CN110321874A - A kind of light-weighted convolutional neural networks pedestrian recognition method - Google Patents
A kind of light-weighted convolutional neural networks pedestrian recognition method Download PDFInfo
- Publication number
- CN110321874A CN110321874A CN201910629693.3A CN201910629693A CN110321874A CN 110321874 A CN110321874 A CN 110321874A CN 201910629693 A CN201910629693 A CN 201910629693A CN 110321874 A CN110321874 A CN 110321874A
- Authority
- CN
- China
- Prior art keywords
- filter
- length
- pedestrian
- filled
- convolutional layer
- 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 43
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 36
- 238000001514 detection method Methods 0.000 claims abstract description 35
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims abstract description 5
- 238000012795 verification Methods 0.000 claims abstract description 4
- 238000005070 sampling Methods 0.000 claims description 19
- 230000000694 effects Effects 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 5
- 238000013461 design Methods 0.000 claims description 4
- 238000007689 inspection Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 2
- 238000013480 data collection Methods 0.000 claims 1
- 238000001914 filtration Methods 0.000 claims 1
- 238000003475 lamination Methods 0.000 claims 1
- 238000012216 screening Methods 0.000 claims 1
- 238000012360 testing method Methods 0.000 abstract description 3
- 238000005516 engineering process Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 108010023321 Factor VII Proteins 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 230000029052 metamorphosis Effects 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000032258 transport Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (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)
Abstract
The present invention discloses a kind of light-weighted convolutional neural networks pedestrian recognition method, and step is: obtaining original pedestrian's data, obtains the pedestrian image with mark information, constructs pedestrian's data set;The input size for designing pedestrian dummy network, clusters data set, selects suitable candidate frame;Image is pre-processed, and expanding data;Convolutional neural networks are constructed, and pretreated image feeding network is trained, obtain the network model with pedestrian's identification function;Image with mark information is sent into network in batches, is trained;It checks the penalty values and accuracy rate on training set, is trained again after regularized learning algorithm rate or increase the number of iterations if not ideal enough, if result ideal tests trained network on verifying collection, network is adjusted according to verification result again.Such recognition methods can improve accuracy of identification, run network model on the hardware platform of lower configuration while ensuring target detection model real-time.
Description
Technical field
The invention belongs to mode identification technology, in particular to a kind of light-weighted convolutional neural networks pedestrian identification side
Method.
Background technique
In recent years, with the rise with AI industry, the continuous development of the industries such as communications and transportation, pedestrian detection technology
Extensive concern has been obtained in fields such as intelligent monitoring, intelligent drivings.In traditional pedestrian detection method, pass through engineer's
Feature extractor, such as Haar, HOG, LBP etc. extract pedestrian's feature from training sample, then are trained with the pedestrian's feature extracted
The classifiers such as SVM carry out pedestrian detection task.Such as using HOG+LBP characteristic processing pedestrian's occlusion issue, Lai Tigao pedestrian
The accuracy rate of detection.With ICF, ACF, gradient magnitude feature, LUV color characteristic etc. proposes that pedestrian's feature obtains preferably
Expression, also achieves preferable achievement in pedestrian's Detection task.But by pedestrian's feature of manual designs have generalization it is poor,
The disadvantages of being difficult to adapt to pedestrian's metamorphosis, it is difficult in practical applications.And in recent years, with the development of deep learning,
Especially extensive application of the convolutional neural networks in computer vision field, in conjunction with the pedestrian recognition method of convolutional neural networks
Many concerns and application are obtained.
It is suitable that " pedestrian detection method under special scenes " of Patent No. 201611146030.9 is selected according to test set
Training set;Supervision message required for detection framework is determined in training set subscript, is completed in training set with confrontation network
Pedestrian's mark;VOC form is converted by training set be delivered to R FCN be trained;Finally by the training pattern of R FCN to specific
Scene pedestrian detects, and the method real-time is insufficient, higher to hsrdware requirements, is difficult to realize in embedded device.
A kind of " real-time railway mouth multi-cam pedestrian inspection based on deep learning of Patent No. 201810457285.X
Survey method " video that is acquired by multiple cameras, camera is overlapped emphasis monitoring area and monitors;Selection needs that detects to take the photograph
As head region and obtain the camera video stream under the region;Video image processing is carried out in GPU processor cluster: first will
Image cutting, then detects the image after cutting with trained pedestrian detection model;It monitors in real time in GPU processor
GPU service condition and be scheduled according to scheduling strategy;Pedestrian area coordinate is calculated, current time stamp is obtained and is sent to use
Family client.A kind of real-time railway mouth multi-cam pedestrian detection method based on deep learning, by video processing and depth
Study combines, and uses the method for camera overlapping monitoring to improve the accuracy rate of railway mouth pedestrian detection.The method needs
GPU processor group is built, deployment is got up more difficult.
