CN108229390A - Rapid pedestrian detection method based on deep learning - Google Patents
Rapid pedestrian detection method based on deep learning Download PDFInfo
- Publication number
- CN108229390A CN108229390A CN201810001077.9A CN201810001077A CN108229390A CN 108229390 A CN108229390 A CN 108229390A CN 201810001077 A CN201810001077 A CN 201810001077A CN 108229390 A CN108229390 A CN 108229390A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- image
- data set
- model
- pedestrian detection
- 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
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The rapid pedestrian detection method based on deep learning of the present invention, includes the following steps:(1)Pedestrian data set of the structure with markup information:The monitor video image under different scenes is acquired, pedestrian detection standard data set is built, and data set is expanded, includes image and its markup information;(2)Using pedestrian's data set of structure to network model pre-training, generation is suitable for the model of pedestrian detection:Pre-training model is finely adjusted using pedestrian's data set, is trained for the pedestrian detection model suitable for constructed data set;(3)The test of the quick pedestrian detection of multi-channel video is carried out to trained model:It based on multi-channel video joining method, treats test image and is pre-processed, input trained model and be detected, obtain the location information of target.The method of the present invention can reduce the influence to pedestrian detection such as illumination, complex background, quickly detect pedestrian, monitor video is analyzed automatically, note abnormalities event in time.
Description
Technical field
The present invention relates to image procossing, video monitoring and security protection be applied to pedestrian detection technical field, specially one
Rapid pedestrian detection method of the kind based on deep learning.
Background technology
Pedestrian detection is exactly computer for given image and video, judges wherein whether there is pedestrian, if there is also
Need to provide the specific location of pedestrian.Pedestrian detection is pedestrian tracking, behavioural analysis, gait analysis, and pedestrian's identification etc. is ground
The basis studied carefully and premise, a good pedestrian detection algorithm can provide strong support and guarantee for the latter.In actual life,
Pedestrian detection is widely used scene.In recent years, numerous individual such as entrance, office, market, bank and public place
Monitoring camera is assembled with, but the work monitored is most to be completed by artificial monitoring.However the work manually monitored
It not only needs to expend a large amount of manpower, and since this work is easy to make observer tired out so that observer may leak
Fall some important information, so as to bring huge economic loss.Using pedestrian detection technology, we, which can realize, allows computer
Automatic each pedestrian detected under monitoring camera.Can also realize based on this specified pedestrian track is carried out with
Track is identified pedestrian's identity and pedestrian behavior is analyzed etc., if it find that the timely automated alarm of unusual condition.This
Sample can greatly reduce manpower, can more improve monitoring accuracy and prevent trouble before it happens, reach us and carry out the original intention of video monitoring.
Human body has both flexible with rigid feature so that pedestrian detection easily by clothes, illumination, block, pedestrian's posture, ruler
The restriction of the complicated factors such as degree, shooting angle, and it is unstable.
Currently used pedestrian detection method includes:Background difference, frame difference method, optical flow method, template matching method and based on machine
Method of device study etc..Aforementioned four kinds of methods are all based on the human body detecting method of image processing techniques, however these methods exist
In face of human body clothing, the bodily form, human posture, human body block, lighting issues and complex background etc. problem the problem of when, institute
Reflected accuracy rate and robustness are very poor.Based on the method for machine learning from the rule of training sample study human body, obtain
Then model is tested on test set.If data and feature can be selected reasonably, it is subject to rational algorithm to carry out
Training, the problems such as can preferably overcoming such as human body diversity, illumination, background diversity, it has also become pedestrian detection now
A kind of mainstream algorithm.
For security protection industry, the quantity of monitor video is more and more, is provided largely to the development of deep learning
Data, but for deep learning, computing capability is a major limitation problem, the time for how saving detection is non-
It is often important.
Invention content
In order to compensate for the shortcomings of the prior art, the present invention provides a kind of simple in structure, easy to use based on depth
The rapid pedestrian detection method of habit can reduce the influence to pedestrian detection such as illumination, complex background, quickly detect pedestrian, from
Dynamic that monitor video is analyzed, note abnormalities event in time, improves the efficiency and robustness of pedestrian detection.
The present invention is achieved through the following technical solutions:
The rapid pedestrian detection method based on deep learning of the present invention, it is characterised in that:Include the following steps:
(1)Pedestrian data set of the structure with markup information:
The monitor video image under different scenes is acquired, builds pedestrian detection standard data set, and data set is expanded, is wrapped
Containing image and its markup information;
(2)Using pedestrian's data set of structure to network model pre-training, generation is suitable for the model of pedestrian detection:
Pre-training model is finely adjusted using data set ready in step 1, is trained for suitable for constructed data set
Pedestrian detection model;
(3)The test of the quick pedestrian detection of multi-channel video is carried out to trained model:
It based on multi-channel video joining method, treats test image and is pre-processed, input trained model and be detected, obtain
To the location information of target.
