | |
| CN106407903A (en) | Multiple dimensioned convolution neural network-based real time human body abnormal behavior identification method | |
| CN110147743A (en) | Real-time online pedestrian analysis and number system and method under a kind of complex scene | |
| CN110287960A (en) | The detection recognition method of curve text in natural scene image | |
| CN107016357A (en) | A kind of video pedestrian detection method based on time-domain convolutional neural networks | |
| CN110188835B (en) | Data-enhanced pedestrian re-identification method based on generative confrontation network model | |
| CN108921875A (en) | A kind of real-time traffic flow detection and method for tracing based on data of taking photo by plane | |
| CN109241913A (en) | In conjunction with the ship detection method and system of conspicuousness detection and deep learning | |
| CN109902806A (en) | Method is determined based on the noise image object boundary frame of convolutional neural networks | |
| CN108717528A (en) | A kind of global population analysis method of more strategies based on depth network | |
| CN106127204A (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN106683091A (en) | Target classification and attitude detection method based on depth convolution neural network | |
| CN108334847A (en) | A kind of face identification method based on deep learning under real scene | |
| CN106408015A (en) | Road fork identification and depth estimation method based on convolutional neural network | |
| CN108764085A (en) | Based on the people counting method for generating confrontation network | |
| CN104504395A (en) | Method and system for achieving classification of pedestrians and vehicles based on neural network | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN110347870A (en) | The video frequency abstract generation method of view-based access control model conspicuousness detection and hierarchical clustering method | |
| CN107038416A (en) | A kind of pedestrian detection method based on bianry image modified HOG features | |
| CN103984963B (en) | Method for classifying high-resolution remote sensing image scenes |
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: 20181214 |