" one kind knows method for distinguishing for pedestrian " of Patent No. 201810103970.2 utilizes video processing and depth
Habit technology, training convolutional neural networks obtain pedestrian's disaggregated model on personal data of being expert at collection, and to be identified by the model realization
The feature extraction of pedestrian, obtains candidate row list finally by characteristic measure and method for reordering in pedestrian and database, real
The identification of existing people.The method is utilized residual error network, therefore there are network parameter amount is big, higher to storage capacity requirement equal to lack
Point.
Paper " a kind of improved convolutional neural networks pedestrian recognition method " proposes a kind of real-time pedestrian detection method.With
The input size of YOLOv3-tiny network foundation change network model;Using clustering method, target frame is carried out to data set
Cluster chooses the candidate frame size and quantity for being suitble to pedestrian detection;Method by increasing certain amount convolutional layer redesigns special
Sign is extracted and target detection network;Finally training obtains detection model on mixed data set.Though the method improves pedestrian's inspection
The precision of survey, but network query function amount is still larger, is disposed in mobile terminal more difficult.
Although mostly existing in conclusion having more research to the pedestrian detection based on convolutional neural networks both at home and abroad
Network parameter is excessive, the more complex problem of network structure, have higher requirements to hardware device and running memory.
Summary of the invention
The purpose of the present invention is to provide a kind of light-weighted convolutional neural networks pedestrian recognition method, can ensure
While target detection model real-time, accuracy of identification is improved, transports network model on the hardware platform of lower configuration
Row.
In order to achieve the above objectives, solution of the invention is:
A kind of light-weighted convolutional neural networks pedestrian recognition method, includes the following steps:
Step 1, original pedestrian's data are obtained, the pedestrian image with mark information is obtained, construct pedestrian's data set;
Step 2, the input size of pedestrian dummy network is designed, and data set is clustered according to the size of design, is selected
Select suitable candidate frame;
Step 3, image is pre-processed, and passes through data enhancement methods expanding data;
Step 4, construct convolutional neural networks, and by by step 3 pretreatment after and expand the figure with mark information
As being sent into network in batches, it is trained, obtains the network model with pedestrian's identification function;
Step 5, after the completion of network training, the penalty values and accuracy rate on training set is checked, are adjusted if not ideal enough
It is trained again after learning rate or increase the number of iterations, completes to train if result ideal, go to step 6;
Step 6, trained network is tested on verifying collection, network is adjusted according to verification result again.
In above-mentioned steps 1, the method for building pedestrian's data set is: to INRIA data set, VOC data set, MS COCO number
It is screened according to collection, obtains the pedestrian image of all ages and classes, different sexes, different shape, and the coordinate by pedestrian in the picture
As mark information, pedestrian's data set is constructed.
In above-mentioned steps 2, the pedestrian image concentrated to data is labeled, and pedestrian is labeled as 0, the seat of recording mark frame
Mark.
In above-mentioned steps 3, data enhancement methods, which are specifically included, to carry out random exposure to image, adjustment image saturation, adjusts
Three kinds of whole picture tone operations.
In above-mentioned steps 4, the convolutional neural networks of building include convolutional layer, pond layer and detection layers, in which:
Convolutional layer structure: the first convolutional layer uses the filter of 16 3 × 3 sizes, and step-length 1 is filled with 1;Second is deep
The filter that separable convolutional layer uses 16 3 × 3 sizes is spent, step-length 1 is filled with 1;Third convolutional layer uses 32 1 × 1
The filter of size, step-length 1, is filled with 0;The size of first pond layer filter is 3 × 3, and step-length 2 is used to down adopt
Sample;4th depth separates the filter that convolutional layer uses 32 3 × 3 sizes, and step-length 1 is filled with 1;5th convolutional layer is adopted
With the filter of 64 1 × 1 sizes, step-length 1 is filled with 0;The size of second pond layer filter is 2 × 2, and step-length is
2, it is used for down-sampling;6th depth separates convolutional layer and uses 64 as the filter of 3 × 3 sizes, and step-length 1 is filled with 1;
7th convolutional layer uses the filter of 128 1 × 1 sizes, and step-length 1 is filled with 0;The size of third down-sampling layer filter
It is 2 × 2, step-length 2 is used for down-sampling;8th depth separates convolutional layer and uses 128 as the filter of 3 × 3 sizes, walks
A length of 1, it is filled with 1;9th convolutional layer uses 256 as the filter of 1 × 1 size, and step-length 1 is filled with 0;4th pond
The size of layer filter is 2 × 2, and step-length 2 is used for down-sampling;The separable convolutional layer of tenth depth uses 256 big for 3 × 3
Small filter, step-length 1, is filled with 1;11st convolutional layer uses 512 as the filter of 1 × 1 size, and step-length 1 is filled out
Fill is 0;The size of 5th pond layer filter is 2 × 2, and step-length 2 is used for down-sampling;12nd depth separates convolutional layer
512 are used as the filter of 3 × 3 sizes, step-length 1 is filled with 1;13rd convolutional layer use 1024 for 1 × 1 size
Filter, step-length 1 is filled with 0;The size of 6th pond layer filter is 2 × 2, step-length 1;14th depth can divide
From the filter that convolutional layer uses 1024 3 × 3 sizes, it is filled with 1;
Detection layers: it is divided into two scales, scale 1: adds three convolutional layers after basic network and predicted, addition
In convolutional layer, the 15th convolutional layer uses the filter of 256 3 × 3 sizes, and step-length 1 is filled with 1;16th convolutional layer is adopted
With the filter of 512 3 × 3 sizes, step-length 1 is filled with 1;17th convolutional layer uses the filter of 18 3 × 3 sizes,
Step-length is 1, is filled with 1;
Scale 2: 2 times of up-samplings are carried out from the 16th convolutional layer, the feature for being 38 × 18 with the output of the 11st convolutional layer
Figure is added, then is predicted after adding three convolution, and the 18th convolutional layer uses the filter of 128 3 × 3 sizes, and step-length is
1, it is filled with 1;19th convolutional layer uses the filter of 256 3 × 3 sizes, and step-length 1 is filled with 1;20th convolutional layer
Using the filter of 18 3 × 3 sizes, step-length 1 is filled with 1.