Step(1)In:Standard pedestrian data set of the structure with markup information, mainly including following three parts:
Pedestrian's picture image under 1-1, acquisition different scenes;Comprising a variety of pedestrian's postures, several scenes, a variety of illumination effects with
And in one day different periods video image so that the data set of construction is capable of the diversity of effective covering scene;
1-2, collected data set is effectively expanded, by image mirrors, angle rotation, size scaling, addition with
The method of machine noise;
1-3, the data images after expansion are generated with corresponding mark and label information, target in markup information, that is, sample
Location information, target generic in label information, that is, sample, classification behaviour are labeled as 1 or inhuman and are labeled as -1.
Step(2)In:Utilize model of the pedestrian's data set training of structure based on Faster R-CNN:
By the labeled data collection of above-mentioned structure, pre-training model is finely adjusted using Faster R-CNN network structures, is instructed
Practice the model suitable for pedestrian detection;Faster R-CNN networks include the RPN convolutional neural networks for generating candidate region
With the Fast R-CNN convolutional neural networks composition for target-recognition.
Step(3)In:The test of the quick pedestrian detection of multi-channel video is carried out to trained model, mainly comprising following
Three parts:
3-1, image to be tested is generated using multi-channel video frame joining method
By the way that the monitor video image of multichannel is spliced, a large-sized image is combined into, it then will be spliced
Image is input in trained model and is detected, and the video frame for being equivalent to multiple channels is carried out at the same time detection, can be effective
Save the detection time of video to be detected;
3-2, the influence that ambient enviroment is reduced using video frame preprocess method
In detection process, the main purpose of image preprocessing is the redundancy eliminated in image, filters out interference, noise, is increased
Detectability for information about by force, so as to improve the reliability of subsequent characteristics extraction and detection;
Available image processing method mainly includes following methods:Histogram equalization, normalization, Gamma corrections;
3-3, it will be input in trained model by pretreated splicing frame and carry out pedestrian detection, output pedestrian position letter
Breath and corresponding confidence level.
Step(2)In:It is finely adjusted on the data set of structure using existing ImageNet models, obtains being suitable for row
The new disaggregated model of people's detection, key step are as follows:
2-1 carries out pre-training using the ZF-Net networks of ImageNet models, and the full articulamentum of the network model is carried out first
Modification, output class are two classes, i.e. pedestrian and non-pedestrian;
2-2 pre-training models on the sample set of structure, Faster R-CNN include the fine tuning RPN nets for extracting candidate region
Network and the Fast R-CNN networks for detecting target;RPN networks handle the image comprising pedestrian of input, generation
Pedestrian's roughing region;And Fast R-CNN networks further differentiate these pedestrian's roughing regions, output pedestrian is final
Location information;RPN networks and Fast R-CNN networks can sharing feature, reduce the time of extraction candidate region, shorten
Time of target detection;
Using the method for pre-training primarily to a good netinit value is obtained, to avoid sunken in subsequent training
Enter local minimum, while also be able to accelerate the convergence rate of network;The follow-up sample set using structure is finely adjusted network,
To ensure that parameter is more suitable for current sample set, the verification and measurement ratio on data set is improved.
Step(3)In:
It in 3-2, is corrected by quick Gamma, each color channel of image is standardized in rgb color space;
Gamma corrects the contrast for having adjusted image, reduces the influence caused by image local shade and illumination variation, alleviates field
The influence of scape and illumination to feature extraction.
Step(3)In, in 3-2, Gamma correction methods, it is assumed that the pixel value of all images is the integer between 0 ~ M, to one
The pixel that a pixel value is i performs following operate:
3-2-1 is normalized:Pixel value is converted into the real number j between 0 ~ 1,;
3-2-2 is pre-compensated for:Pre-compensation value, wherein Gamma represents the gamma values chosen, in practical application
In, it should gamma values are adjusted in a certain range according to actual conditions, to obtain optimum efficiency;
3-2-3 renormalizations:The integer value between 0 ~ M will be changed to by the real number value contravariant of precompensation,;Aforesaid operations are performed to each pixel in image, i.e., can be obtained by Gamma
Image value after transformation;Simple way is by the corresponding Gamma changing values of all integer calculations between 0 ~ M, is corresponded to
Value be stored in the Gamma pre-established a correction look-up table, using the table to image of any pixel value between 0 ~ M into
Row Gamma is corrected.
The invention has the advantages that
1. the present invention on the basis of existing standard pedestrian's data set, increases the data under current environment, by acquiring a variety of rings
Video image under border, and data set is effectively expanded by the method that a variety of data enhance, not only increase data
Amount, at the same also cover it is more under the conditions of data, effectively prevent network over-fitting, improve the accuracy of pedestrian detection.
2. method of the present invention by effectively being spliced to multi-channel video frame, effectively increases multi-channel video pedestrian
The speed of detection.Interior at the same time, detection method proposed by the present invention can detect multi-channel video simultaneously, compared to tradition one
The detection method of road video, effectively saves detection time.
3. by the present invention in that with image pretreatment operation, the quality of image to be detected under varying environment is effectively had adjusted,
The influence caused by image local shade and illumination variation is reduced, alleviates the shadow of scene and illumination to pedestrian's feature extraction
It rings, improves the robustness of pedestrian detection.
