CN107767416A - The recognition methods of pedestrian's direction in a kind of low-resolution image - Google Patents
The recognition methods of pedestrian's direction in a kind of low-resolution image Download PDFInfo
- Publication number
- CN107767416A CN107767416A CN201710791630.9A CN201710791630A CN107767416A CN 107767416 A CN107767416 A CN 107767416A CN 201710791630 A CN201710791630 A CN 201710791630A CN 107767416 A CN107767416 A CN 107767416A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- image
- leg
- towards
- mrow
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 30
- 238000005070 sampling Methods 0.000 claims abstract description 22
- 238000012545 processing Methods 0.000 claims abstract description 16
- 239000000284 extract Substances 0.000 claims abstract description 12
- 238000003062 neural network model Methods 0.000 claims abstract description 8
- 230000003321 amplification Effects 0.000 claims abstract description 5
- 238000003199 nucleic acid amplification method Methods 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 6
- 238000000605 extraction Methods 0.000 claims description 3
- 238000005303 weighing Methods 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims description 2
- 239000000203 mixture Substances 0.000 claims description 2
- 230000004044 response Effects 0.000 claims description 2
- 238000010586 diagram Methods 0.000 claims 3
- 230000001537 neural effect Effects 0.000 claims 1
- 238000000926 separation method Methods 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 abstract description 15
- 238000013135 deep learning Methods 0.000 abstract description 7
- 238000011160 research Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000000153 supplemental effect Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013497 data interchange Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000000686 essence Substances 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000003475 lamination Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000001144 postural effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30196—Human being; Person
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a kind of recognition methods of pedestrian's direction in low-resolution image, comprise the steps：(1) pedestrian is made towards data set and label, three kinds of generation pedestrian head, leg, whole body training samples；(2) mirror image processing and multi-scale sampling operation amplification training dataset are carried out to image；(3) it is trained using framework caffe, obtains corresponding neural network model；(4) neural network model for utilizing (3) to train extracts pedestrian image triple channel feature respectively；(5) classified using Softmax graders, obtained final towards recognition result.The invention provides a kind of generation method of the pedestrian towards image data set, the operation such as the mirror image processing taken and multi-scale sampling can effectively expand small data set, mitigate over-fitting；The convolutional neural networks model for the identification pedestrian's direction trained by deep learning can extract the feature of more discrimination, and the weighting to pedestrian image triple channel feature causes more accurate reasonable towards classification results.
Description
Technical field
The present invention relates to the research field of image procossing, the identification of pedestrian's direction in more particularly to a kind of low-resolution image
Method.
Background technology
In research fields such as image recognition, segmentation, reconstructions, the direction of objects in images is to seek as a kind of supplemental characteristic
Look for key point, description posture, the important foundation for judging movement tendency.Object is believed towards multiple image is relied primarily in present image
Breath, motion track information or camera calibration information obtain, for object in single-frame images relative towards the research of identification
It is less.Tulsiani etc. article " Pose Induction for Novel Object Categories " (IEEE
International Conference on Computer Vision,2015:64-72) propose the method by concluding, profit
The direction of no tag object class is predicted with the small-scale seed set of tagged asset class.This method propose based on CNN
The postural assessment system of (Convolutional Neural Network, convolutional neural networks), can be different classes of in utilization
Between object while similitude, implicitly learn the General Expression to be come in handy to predicting the direction of different classes of object.This
Kind method trains the convolution that " the broad category device " come has been played in CNN and is extracting various feature, in terms of processing picture displacement
Advantage, yielded good result on some data sets, but its height to tape label seed set relies on and locating
When managing the image larger with the image difference in tape label seed set, direction can not be concluded well.
In practical application scene, due to the influence of the environmental factors such as video camera setting, illumination, pedestrian image generally has
The characteristics of visual angle is various, serious shielding, relatively low resolution ratio, make object of the pedestrian towards in than resolution normal image in resolution image
Direction is more difficult, so some the pedestrian image processing methods of currently used pedestrian's orientation information as supplemental characteristic, to row
The judgement of people's direction generally requires to use camera calibration information, pedestrian movement's trace information etc., but our images for grasping
Data often lack corresponding calibration and space time information so that these methods have larger limitation in actual applications.For
This, the present invention proposes a kind of recognition methods of pedestrian's direction in low-resolution image, and pedestrian's direction in single-frame images is entered
Row identification.
The content of the invention
The shortcomings that it is a primary object of the present invention to overcome prior art and deficiency, there is provided row in a kind of low-resolution image
The recognition methods of people's direction, the convolutional neural networks model for the identification pedestrian's direction trained by deep learning can extract more
The feature of discrimination, the weighting to pedestrian image triple channel feature make it that classification results are more accurate reasonable.