In above-mentioned detection layers, the characteristic pattern of scale 1 exports the classification and location information of target having a size of 19 × 9;Scale
Two characteristic pattern exports the classification and location information of target having a size of 38 × 18.
In above-mentioned steps 5, in the training process, the penalty values and accuracy rate on training set are calculated, using under stochastic gradient
It drops algorithm and carries out backpropagation, realize that parameter updates.
In above-mentioned steps 5, the penalty values and accuracy rate on training set are checked, it is expected that loss value is not higher than 20%, it is expected that standard
True rate is not less than 90%, and recall rate is not less than 85%, and omission factor is not higher than 15%.
The particular content of above-mentioned steps 6 is: if effect is verified very well and shows bad on collection on training set, reducing
The number of iterations changes decay value, again with training set training network adjusted, return step 5.
After adopting the above scheme, compared with prior art, the present invention having the advantage that
(1) based on Tiny-YOLO, change the Aspect Ratio of network model input, extracted using network more horizontal
To feature, selects the rectangular image of 608 × 288 resolution ratio as network inputs, increase the extraction effect of transverse features, for row
The target frame that people's detection data is concentrated carries out K-means clustering, facilitates the ability for the pedestrian's identification for promoting network.
2, construction INRIA+VOC+MS COCO mixes pedestrian detection data set, the training on mixing pedestrian detection data set
Network model keeps the generalization of network model stronger, and detection effect is more preferable.
3,3 × 3 common convolutional layers are replaced with into the separable convolutional layer of 3 × 3 depth to reduce network model calculation amount, together
Shi Jiashen network.Due to 3 × 3 depth separate convolutional layer addition, by calculation amount be reduced to common convolution model three/
One, reduce pedestrian detection model running to memory capacity and running memory requirement, guarantees while increasing pedestrian's recognition capability
Model real-time.
The invention proposes a kind of light-weighted convolutional neural networks pedestrian recognition method technologies, using Tiny-YOLO as base
Plinth changes the input size of network model, in conjunction with pedestrian's size characteristics in image, using clustering method, to data set into
Row target frame cluster, chooses the candidate frame size and quantity for being suitble to pedestrian detection;Common convolutional layer is replaced with depth to separate
Convolutional layer redesigns feature extraction and target detection network;The training on mixed data set, enhances model generalization.
Detailed description of the invention
Fig. 1 is flow chart of the invention;
Fig. 2 is inventive network model structure, includes network layer and every layer of output;
Wherein, Input: input picture;Conv: common convolutional layer;Depthwise Conv: depth separates convolutional layer;
Ped-Detection: detection layers.
Specific embodiment
Below with reference to attached drawing, technical solution of the present invention and beneficial effect are described in detail.