4. by using the network model of deep learning, solve the deficiency of manual extraction characteristics of image, improve pedestrian
The accuracy rate of detection.
Description of the drawings
Fig. 1 is whole 3 steps flow chart schematic diagrams of invention.Fig. 2 is the step of the present invention(1)Flow show
It is intended to.Fig. 3 is the step of the present invention(3)Flow diagram.Fig. 4 is the whole detailed flow diagram of the present invention.
Specific embodiment
A kind of specific embodiment of the attached drawing for the present invention.
A kind of rapid pedestrian detection method based on deep learning disclosed in this invention is mainly comprising establishment sample set, instruction
Practice and quickly detect three phases.Key content wherein of the invention is quick detection-phase, and the part is by image to be tested
Spliced, and carry out the influence caused by image procossing reduction image local shade and illumination variation, improve pedestrian detection
Speed and robustness.
Step is as follows:
(1)Pedestrian data set of the structure with markup information;
The monitor video image under different scenes is acquired, builds pedestrian detection standard data set, and data set is expanded, is wrapped
Containing image and its markup information.
(2)Using pedestrian's data set of structure to network model pre-training, generation is suitable for the model of pedestrian detection;
Pre-training model is finely adjusted using data set ready in step 1, is trained for suitable for constructed data set
Pedestrian detection model.
(3)The test of the quick pedestrian detection of multi-channel video is carried out to trained model.
It based on multi-channel video joining method, treats test image and is pre-processed, input trained model and examined
It surveys, obtains the location information of target.
Step(1)In:Standard pedestrian data set of the structure with markup information, mainly including following three parts:
Pedestrian's picture image under 1-1 acquisition different scenes;
1-2 effectively expands collected data set, such as passes through the side such as mirror image, rotation, scaling, addition random noise
Method;
1-3 gives birth to the markup information of paired data collection image.
Step(2)In:Utilize model of the pedestrian's data set training of structure based on Faster R-CNN:
By the labeled data collection of above-mentioned structure, pre-training model is finely adjusted using Faster R-CNN network structures, is instructed
Practice the model suitable for pedestrian detection.Faster R-CNN networks include the RPN convolutional neural networks for generating candidate region
With the Fast R-CNN convolutional neural networks composition for target-recognition.
Step(3):The test of the quick pedestrian detection of multi-channel video is carried out to trained model, mainly comprising following three
Part:
3-1 generates image to be tested using multi-channel video frame splicing
For security protection industry, the quantity of monitor video is more and more, and a large amount of data are provided to the development of deep learning,
But for deep learning, computing capability is a major limitation problem, and the time for how saving detection is very heavy
It wants.
By the way that the video image of multichannel is spliced, a large-sized image is combined into, it then will be spliced
Image is input in trained model and is detected, and can effectively save the detection time of video to be detected.Assuming that by N number of logical
The video frame in road is spliced into a frame image, is input in model, and the video frame for being equivalent to N number of channel is carried out at the same time detection, can have
Effect saves detection time.
3-2 reduces the influence of ambient enviroment using video frame preprocess method
In detection process, image to be detected of input can lead to picture quality because of limitation and the random disturbances of various conditions
Difference, direct use cannot get a desired effect.The main purpose of image preprocessing is the redundancy eliminated in image, is filtered out
Interference, noise, enhance detectability for information about, so as to improve the reliability of subsequent characteristics extraction and detection.
The present invention is corrected by quick Gamma, and standard is carried out to each color channel of image in rgb color space
Change.Gamma corrects the contrast for having adjusted image, reduces the influence caused by image local shade and illumination variation, mitigates
The influence of scene and illumination to feature extraction.
3-3 will be input in trained model by pretreated splicing frame and carry out pedestrian detection, output pedestrian position
Confidence ceases and corresponding confidence level.
Step(1)In, pedestrian data set of the structure with markup information.
Monitor video image under 1-1 acquisition different scenes, as possible comprising a variety of pedestrian's postures, several scenes, Duo Zhongguang
According to influence and in one day different periods video image so that the data set of construction is capable of the diversity of effective covering scene.
For 1-2 for deep learning, big data is being effectively ensured for its effect, and collected sample data set is passed through
Different methods are effectively expanded.
Common data set extending method is including but not limited to following:Image mirrors, angle rotation, size scaling and addition
The modes such as random noise.
Mark and label information corresponding to the data set generation after expansion 1-3, target in markup information, that is, sample
Location information, target generic in label information, that is, sample, classification behaviour in the present invention is labeled as 1 or inhuman being labeled as-
1。
Step(2)In, pre-training is carried out to the model based on Faster R-CNN algorithms using pedestrian's data set of structure,
Generate final pedestrian detection model.
It is finely adjusted, obtained suitable for the new of pedestrian detection on the data set of structure using existing ImageNet models
Disaggregated model.Key step is as follows:
Pre-training is carried out using the ZF-Net networks of ImageNet models in the 2-1 present invention, the complete of the network model is connected first
It connects layer to modify, output class is two classes, i.e. pedestrian and non-pedestrian;
2-2 pre-training models on the sample set of structure, including fine tuning RPN networks and Fast R-CNN networks.RPN networks pair
The image comprising pedestrian of input is handled, generation pedestrian's roughing region;And Fast R-CNN networks are to these pedestrian's roughings
Region is further differentiated, exports the final location information of pedestrian.