In order to achieve the above object, the present invention adopts the following technical scheme that：
The recognition methods of pedestrian's direction, comprises the steps in a kind of low-resolution image of the present invention：
The recognition methods of pedestrian's direction, comprises the steps in a kind of low-resolution image of the present invention：
(1) pedestrian is made towards data set and label, three kinds of generation pedestrian head, leg, whole body training samples；
(2) mirror image processing and multi-scale sampling operation amplification training dataset are carried out to image；
(3) towards label depth is inputted three kinds of pedestrian head after amplification, leg, whole body image training samples and its respectively
It is trained in degree learning framework caffe, obtains corresponding neural network model；
(4) image of identification pedestrian's direction is treated, the neural network model trained using step (3) extracts pedestrian respectively
Image three channel feature；
(5) feature of different passages is weighted according to pedestrian image structure priori, utilizes Softmax graders
Classified, obtained final towards recognition result.
As preferential technical scheme, in the step (1), pedestrian towards data set and label making concrete operations such as
Under：
Four directions of pedestrian are respectively in definition image：Front, the back side, the left side, the right side, it is that every pedestrian image is distributed
One towards label, according to the priori of pedestrian image structure and identification target, proportionally cut pedestrian head, trunk,
Leg subgraph, pedestrian head and leg subgraph are chosen, assigned same with former pedestrian image towards label, generation pedestrian's head
Three kinds of portion, leg, whole body training samples.
As preferential technical scheme, in the step (2), the method for training dataset is expanded such as using mirror image processing
Under：
The image concentrated to data makees mirror image processing in horizontal direction, i.e., the pixel value of the right and left in swap image,
The mirror image of original image is obtained, to the original image G that size is MXN, pixel values of its mirror image G' in (i, j) position is：
G'[i, j]=G [N-i+1, j]
For the pedestrian image towards label for front or the back side, assign mirror image and marked with original image identical direction
Label, for the label on the mirror image right side or the left side corresponding to towards the pedestrian image that label is the left side or the right side, assigning respectively,
All mirror images are added into data set.
As preferential technical scheme, in the step (2), the method for training dataset is expanded such as using multi-scale sampling
Under：
Multiple dimensioned linear space filtering process, i.e., each pixel in pedestrian image are carried out to the image that data are concentrated
In the contiguous range of point (x, y), each pixel and corresponding multiplication, result is added up to obtain point (x, y) in neighborhood
The response at place, the matrix of these coefficients composition is corresponding wave filter, can be with by the adjustment to wave filter size and numerical value
The filter result of image different scale is obtained, realizes the purpose of multi-scale sampling, the multi-scale sampling image obtained is assigned
With original image identical towards label, added in data set.
As preferential technical scheme, in the step (2), the specific behaviour of training dataset is expanded using multi-scale sampling
Make as follows：
The filter result of image different scale is obtained by the adjustment to wave filter size and numerical value, realizes multi-scale sampling
Purpose, using motion blur build wave filter f, blurred length len, obscure angle be θ motion blur i.e. represent with
Len pixel is moved on the direction that horizontal direction angle is θ, wave filter f building process is as described below：
The size that (2-1) constructs two-dimensional matrix a f, f should meet that it is len that can accommodate a length just, and slope is
Tan θ line segment l, it is assumed that the size of matrix is axb, then a=len*cos θ+1, b=len*sin θ；
(2-2) calculates the minimum range N_D of the position and line segment l to the position (i, j) in matrix f：
N_D=jcos θ-isin θ
(2-3) calculates the coefficient value at (i, j) place according to minimum range N_D：
F (I, j)=max (1-N_D, 0)；
(2-4) f is normalized：
Using multiple pixel sizes as blurred length, different wave filters is built respectively with multiple angles image is filtered
Processing, the multi-scale sampling image obtained is added in data set, all labels are identical with original image.
As preferential technical scheme, in the step (4), extraction pedestrian image triple channel feature concrete operations are as follows：
Definition pedestrian head, leg, whole body images are characterized as pedestrian's triple channel feature, by the row of the direction to be detected cut
The convolution for pedestrian head, leg, whole body images that head part, leg, whole body images difference input step (3) train is refreshing
Through network, head is respectively extracted, three leg, whole body passages recognize feature.
It is as follows to the characteristic weighing formula of different passages in the step (5) as preferential technical scheme：
F (X)=ω1f(head)+ω2f(body)+ω3f(leg)
Wherein：F (head), f (body), f (leg) are respectively pedestrian's head that the neural network model corresponding to extracts
Portion, whole body and leg characteristics of image, ωi, i=1,2,3 be weight coefficient, and f (X) is the final spy for differentiating pedestrian's direction
Sign.