As shown in Figure 1, the present invention provides a kind of light-weighted convolutional neural networks pedestrian recognition method, including walk as follows
It is rapid:
Step 1, original pedestrian's data are obtained, INRIA data set, VOC data set, MS COCO data set are sieved
Choosing obtains the pedestrian image of all ages and classes, different sexes, different shape, and believes the coordinate of pedestrian in the picture as label
Breath constructs pedestrian's data set;
Step 2, the input size of pedestrian dummy network is designed, and data set is clustered according to the size of design, is selected
Select suitable candidate frame;In the present embodiment, facilitate pedestrian's identification due to increasing transverse features expression, change pedestrian's mould
The Aspect Ratio of type network inputs extracts more transverse features using rectangle input network, while in order to avoid input picture
Influence of the resolution ratio to networked-induced delay, selects 608 × 288 image in different resolution as network inputs;The pedestrian that data are concentrated
Image is labeled, and pedestrian is labeled as 0, the coordinate of recording mark frame;
Step 3, image is pre-processed, and passes through data enhancement methods expanding data;Wherein, the data enhancing side
Method, which is specifically included, carries out three kinds of picture tone random exposure, adjustment image saturation, adjustment operations to image, is enhanced by data
To the supplement of sample data, so that the network finally trained has better robust for the pedestrian detection under varying environment
Property, it also can effectively prevent the over-fitting of network;
Step 4, convolutional neural networks are constructed, and pretreated image feeding network is trained, obtain that there is row
The network model of people's identification function;
As shown in Fig. 2, the convolutional neural networks of building include convolutional layer, pond layer and detection layers, in which:
Convolutional layer structure: the first convolutional layer uses the filter of 16 3 × 3 sizes, and step-length 1 is filled with 1;Second is deep
The filter that separable convolutional layer uses 16 3 × 3 sizes is spent, step-length 1 is filled with 1;Third convolutional layer uses 32 1 × 1
The filter of size, step-length 1, is filled with 0;The size of first pond layer filter is 3 × 3, and step-length 2 is used to down adopt
Sample;4th depth separates the filter that convolutional layer uses 32 3 × 3 sizes, and step-length 1 is filled with 1;5th convolutional layer is adopted
With the filter of 64 1 × 1 sizes, step-length 1 is filled with 0;The size of second pond layer filter is 2 × 2, and step-length is
2, it is used for down-sampling;6th depth separates convolutional layer and uses 64 as the filter of 3 × 3 sizes, and step-length 1 is filled with 1;
7th convolutional layer uses the filter of 128 1 × 1 sizes, and step-length 1 is filled with 0;The size of third down-sampling layer filter
It is 2 × 2, step-length 2 is used for down-sampling;8th depth separates convolutional layer and uses 128 as the filter of 3 × 3 sizes, walks
A length of 1, it is filled with 1;9th convolutional layer uses 256 as the filter of 1 × 1 size, and step-length 1 is filled with 0;4th pond
The size of layer filter is 2 × 2, and step-length 2 is used for down-sampling;The separable convolutional layer of tenth depth uses 256 big for 3 × 3
Small filter, step-length 1, is filled with 1;11st convolutional layer uses 512 as the filter of 1 × 1 size, and step-length 1 is filled out
Fill is 0;The size of 5th pond layer filter is 2 × 2, and step-length 2 is used for down-sampling;12nd depth separates convolutional layer
512 are used as the filter of 3 × 3 sizes, step-length 1 is filled with 1;13rd convolutional layer use 1024 for 1 × 1 size
Filter, step-length 1 is filled with 0;The size of 6th pond layer filter is 2 × 2, step-length 1;14th depth can divide
From the filter that convolutional layer uses 1024 3 × 3 sizes, it is filled with 1.
Detection layers: it is divided into two scales, scale 1: adds three convolutional layers after basic network and predicted.Addition
In convolutional layer, the 15th convolutional layer uses the filter of 256 3 × 3 sizes, and step-length 1 is filled with 1;16th convolutional layer is adopted
With the filter of 512 3 × 3 sizes, step-length 1 is filled with 1;17th convolutional layer uses the filter of 18 3 × 3 sizes,
Step-length is 1, is filled with 1.The characteristic pattern of scale one exports the classification and location information of target having a size of 19 × 9, to larger ruler
Very little pedestrian has preferable recognition capability;
Scale 2: 2 times of up-samplings are carried out from the 16th convolutional layer, the feature for being 38 × 18 with the output of the 11st convolutional layer
Figure is added, then is predicted after adding three convolution, and the 18th convolutional layer uses the filter of 128 3 × 3 sizes, and step-length is
1, it is filled with 1;19th convolutional layer uses the filter of 256 3 × 3 sizes, and step-length 1 is filled with 1;20th convolutional layer
Using the filter of 18 3 × 3 sizes, step-length 1 is filled with 1.The characteristic pattern of scale two exports target having a size of 38 × 18
Classification and location information, have preferable recognition capability to the pedestrian of smaller size.
In step 4, the number of iterations and Study rate parameter appropriate are set, wherein the number of iterations should be ensured that in iteration time
Network can restrain in number, i.e. accuracy rate and penalty values of the network on training set tends towards stability.
Step 5, the image of tape label information is sent into network in batches, is trained.The wherein amount of images that every batch of includes,
It is specifically formulated in conjunction with hardware facility;Used here as GPU model GTX Titan Xp, video card memory 12g, every batch of is sent into 64;
In training process, the penalty values and accuracy rate on training set are calculated, backpropagation is carried out using stochastic gradient descent algorithm, is realized
Parameter updates;
Step 6, after the completion of network training, the penalty values and accuracy rate on training set are checked, it is expected that loss value is not higher than
20%, it is expected that accuracy rate is not less than 90%, recall rate is not less than 85%, and omission factor is not higher than 15%.If not ideal enough pass through
Regularized learning algorithm rate or increase the number of iterations are trained again, if result ideal completes training.