Using the method for pre-training primarily to a good netinit value is obtained, to avoid subsequent training
In be absorbed in local minimum, while also be able to accelerate network convergence rate.The follow-up sample set using structure carries out network
Fine tuning, to ensure that parameter is more suitable for current sample set, improves the verification and measurement ratio on data set.
Step(2)Used model is realized based on Faster RCNN frames.Faster RCNN are included and are used for
Extract the RPN networks of candidate region and the Fast RCNN networks for detecting target.RPN networks and Fast RCNN networks can
Sharing feature greatly reduces the time of extraction candidate region, shortens the time of target detection.
Step(3)In, the test of the quick pedestrian detection of multi-channel video is carried out to trained pedestrian detection model.
3-1, image to be tested is generated using picture frame splicing to multi-channel video frame
With safety monitoring high definition, intelligent, integrated fast development, monitor video resource is more and more, and monitoring big data is
The development of deep learning provides fertile soil.In addition to big data, another main problem of influence depth study is computing capability
Limitation, the deep learning based on big data need stronger computing capability, and it is extremely important for how saving the time of detection
's.
For pedestrian detection, by the way that the feature of multi-path monitoring video is effectively spliced, be combined into one it is to be measured
Attempt picture, spliced image is input in trained model and is detected, can effectively save multi-channel video pedestrian inspection
The time of survey.
Assuming that the video frame of N number of channel is spliced into a frame image, it is input in model, carries out pedestrian detection, be equivalent to N
The video of a channel is carried out at the same time detection, originally the time spent in one channel of detection, can detect regarding for N number of channel simultaneously
Frequently, the detection time of multi-channel video is effectively saved.In the present embodiment, the value of N needs to consider the splicing of multi-channel video frame
Influence of the image afterwards to pedestrian detection, while in view of splicing frame to factors such as the deformation of primitive frame, it is proposed that value is nature
Several squares.Value suggestion in the present embodiment is it is not limited to N values are 4.
3-2, in view of spliced image in different illumination and varying environment, the pretreatment operation for carrying out video frame subtracts
The influence of few ambient enviroment.
In detection process, image to be detected of input can lead to image because of limitation and the random disturbances of various conditions
Of poor quality, direct use cannot get a desired effect.The main purpose of image preprocessing is the redundancy eliminated in image,
Interference, noise are filtered out, enhances detectability for information about, so as to improve the reliability of subsequent characteristics extraction and detection.
Available image processing method includes but not limited to following methods:Histogram equalization, normalization, Gamma corrections
The methods of.In the present embodiment by taking Gamma is corrected as an example, key step is introduced:
Quick Gamma bearing calibrations, are standardized each color channel of image in rgb color space.Gamma is corrected
The contrast of image is had adjusted, the influence caused by image local shade and illumination variation is reduced, alleviates scene and illumination
Influence to feature extraction.
Assuming that the pixel value of all images is the integer between 0 ~ M, following grasp is performed to the pixel that a pixel value is i
Make:
3-2-1 is normalized:Pixel value is converted into the real number j between 0 ~ 1,;
3-2-2 is pre-compensated for:Pre-compensation value, wherein Gamma represents the gamma values chosen, in practical application
In, it should gamma value is adjusted in a certain range according to actual conditions, to obtain optimum efficiency;
3-2-3 renormalizations:The integer value between 0 ~ M will be changed to by the real number value contravariant of precompensation,.To in image each pixel perform an aforesaid operations, i.e., can obtain by
Image value after Gamma transformation.Simple way is by the corresponding Gamma changing values of all integer calculations between 0 ~ M, by it
Corresponding value is stored in the Gamma pre-established a correction look-up table, uses figure of the table to any pixel value between 0 ~ M
As carrying out Gamma corrections.
3-3 is last, will splice and is input in network model by the stitching image of image procossing, carries out pedestrian detection,
Export position and the confidence score of pedestrian.
Claims (7)
1. a kind of rapid pedestrian detection method based on deep learning, it is characterised in that:Include the following steps:
(1)Pedestrian data set of the structure with markup information:
The monitor video image under different scenes is acquired, builds pedestrian detection standard data set, and data set is expanded, is wrapped
Containing image and its markup information;
(2)Using pedestrian's data set of structure to network model pre-training, generation is suitable for the model of pedestrian detection:
Pre-training model is finely adjusted using data set ready in step 1, is trained for suitable for constructed data set
Pedestrian detection model;
(3)The test of the quick pedestrian detection of multi-channel video is carried out to trained model:
It based on multi-channel video joining method, treats test image and is pre-processed, input trained model and be detected, obtain
To the location information of target.