The present invention compared with prior art, has the following advantages that and beneficial effect：
(1), the invention provides a kind of generation method of the pedestrian towards image data set, the mirror image processing taken and more chis
The operations such as degree sampling can effectively expand small data set, mitigate over-fitting, while strengthen the Shandong to low-resolution image processing
Rod；
(2), the present invention proposes pedestrian image triple channel feature, passes through the convolutional neural networks model extraction that trains
Important pedestrian's architectural feature such as head, leg, whole body, more information for determining pedestrian's direction in image can be obtained, to low resolution
In rate image pedestrian's direction identification and classification provide necessary and sufficient information；
(3), the convolutional neural networks for identification pedestrian's direction that the present invention is trained by deep learning can extract more area
The feature of indexing, the weighting to different characteristic make it that classification results are more accurate reasonable.
(4), the present invention can preferably be located by identifying pedestrian's orientation information to the visual angle change of pedestrian image
Reason, contribute to pedestrian's gesture recognition, the again higher level application such as identification.
Brief description of the drawings
Fig. 1 is the overall flow of the recognition methods key step of pedestrian's direction in low-resolution image disclosed by the invention
Figure.
Fig. 2 is the network structure of deep learning framework used in the present invention.
Embodiment
With reference to embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are unlimited
In this.
Embodiment
The schematic flow sheet of the present invention is as shown in Figure 1.Concrete technical scheme is：First make pedestrian towards image data set and
Label, three kinds of generation pedestrian head, leg, whole body training samples, mirror image processing is carried out to image and multi-scale sampling operation is expanded
Increase training dataset, be then trained respectively using the deep learning framework caffe samples different to three kinds, using training
Three kinds of convolutional neural networks extract pedestrian image head to be detected, leg, whole body triple channel feature, according to pedestrian image structure
Priori is weighted to the feature of different passages, is classified using Softmax graders.
Specifically include following main technical points：
1. pedestrian is towards the making of data set and label
Four directions of pedestrian are respectively in present invention definition image：Front, the back side, the left side, the right side, wherein, training set
In each towards there is 300 pedestrian images, totally 1200 images, it is each towards there is 50 pedestrian images in test set, totally 200
Image.It is that every pedestrian image is distributed accordingly towards label, according to the priori of pedestrian image structure and identification target, presses
According to 1:2:3 ratio cuts pedestrian head, trunk, leg subgraph, is assigned and former pedestrian image to head and leg subgraph
Same label, three kinds of generation pedestrian head, leg, whole body training samples.
2. mirror image processing
The image concentrated to data makees the mirror image processing in horizontal direction, and the size for obtaining original image G first is MXN, then
Value calculation formula of the mirror image G' in (i, j) position is as follows：
G'[i, j]=G [N-i+1, j]
For the pedestrian image towards label for front or the back side, mirror image and original image identical label are assigned, it is right
The label on the mirror image right side or the left side corresponding to towards the pedestrian image that label is the left side or the right side, assign respectively, by institute
There is mirror image to add data set, make one times of training dataset popularization, totally 2400 images.
3. multi-scale sampling
The filter result of image different scale is obtained by the adjustment to wave filter size and numerical value, it is possible to achieve multiple dimensioned
The purpose of sampling.In the present embodiment, in order to reflect the relative motion of pedestrian and video camera, wave filter is built using motion blur
f.Blurred length is len, and the motion blur that fuzzy angle is θ represents to be that len is moved on θ direction with horizontal direction angle
Individual pixel, wave filter f building process are as described below：
(1) constructing two-dimensional matrix a f, f size should meet that it is len, slope tan that can accommodate a length just
θ line segment l, it is assumed that the size of matrix is axb, then a=len*cos θ+1, b=len*sin θ；
(2) to the position (i, j) in matrix f, the minimum range N_D of the position and line segment l is calculated：
N_D=jcos θ-isin θ
(3) coefficient value at (i, j) place is calculated according to minimum range N_D：
F (i, j)=max (1-N_D, 0)；
(4) f is normalized：
Using 9 pixel sizes as blurred length, with -10 °, -5 °, 0 °, 5 °, 10 ° of five angles build different filters respectively
Ripple device is filtered processing to image, the multi-scale sampling image obtained is added in data set, all labels and artwork
As identical.
After multi-scale sampling image is added into data set, 6 times of training set popularization, increased to by 2400 images
14400 images, form final training dataset.