Step 7, trained network is tested on verifying collection, network is adjusted according to verification result again.If
Effect shows bad very well and on verifying collection on training set, can reduce the number of iterations or attempt to change decay value, use again
Training set training network adjusted, that is, return to step 5;
Step 8, trained network model is subjected to batch testing on verifying collection, final network model loss is
15%, mAP reach 92%, and recall rate reaches 93%, omission factor 7%, to the row of various sizes of pedestrian and partial occlusion
People has better detection effect.
The above examples only illustrate the technical idea of the present invention, and this does not limit the scope of protection of the present invention, all
According to the technical idea provided by the invention, any changes made on the basis of the technical scheme each falls within the scope of the present invention
Within.
Claims (9)
1. a kind of light-weighted convolutional neural networks pedestrian recognition method, it is characterised in that include the following steps:
Step 1, original pedestrian's data are obtained, the pedestrian image with mark information is obtained, construct pedestrian's data set;
Step 2, the input size of pedestrian dummy network is designed, and pedestrian's data set is clustered according to the size of design, is selected
Select suitable candidate frame;
Step 3, image is pre-processed, and passes through data enhancement methods expanding data;
Step 4, construct convolutional neural networks, and by by step 3 pretreatment after and expand the image with mark information divide
It criticizes and is sent into network, be trained, obtain the network model with pedestrian's identification function;
Step 5, after the completion of network training, the penalty values and accuracy rate on training set, the regularized learning algorithm if not ideal enough are checked
It is trained again after rate or increase the number of iterations, completes to train if result ideal, go to step 6;
Step 6, trained network is tested on verifying collection, network is adjusted according to verification result again.
2. a kind of light-weighted convolutional neural networks pedestrian recognition method as described in claim 1, it is characterised in that: the step
In rapid 1, the method for building pedestrian's data set is: screening, obtains to INRIA data set, VOC data set, MS COCO data set
To the pedestrian image of all ages and classes, different sexes, different shape, and using the coordinate of pedestrian in the picture as mark information, structure
Construction Bank's personal data collection.
3. a kind of light-weighted convolutional neural networks pedestrian recognition method as described in claim 1, it is characterised in that: the step
In rapid 2, the pedestrian image concentrated to data is labeled, and pedestrian is labeled as 0, the coordinate of recording mark frame.
4. a kind of light-weighted convolutional neural networks pedestrian recognition method as described in claim 1, it is characterised in that: the step
In rapid 3, data enhancement methods, which are specifically included, carries out random exposure, adjustment image saturation, three kinds of picture tone of adjustment to image
Operation.
5. a kind of light-weighted convolutional neural networks pedestrian recognition method as described in claim 1, it is characterised in that: the step
In rapid 4, the convolutional neural networks of building include convolutional layer, pond layer and detection layers, in which:
Convolutional layer structure: the first convolutional layer uses the filter of 16 3 × 3 sizes, and step-length 1 is filled with 1;Second depth can
The filter that convolutional layer uses 16 3 × 3 sizes is separated, step-length 1 is filled with 1;Third convolutional layer uses 32 1 × 1 sizes
Filter, step-length 1 is filled with 0;The size of first pond layer filter is 3 × 3, and step-length 2 is used for down-sampling;The
Four depth separate the filter that convolutional layer uses 32 3 × 3 sizes, and step-length 1 is filled with 1;5th convolutional layer uses 64
The filter of 1 × 1 size, step-length 1, is filled with 0;The size of second pond layer filter is 2 × 2, and step-length 2 is used for
Down-sampling;6th depth separates convolutional layer and uses 64 as the filter of 3 × 3 sizes, and step-length 1 is filled with 1;Volume seven
Lamination uses the filter of 128 1 × 1 sizes, and step-length 1 is filled with 0;The size of third down-sampling layer filter is 2 × 2,
Step-length is 2, is used for down-sampling;8th depth separates convolutional layer and uses 128 as the filter of 3 × 3 sizes, and step-length 1 is filled out
Fill is 1;9th convolutional layer uses 256 as the filter of 1 × 1 size, and step-length 1 is filled with 0;4th pond layer filter
Size be 2 × 2, step-length 2, be used for down-sampling;Tenth depth separate convolutional layer use 256 for the filtering of 3 × 3 sizes
Device, step-length 1, is filled with 1;11st convolutional layer uses 512 as the filter of 1 × 1 size, and step-length 1 is filled with 0;The
The size of five pond layer filter is 2 × 2, and step-length 2 is used for down-sampling;12nd depth separates convolutional layer and uses 512
For the filter of 3 × 3 sizes, step-length 1 is filled with 1;13rd convolutional layer use 1024 for the filter of 1 × 1 size,
Step-length is 1, is filled with 0;The size of 6th pond layer filter is 2 × 2, step-length 1;14th depth separates convolutional layer and adopts
With the filter of 1024 3 × 3 sizes, it is filled with 1;
Detection layers: it is divided into two scales, scale 1: adds three convolutional layers after basic network and predicted, the convolution of addition
In layer, the 15th convolutional layer uses the filter of 256 3 × 3 sizes, and step-length 1 is filled with 1;16th convolutional layer uses
The filter of 512 3 × 3 sizes, step-length 1, is filled with 1;17th convolutional layer uses the filter of 18 3 × 3 sizes, step
A length of 1, it is filled with 1;
Scale 2: 2 times of up-samplings are carried out from the 16th convolutional layer, the characteristic pattern phase for being 38 × 18 with the output of the 11st convolutional layer
Add, then predicted after adding three convolution, the 18th convolutional layer uses the filter of 128 3 × 3 sizes, and step-length 1 is filled out
Fill is 1;19th convolutional layer uses the filter of 256 3 × 3 sizes, and step-length 1 is filled with 1;20th convolutional layer uses
The filter of 18 3 × 3 sizes, step-length 1, is filled with 1.