2. the rapid pedestrian detection method according to claim 1 based on deep learning, it is characterised in that:
Step(1)In:Standard pedestrian data set of the structure with markup information, mainly including following three parts:
Pedestrian's picture image under 1-1, acquisition different scenes;Comprising a variety of pedestrian's postures, several scenes, a variety of illumination effects with
And in one day different periods video image so that the data set of construction is capable of the diversity of effective covering scene;
1-2, collected data set is effectively expanded, by image mirrors, angle rotation, size scaling, addition with
The method of machine noise;
1-3, the data images after expansion are generated with corresponding mark and label information, target in markup information, that is, sample
Location information, target generic in label information, that is, sample, classification behaviour are labeled as 1 or inhuman and are labeled as -1.
3. the rapid pedestrian detection method according to claim 1 based on deep learning, it is characterised in that:
Step(2)In:Utilize model of the pedestrian's data set training of structure based on Faster R-CNN:
By the labeled data collection of above-mentioned structure, pre-training model is finely adjusted using Faster R-CNN network structures, is instructed
Practice the model suitable for pedestrian detection;Faster R-CNN networks include the RPN convolutional neural networks for generating candidate region
With the fast R-CNN convolutional neural networks composition for target-recognition.
4. the rapid pedestrian detection method according to claim 1 based on deep learning, it is characterised in that:
Step(3)In:The test of the quick pedestrian detection of multi-channel video is carried out to trained model, mainly comprising following three
Point:
3-1, image to be tested is generated using multi-channel video frame joining method
By the way that the monitor video image of multichannel is spliced, a large-sized image is combined into, it then will be spliced
Image is input in trained model and is detected, and the video frame for being equivalent to multiple channels is carried out at the same time detection, can be effective
Save the detection time of video to be detected;
3-2, the influence that ambient enviroment is reduced using video frame preprocess method
In detection process, the main purpose of image preprocessing is the redundancy eliminated in image, filters out interference, noise, is increased
Detectability for information about by force, so as to improve the reliability of subsequent characteristics extraction and detection;
Available image processing method mainly includes following methods:Histogram equalization, normalization, Gamma corrections;
3-3, it will be input in trained model by pretreated splicing frame and carry out pedestrian detection, output pedestrian position letter
Breath and corresponding confidence level.
5. the rapid pedestrian detection method according to claim 3 based on deep learning, it is characterised in that:
Step(2)In:
It is finely adjusted on the data set of structure using existing ImageNet models, obtains new point suitable for pedestrian detection
Class model, key step are as follows:
2-1 carries out pre-training using the ZF-Net networks of ImageNet models, and the full articulamentum of the network model is carried out first
Modification, output class are two classes, i.e. pedestrian and non-pedestrian;
2-2 pre-training models on the sample set of structure, Faster R-CNN include the fine tuning RPN nets for extracting candidate region
Network and the Fast R-CNN networks for detecting target;RPN networks handle the image comprising pedestrian of input, generation
Pedestrian's roughing region;And Fast R-CNN networks further differentiate these pedestrian's roughing regions, output pedestrian is final
Location information;RPN networks and Fast RCNN networks can sharing feature, reduce the time of extraction candidate region, shorten
The time of target detection;
Using the method for pre-training primarily to a good netinit value is obtained, to avoid sunken in subsequent training
Enter local minimum, while also be able to accelerate the convergence rate of network;The follow-up sample set using structure is finely adjusted network,
To ensure that parameter is more suitable for current sample set, the verification and measurement ratio on data set is improved.
6. the rapid pedestrian detection method according to claim 4 based on deep learning, it is characterised in that:
Step(3)In:
It in 3-2, is corrected by quick Gamma, each color channel of image is standardized in rgb color space;
Gamma corrects the contrast for having adjusted image, reduces the influence caused by image local shade and illumination variation, alleviates field
The influence of scape and illumination to feature extraction.