6. a kind of light-weighted convolutional neural networks pedestrian recognition method as claimed in claim 5, it is characterised in that: the inspection
It surveys in layer, the characteristic pattern of scale 1 exports the classification and location information of target having a size of 19 × 9;The characteristic pattern size of scale two
It is 38 × 18, exports the classification and location information of target.
7. a kind of light-weighted convolutional neural networks pedestrian recognition method as described in claim 1, it is characterised in that: the step
In rapid 5, in the training process, the penalty values and accuracy rate on training set are calculated, is carried out using stochastic gradient descent algorithm reversed
It propagates, realizes that parameter updates.
8. a kind of light-weighted convolutional neural networks pedestrian recognition method as described in claim 1, it is characterised in that: the step
In rapid 5, the penalty values and accuracy rate on training set are checked, it is expected that loss value is not higher than 20%, it is expected that accuracy rate is not less than 90%,
Recall rate is not less than 85%, and omission factor is not higher than 15%.
9. a kind of light-weighted convolutional neural networks pedestrian recognition method as described in claim 1, it is characterised in that: the step
Rapid 6 particular content is: if effect is verified very well and shows bad on collection on training set, reducing the number of iterations or change
Decay value, again with training set training network adjusted, return step 5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910629693.3A CN110321874A (en) | 2019-07-12 | 2019-07-12 | A kind of light-weighted convolutional neural networks pedestrian recognition method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910629693.3A CN110321874A (en) | 2019-07-12 | 2019-07-12 | A kind of light-weighted convolutional neural networks pedestrian recognition method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110321874A true CN110321874A (en) | 2019-10-11 |
Family
ID=68122106
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910629693.3A Pending CN110321874A (en) | 2019-07-12 | 2019-07-12 | A kind of light-weighted convolutional neural networks pedestrian recognition method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110321874A (en) |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110942144A (en) * | 2019-12-05 | 2020-03-31 | 深圳牛图科技有限公司 | Neural network construction method integrating automatic training, checking and reconstructing |
CN111126310A (en) * | 2019-12-26 | 2020-05-08 | 华侨大学 | Pedestrian gender identification method based on scene migration |
CN111144272A (en) * | 2019-12-24 | 2020-05-12 | 珠海格力电器股份有限公司 | Multi-scale gait recognition method and system and intelligent household appliance |
CN111401241A (en) * | 2020-03-16 | 2020-07-10 | 中国民用航空飞行学院 | Marine civil aircraft searching method and device based on convolutional neural network |
CN111401215A (en) * | 2020-03-12 | 2020-07-10 | 杭州涂鸦信息技术有限公司 | Method and system for detecting multi-class targets |
CN111507271A (en) * | 2020-04-20 | 2020-08-07 | 北京理工大学 | Airborne photoelectric video target intelligent detection and identification method |
CN111582091A (en) * | 2020-04-27 | 2020-08-25 | 西安交通大学 | Pedestrian identification method based on multi-branch convolutional neural network |
CN111898699A (en) * | 2020-08-11 | 2020-11-06 | 海之韵(苏州)科技有限公司 | Automatic detection and identification method for hull target |
CN111931615A (en) * | 2020-07-28 | 2020-11-13 | 五邑大学 | Robot target identification method, system, device and storage medium |
CN112365740A (en) * | 2020-11-30 | 2021-02-12 | 北京停简单信息技术有限公司 | Alarm display method and device |
CN112487911A (en) * | 2020-11-24 | 2021-03-12 | 中国信息通信科技集团有限公司 | Real-time pedestrian detection method and device based on improved yolov3 in intelligent monitoring environment |
CN112784756A (en) * | 2021-01-25 | 2021-05-11 | 南京邮电大学 | Human body identification tracking method |
CN112836619A (en) * | 2021-01-28 | 2021-05-25 | 合肥英睿系统技术有限公司 | Embedded vehicle-mounted far infrared pedestrian detection method, system, equipment and storage medium |
CN112836657A (en) * | 2021-02-08 | 2021-05-25 | 中国电子科技集团公司第三十八研究所 | Pedestrian detection method and system based on lightweight YOLOv3 |
CN113343949A (en) * | 2021-08-03 | 2021-09-03 | 中国航空油料集团有限公司 | Pedestrian detection model training method for universal embedded platform |
CN113408423A (en) * | 2021-06-21 | 2021-09-17 | 西安工业大学 | Aquatic product target real-time detection method suitable for TX2 embedded platform |
CN113468992A (en) * | 2021-06-21 | 2021-10-01 | 四川轻化工大学 | Construction site safety helmet wearing detection method based on lightweight convolutional neural network |
CN113723526A (en) * | 2021-08-31 | 2021-11-30 | 电子科技大学 | Method for identifying different types of craters |
CN114623727A (en) * | 2022-03-14 | 2022-06-14 | 北京理工大学 | Laser imaging short-range detection target identification method |
CN112487911B (en) * | 2020-11-24 | 2024-05-24 | 中国信息通信科技集团有限公司 | Real-time pedestrian detection method and device based on improvement yolov under intelligent monitoring environment |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107644221A (en) * | 2017-10-31 | 2018-01-30 | 南京航空航天大学 | Convolutional neural networks traffic sign recognition method based on compression of parameters |
CN108288047A (en) * | 2018-02-07 | 2018-07-17 | 成都新舟锐视科技有限公司 | A kind of pedestrian/vehicle checking method |
CN109919045A (en) * | 2019-02-18 | 2019-06-21 | 北京联合大学 | Small scale pedestrian detection recognition methods based on concatenated convolutional network |
-
2019
- 2019-07-12 CN CN201910629693.3A patent/CN110321874A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107644221A (en) * | 2017-10-31 | 2018-01-30 | 南京航空航天大学 | Convolutional neural networks traffic sign recognition method based on compression of parameters |
CN108288047A (en) * | 2018-02-07 | 2018-07-17 | 成都新舟锐视科技有限公司 | A kind of pedestrian/vehicle checking method |
CN109919045A (en) * | 2019-02-18 | 2019-06-21 | 北京联合大学 | Small scale pedestrian detection recognition methods based on concatenated convolutional network |
Non-Patent Citations (1)
Title |
---|
陈聪等: "一种改进的卷积神经网络行人识别方法", 《HTTP://KNS.CNKI.NET/KCMS/DETAIL/23.1191.U.20181017.1655.002.HTM1》 * |
Cited By (33)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110942144B (en) * | 2019-12-05 | 2023-05-02 | 深圳牛图科技有限公司 | Neural network construction method integrating automatic training, checking and reconstruction |
CN110942144A (en) * | 2019-12-05 | 2020-03-31 | 深圳牛图科技有限公司 | Neural network construction method integrating automatic training, checking and reconstructing |
CN111144272A (en) * | 2019-12-24 | 2020-05-12 | 珠海格力电器股份有限公司 | Multi-scale gait recognition method and system and intelligent household appliance |
CN111126310B (en) * | 2019-12-26 | 2023-03-24 | 华侨大学 | Pedestrian gender identification method based on scene migration |
CN111126310A (en) * | 2019-12-26 | 2020-05-08 | 华侨大学 | Pedestrian gender identification method based on scene migration |
CN111401215A (en) * | 2020-03-12 | 2020-07-10 | 杭州涂鸦信息技术有限公司 | Method and system for detecting multi-class targets |
CN111401215B (en) * | 2020-03-12 | 2023-10-31 | 杭州涂鸦信息技术有限公司 | Multi-class target detection method and system |
CN111401241A (en) * | 2020-03-16 | 2020-07-10 | 中国民用航空飞行学院 | Marine civil aircraft searching method and device based on convolutional neural network |
CN111507271B (en) * | 2020-04-20 | 2021-01-12 | 北京理工大学 | Airborne photoelectric video target intelligent detection and identification method |
CN111507271A (en) * | 2020-04-20 | 2020-08-07 | 北京理工大学 | Airborne photoelectric video target intelligent detection and identification method |
CN111582091A (en) * | 2020-04-27 | 2020-08-25 | 西安交通大学 | Pedestrian identification method based on multi-branch convolutional neural network |
CN111582091B (en) * | 2020-04-27 | 2023-05-02 | 西安交通大学 | Pedestrian recognition method based on multi-branch convolutional neural network |
CN111931615A (en) * | 2020-07-28 | 2020-11-13 | 五邑大学 | Robot target identification method, system, device and storage medium |
CN111931615B (en) * | 2020-07-28 | 2024-01-09 | 五邑大学 | Robot target recognition method, system, device and storage medium |
CN111898699B (en) * | 2020-08-11 | 2024-05-10 | 海之韵(苏州)科技有限公司 | Automatic detection and identification method for ship body target |
CN111898699A (en) * | 2020-08-11 | 2020-11-06 | 海之韵(苏州)科技有限公司 | Automatic detection and identification method for hull target |
CN112487911B (en) * | 2020-11-24 | 2024-05-24 | 中国信息通信科技集团有限公司 | Real-time pedestrian detection method and device based on improvement yolov under intelligent monitoring environment |
CN112487911A (en) * | 2020-11-24 | 2021-03-12 | 中国信息通信科技集团有限公司 | Real-time pedestrian detection method and device based on improved yolov3 in intelligent monitoring environment |
CN112365740A (en) * | 2020-11-30 | 2021-02-12 | 北京停简单信息技术有限公司 | Alarm display method and device |