7. the rapid pedestrian detection method according to claim 6 based on deep learning, it is characterised in that:
Step(3)In, in 3-2, Gamma correction methods, it is assumed that the pixel value of all images is the integer between 0 ~ M, to a picture
Element value performs following operate for the pixel of i:
3-2-1 is normalized:Pixel value is converted into the real number j between 0 ~ 1,;
3-2-2 is pre-compensated for:Pre-compensation value, the gamma values of wherein Gamma expression selections, in practical applications,
Gamma values should be adjusted in a certain range according to actual conditions, to obtain optimum efficiency;
3-2-3 renormalizations:The integer value between 0 ~ M will be changed to by the real number value contravariant of precompensation,;Aforesaid operations are performed to each pixel in image, i.e., can be obtained by Gamma
Image value after transformation;Simple way is by the corresponding Gamma changing values of all integer calculations between 0 ~ M, is corresponded to
Value be stored in the Gamma pre-established a correction look-up table, using the table to image of any pixel value between 0 ~ M into
Row Gamma is corrected.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810001077.9A CN108229390A (en) | 2018-01-02 | 2018-01-02 | Rapid pedestrian detection method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810001077.9A CN108229390A (en) | 2018-01-02 | 2018-01-02 | Rapid pedestrian detection method based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108229390A true CN108229390A (en) | 2018-06-29 |
Family
ID=62642574
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810001077.9A Pending CN108229390A (en) | 2018-01-02 | 2018-01-02 | Rapid pedestrian detection method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108229390A (en) |
Cited By (30)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109086712A (en) * | 2018-07-27 | 2018-12-25 | 深圳四海万联科技有限公司 | Working time statistical method, device, medium and the equipment of engineering machinery |
CN109409309A (en) * | 2018-11-05 | 2019-03-01 | 电子科技大学 | A kind of intelligent alarm system and method based on human testing |
CN109472229A (en) * | 2018-10-30 | 2019-03-15 | 福州大学 | Shaft tower Bird's Nest detection method based on deep learning |
CN109587093A (en) * | 2018-10-30 | 2019-04-05 | 东南大学 | A kind of deep learning network ofdm signal detection method based on cascade structure |
CN109635717A (en) * | 2018-12-10 | 2019-04-16 | 天津工业大学 | A kind of mining pedestrian detection method based on deep learning |
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 |
CN109816014A (en) * | 2019-01-22 | 2019-05-28 | 天津大学 | Generate method of the deep learning target detection network training with labeled data collection |
CN109886245A (en) * | 2019-03-02 | 2019-06-14 | 山东大学 | A kind of pedestrian detection recognition methods based on deep learning cascade neural network |
CN109902558A (en) * | 2019-01-15 | 2019-06-18 | 安徽理工大学 | A kind of human health deep learning prediction technique based on CNN-LSTM |
CN109919135A (en) * | 2019-03-27 | 2019-06-21 | 华瑞新智科技(北京)有限公司 | Behavioral value method, apparatus based on deep learning |
CN110147738A (en) * | 2019-04-29 | 2019-08-20 | 中国人民解放军海军特色医学中心 | A kind of driver fatigue monitoring and pre-alarming method and system |
CN110348376A (en) * | 2019-07-09 | 2019-10-18 | 华南理工大学 | A kind of pedestrian's real-time detection method neural network based |
CN110399690A (en) * | 2019-07-31 | 2019-11-01 | 佳都新太科技股份有限公司 | Subway station pedestrian simulation method, apparatus, electronic equipment and storage medium |
CN110503002A (en) * | 2019-07-26 | 2019-11-26 | 福州瑞芯微电子股份有限公司 | A kind of method for detecting human face and storage medium |
CN110781806A (en) * | 2019-10-23 | 2020-02-11 | 浙江工业大学 | Pedestrian detection tracking method based on YOLO |
CN110866487A (en) * | 2019-11-12 | 2020-03-06 | 复旦大学 | Large-scale pedestrian detection and re-identification sample set construction method and device |
CN110889367A (en) * | 2019-11-22 | 2020-03-17 | 贵州科学院(贵州省应用技术研究院) | Deep learning-based kitchen worker wearing standard identification method |
CN111222502A (en) * | 2019-12-28 | 2020-06-02 | 中国船舶重工集团公司第七一七研究所 | Infrared small target image labeling method and system |
CN111597992A (en) * | 2020-05-15 | 2020-08-28 | 哈尔滨工业大学 | Scene object abnormity identification method based on video monitoring |
CN111738103A (en) * | 2020-06-04 | 2020-10-02 | 山东大学 | Pedestrian tracking method and system with continuous autonomous learning capability |
CN112101305A (en) * | 2020-05-12 | 2020-12-18 | 杭州宇泛智能科技有限公司 | Multi-path image processing method and device and electronic equipment |
CN112183235A (en) * | 2020-09-07 | 2021-01-05 | 根尖体育科技(北京)有限公司 | Automatic control method for video acquisition aiming at sport places |
CN112712489A (en) * | 2020-12-31 | 2021-04-27 | 北京澎思科技有限公司 | Method, system and computer readable storage medium for image processing |
CN112733685A (en) * | 2020-12-31 | 2021-04-30 | 北京澎思科技有限公司 | Face tracking method, system and computer readable storage medium |
CN112884803A (en) * | 2020-08-18 | 2021-06-01 | 眸芯科技(上海)有限公司 | Real-time intelligent monitoring target detection method and device based on DSP |
CN113256499A (en) * | 2021-07-01 | 2021-08-13 | 北京世纪好未来教育科技有限公司 | Image splicing method, device and system |
CN113612922A (en) * | 2021-07-29 | 2021-11-05 | 重庆赛迪奇智人工智能科技有限公司 | Video processing method and device, electronic equipment and computer readable storage medium |
CN114299429A (en) * | 2021-12-24 | 2022-04-08 | 宁夏广天夏电子科技有限公司 | Human body recognition method, system and device based on deep learning |
CN114743257A (en) * | 2022-01-23 | 2022-07-12 | 中国电子科技集团公司第十研究所 | Method for detecting and identifying image target behaviors |
CN115359094A (en) * | 2022-09-05 | 2022-11-18 | 珠海安联锐视科技股份有限公司 | Moving target detection method based on deep learning |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101442653A (en) * | 2008-12-18 | 2009-05-27 | 北京航空航天大学 | Multichannel video play and large screen display method |
CN102256111A (en) * | 2011-07-17 | 2011-11-23 | 西安电子科技大学 | Multi-channel panoramic video real-time monitoring system and method |
CN103425967A (en) * | 2013-07-21 | 2013-12-04 | 浙江大学 | Pedestrian flow monitoring method based on pedestrian detection and tracking |
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 |
CN106203506A (en) * | 2016-07-11 | 2016-12-07 | 上海凌科智能科技有限公司 | A kind of pedestrian detection method based on degree of depth learning art |
US20170169313A1 (en) * | 2015-12-14 | 2017-06-15 | Samsung Electronics Co., Ltd. | Image processing apparatus and method based on deep learning and neural network learning |
-
2018
- 2018-01-02 CN CN201810001077.9A patent/CN108229390A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101442653A (en) * | 2008-12-18 | 2009-05-27 | 北京航空航天大学 | Multichannel video play and large screen display method |
CN102256111A (en) * | 2011-07-17 | 2011-11-23 | 西安电子科技大学 | Multi-channel panoramic video real-time monitoring system and method |
CN103425967A (en) * | 2013-07-21 | 2013-12-04 | 浙江大学 | Pedestrian flow monitoring method based on pedestrian detection and tracking |
US20170169313A1 (en) * | 2015-12-14 | 2017-06-15 | Samsung Electronics Co., Ltd. | Image processing apparatus and method based on deep learning and neural network learning |
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 |
CN106203506A (en) * | 2016-07-11 | 2016-12-07 | 上海凌科智能科技有限公司 | A kind of pedestrian detection method based on degree of depth learning art |
Non-Patent Citations (1)
Title |
---|
曾嘉亮: "Gamma校正的快速算法及其C语言实现", 《信息技术》 * |
Cited By (39)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109086712A (en) * | 2018-07-27 | 2018-12-25 | 深圳四海万联科技有限公司 | Working time statistical method, device, medium and the equipment of engineering machinery |
CN109472229A (en) * | 2018-10-30 | 2019-03-15 | 福州大学 | Shaft tower Bird's Nest detection method based on deep learning |
CN109587093A (en) * | 2018-10-30 | 2019-04-05 | 东南大学 | A kind of deep learning network ofdm signal detection method based on cascade structure |
CN109409309A (en) * | 2018-11-05 | 2019-03-01 | 电子科技大学 | A kind of intelligent alarm system and method based on human testing |
CN109635717A (en) * | 2018-12-10 | 2019-04-16 | 天津工业大学 | A kind of mining pedestrian detection method based on deep learning |
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 |
CN109766790B (en) * | 2018-12-24 | 2022-08-23 | 重庆邮电大学 | Pedestrian detection method based on self-adaptive characteristic channel |
CN109902558B (en) * | 2019-01-15 | 2022-12-20 | 安徽理工大学 | CNN-LSTM-based human health deep learning prediction method |
CN109902558A (en) * | 2019-01-15 | 2019-06-18 | 安徽理工大学 | A kind of human health deep learning prediction technique based on CNN-LSTM |
CN109816014A (en) * | 2019-01-22 | 2019-05-28 | 天津大学 | Generate method of the deep learning target detection network training with labeled data collection |
CN109886245A (en) * | 2019-03-02 | 2019-06-14 | 山东大学 | A kind of pedestrian detection recognition methods based on deep learning cascade neural network |
CN109919135A (en) * | 2019-03-27 | 2019-06-21 | 华瑞新智科技(北京)有限公司 | Behavioral value method, apparatus based on deep learning |
CN110147738A (en) * | 2019-04-29 | 2019-08-20 | 中国人民解放军海军特色医学中心 | A kind of driver fatigue monitoring and pre-alarming method and system |
CN110348376A (en) * | 2019-07-09 | 2019-10-18 | 华南理工大学 | A kind of pedestrian's real-time detection method neural network based |
CN110348376B (en) * | 2019-07-09 | 2021-05-14 | 华南理工大学 | Pedestrian real-time detection method based on neural network |
CN110503002A (en) * | 2019-07-26 | 2019-11-26 | 福州瑞芯微电子股份有限公司 | A kind of method for detecting human face and storage medium |
CN110503002B (en) * | 2019-07-26 | 2021-11-26 | 瑞芯微电子股份有限公司 | Face detection method and storage medium |
CN110399690A (en) * | 2019-07-31 | 2019-11-01 | 佳都新太科技股份有限公司 | Subway station pedestrian simulation method, apparatus, electronic equipment and storage medium |
CN110781806A (en) * | 2019-10-23 | 2020-02-11 | 浙江工业大学 | Pedestrian detection tracking method based on YOLO |
CN110866487B (en) * | 2019-11-12 | 2023-01-17 | 复旦大学 | Large-scale pedestrian detection and re-identification sample set construction method and device |
CN110866487A (en) * | 2019-11-12 | 2020-03-06 | 复旦大学 | Large-scale pedestrian detection and re-identification sample set construction method and device |