CN112784756A (en) * | 2021-01-25 | 2021-05-11 | 南京邮电大学 | Human body identification tracking method |
CN112784756B (en) * | 2021-01-25 | 2022-08-26 | 南京邮电大学 | Human body identification tracking method |
CN112836619A (en) * | 2021-01-28 | 2021-05-25 | 合肥英睿系统技术有限公司 | Embedded vehicle-mounted far infrared pedestrian detection method, system, equipment and storage medium |
CN112836657B (en) * | 2021-02-08 | 2023-04-18 | 中国电子科技集团公司第三十八研究所 | Pedestrian detection method and system based on lightweight YOLOv3 |
CN112836657A (en) * | 2021-02-08 | 2021-05-25 | 中国电子科技集团公司第三十八研究所 | Pedestrian detection method and system based on lightweight YOLOv3 |
CN113468992B (en) * | 2021-06-21 | 2022-11-04 | 四川轻化工大学 | Construction site safety helmet wearing detection method based on lightweight convolutional neural network |
CN113408423B (en) * | 2021-06-21 | 2023-09-05 | 西安工业大学 | Aquatic product target real-time detection method suitable for TX2 embedded platform |
CN113468992A (en) * | 2021-06-21 | 2021-10-01 | 四川轻化工大学 | Construction site safety helmet wearing detection method based on lightweight convolutional neural network |
CN113408423A (en) * | 2021-06-21 | 2021-09-17 | 西安工业大学 | Aquatic product target real-time detection method suitable for TX2 embedded platform |
CN113343949B (en) * | 2021-08-03 | 2021-10-29 | 中国航空油料集团有限公司 | Pedestrian detection model training method for universal embedded platform |
CN113343949A (en) * | 2021-08-03 | 2021-09-03 | 中国航空油料集团有限公司 | Pedestrian detection model training method for universal embedded platform |
CN113723526B (en) * | 2021-08-31 | 2023-04-18 | 电子科技大学 | Method for identifying different types of craters |
CN113723526A (en) * | 2021-08-31 | 2021-11-30 | 电子科技大学 | Method for identifying different types of craters |
CN114623727A (en) * | 2022-03-14 | 2022-06-14 | 北京理工大学 | Laser imaging short-range detection target identification method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110321874A (en) | A kind of light-weighted convolutional neural networks pedestrian recognition method | |
CN108256544B (en) | Picture classification method and device, robot | |
CN109344736B (en) | Static image crowd counting method based on joint learning | |
CN108399380A (en) | A kind of video actions detection method based on Three dimensional convolution and Faster RCNN | |
CN108108751B (en) | Scene recognition method based on convolution multi-feature and deep random forest | |
CN110175613A (en) | Street view image semantic segmentation method based on Analysis On Multi-scale Features and codec models | |
CN110287960A (en) | The detection recognition method of curve text in natural scene image | |
CN107871101A (en) | A kind of method for detecting human face and device | |
CN110197182A (en) | Remote sensing image semantic segmentation method based on contextual information and attention mechanism | |
CN103258214B (en) | Based on the Classifying Method in Remote Sensing Image of image block Active Learning | |
CN106599869A (en) | Vehicle attribute identification method based on multi-task convolutional neural network | |
CN107742107A (en) | Facial image sorting technique, device and server | |
CN110059741A (en) | Image-recognizing method based on semantic capsule converged network | |
CN106599925A (en) | Plant leaf identification system and method based on deep learning | |
CN110689012A (en) | End-to-end natural scene text recognition method and system | |
CN111652273B (en) | Deep learning-based RGB-D image classification method | |
CN110197152A (en) | A kind of road target recognition methods for automated driving system | |
CN110175249A (en) | A kind of search method and system of similar pictures | |
CN109241858A (en) | A kind of passenger flow density detection method and device based on rail transit train | |
CN108549794A (en) | A kind of secondary protein structure prediction method | |
CN108509833A (en) | A kind of face identification method, device and equipment based on structured analysis dictionary | |
CN106156798A (en) | Scene image classification method based on annular space pyramid and Multiple Kernel Learning | |
CN106874913A (en) | A kind of vegetable detection method | |
CN107766828A (en) | UAV Landing Geomorphological Classification method based on wavelet convolution neutral net | |
CN109978074A (en) | Image aesthetic feeling and emotion joint classification method and system based on depth multi-task learning |
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: 20191011 |