CN110889367A (en) * | 2019-11-22 | 2020-03-17 | 贵州科学院(贵州省应用技术研究院) | Deep learning-based kitchen worker wearing standard identification method |
CN111222502B (en) * | 2019-12-28 | 2023-05-12 | 中国船舶重工集团公司第七一七研究所 | Infrared small target image labeling method and system |
CN111222502A (en) * | 2019-12-28 | 2020-06-02 | 中国船舶重工集团公司第七一七研究所 | Infrared small target image labeling method and system |
CN112101305A (en) * | 2020-05-12 | 2020-12-18 | 杭州宇泛智能科技有限公司 | Multi-path image processing method and device and electronic equipment |
CN111597992A (en) * | 2020-05-15 | 2020-08-28 | 哈尔滨工业大学 | Scene object abnormity identification method based on video monitoring |
CN111597992B (en) * | 2020-05-15 | 2023-04-18 | 哈尔滨工业大学 | Scene object abnormity identification method based on video monitoring |
CN111738103A (en) * | 2020-06-04 | 2020-10-02 | 山东大学 | Pedestrian tracking method and system with continuous autonomous learning capability |
CN111738103B (en) * | 2020-06-04 | 2023-01-13 | 山东大学 | Pedestrian tracking method and system with continuous autonomous learning capability |
CN112884803A (en) * | 2020-08-18 | 2021-06-01 | 眸芯科技(上海)有限公司 | Real-time intelligent monitoring target detection method and device based on DSP |
CN112884803B (en) * | 2020-08-18 | 2023-11-24 | 眸芯科技(上海)有限公司 | Real-time intelligent monitoring target detection method and device based on DSP |
CN112183235A (en) * | 2020-09-07 | 2021-01-05 | 根尖体育科技(北京)有限公司 | Automatic control method for video acquisition aiming at sport places |
CN112733685A (en) * | 2020-12-31 | 2021-04-30 | 北京澎思科技有限公司 | Face tracking method, system and computer readable storage medium |
CN112712489A (en) * | 2020-12-31 | 2021-04-27 | 北京澎思科技有限公司 | Method, system and computer readable storage medium for image processing |
CN113256499A (en) * | 2021-07-01 | 2021-08-13 | 北京世纪好未来教育科技有限公司 | Image splicing method, device and system |
CN113612922A (en) * | 2021-07-29 | 2021-11-05 | 重庆赛迪奇智人工智能科技有限公司 | Video processing method and device, electronic equipment and computer readable storage medium |
CN114299429A (en) * | 2021-12-24 | 2022-04-08 | 宁夏广天夏电子科技有限公司 | Human body recognition method, system and device based on deep learning |
CN114743257A (en) * | 2022-01-23 | 2022-07-12 | 中国电子科技集团公司第十研究所 | Method for detecting and identifying image target behaviors |
CN115359094A (en) * | 2022-09-05 | 2022-11-18 | 珠海安联锐视科技股份有限公司 | Moving target detection method based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108229390A (en) | Rapid pedestrian detection method based on deep learning | |
CN105069746B (en) | Video real-time face replacement method and its system based on local affine invariant and color transfer technology | |
CN109284733B (en) | Shopping guide negative behavior monitoring method based on yolo and multitask convolutional neural network | |
CN104050471B (en) | Natural scene character detection method and system | |
CN103632136B (en) | Human-eye positioning method and device | |
CN100423020C (en) | Human face identifying method based on structural principal element analysis | |
CN105741375B (en) | A kind of infrared image Work attendance method of big field-of-view binocular vision | |
CN108446729A (en) | Egg embryo classification method based on convolutional neural networks | |
CN110059586B (en) | Iris positioning and segmenting system based on cavity residual error attention structure | |
CN108596102B (en) | RGB-D-based indoor scene object segmentation classifier construction method | |
CN107644418B (en) | Optic disk detection method and system based on convolutional neural networks | |
CN108647625A (en) | A kind of expression recognition method and device | |
CN107229929A (en) | A kind of license plate locating method based on R CNN | |
CN101828201A (en) | Image processing device and method, and learning device, method, and program | |
CN111832443B (en) | Construction method and application of construction violation detection model | |
CN101739546A (en) | Image cross reconstruction-based single-sample registered image face recognition method | |
CN102915446A (en) | Plant disease and pest detection method based on SVM (support vector machine) learning | |
CN112308156B (en) | Two-stage image change detection method based on counterstudy | |
CN109977887A (en) | A kind of face identification method of anti-age interference | |
CN111489330B (en) | Weak and small target detection method based on multi-source information fusion | |
CN106934380A (en) | A kind of indoor pedestrian detection and tracking based on HOG and MeanShift algorithms | |
CN109740459A (en) | A kind of image difference control methods, system and self-service device | |
CN108776777A (en) | The recognition methods of spatial relationship between a kind of remote sensing image object based on Faster RCNN | |
CN111815579A (en) | Image change detection method and device and computer readable storage medium | |
CN106909883A (en) | A kind of modularization hand region detection method and device based on ROS |
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 |
Application publication date: 20180629 |
|
RJ01 | Rejection of invention patent application after publication |