4. training convolutional neural networks
The present invention extracts pedestrian towards feature using deep learning framework caffe training convolutional neural networks, and caffe is one
Individual clear efficient deep learning framework, there is provided be available for each channel type of user-defined model, can be easily extended to
In new task and setting.Caffe frameworks mainly have a Solver, Net, Layer, five components of Blob, Proto, wherein
Solver is responsible for carrying out the training of depth network, and each Solver includes a training network object and a test network pair
As network N et is made up of several Layer, and Layer is used for defining the details such as convolution, pond, each Layer input and defeated
Go out and stored with Blob matrix structures, and Proto is a kind of data interchange format, for the structure definition of network model, storage and
Read.Trained using caffe as follows towards the convolutional neural networks specific implementation details of identification for pedestrian：
(1) image is concentrated to be converted into lmdb three pedestrian head, leg and whole body position subgraphs in data set first
Form, and upset the different orders towards image, then calculate the average image of subgraph image set, the subgraph by band towards label
Collection image and average image are trained as caffe input.
(2) as shown in Fig. 2 network structure uses the network structure for being used to train Imagenet provided in caffe, including
Five convolutional layers and three full articulamentums, first convolutional layer include the convolution kernel that 96 sizes are 11*11*3, second convolution
Layer is using the output of first convolutional layer as input, and the convolution kernel for being 5*5*48 with 256 sizes carries out convolution, the two volumes
The output of lamination has all carried out maximum pondization and LRN standardization, followed by three no carry out pond and standardization three
Convolution kernel that convolution kernel that convolutional layer is respectively 3*3*256 comprising 384 sizes, 384 sizes are 3*3*192,256 3*
3*192 convolution kernel.Three full articulamentums respectively have 4096,4096,4 neurons.
(3) in the training process, optimized parameter is solved using stochastic gradient descent algorithm, often trains 500 times and carry out once
Test, learning rate 0.001, maximum iteration are 20000 times, are trained on GPU.
5. extract feature
The pin that head, leg, whole body images the difference input step (3) of the pedestrian of the direction to be detected cut is trained
To pedestrian head, leg, whole body images convolutional neural networks, respectively using the dimensional vector of output 4096 of fc7 layers as each logical
The feature in road.
6. weighting classification
It is as follows to the characteristic weighing formula of different passages：
F (X)=ω1f(head)+ω2f(body)+ω3f(leg)
Wherein：F (head), f (body), f (leg) are respectively the row that the convolutional neural networks trained with (3) extract
Head part, whole body and leg feature, ωi, i=1,2,3 be weight coefficient, according to pedestrian image structure priori, takes ω1=
1/6, ω2=1/2, ω3=1, f (X) are the final feature for differentiating pedestrian's direction.
The present invention is classified using Softmax graders, and the grader is applied to solve more classification problems, tag along sort
More than two values can be taken, in pedestrian towards in identification mission, it is y to take tag along sort(i)={ 0,1,2,3 } representative front,
The back side, the left side, the right side }, in Softmax recurrence, the feature f (X) of input is categorized as classification j (j=0,1,2,3) probability
For：
Wherein, θl, l=1 ..., k are the parameters of the sorter model obtained in training convolutional neural networks, probable value
That maximum classification is recognition result.
Above-described embodiment is the preferable embodiment of the present invention, but embodiments of the present invention are not by above-described embodiment
Limitation, other any Spirit Essences without departing from the present invention with made under principle change, modification, replacement, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (7)
1. the recognition methods of pedestrian's direction in a kind of low-resolution image, it is characterised in that comprise the steps：
(1) pedestrian is made towards data set and label, three kinds of generation pedestrian head, leg, whole body training samples；
(2) mirror image processing and multi-scale sampling operation amplification training dataset are carried out to image；
(3) towards label depth is inputted three kinds of pedestrian head after amplification, leg, whole body image training samples and its respectively
Practise and being trained in framework caffe, obtain corresponding neural network model；
(4) image of identification pedestrian's direction is treated, the neural network model trained using step (3) extracts pedestrian image respectively
Triple channel feature；
(5) feature of different passages is weighted according to pedestrian image structure priori, carried out using Softmax graders
Classification, is obtained final towards recognition result.
2. according to claim 1 in low-resolution image pedestrian's direction recognition methods, it is characterised in that the step
(1) in, the making concrete operations of pedestrian towards data set and label are as follows：
Four directions of pedestrian are respectively in definition image：Front, the back side, the left side, the right side, it is that every pedestrian image distributes one
Towards label, according to the priori of pedestrian image structure and identification target, pedestrian head, trunk, leg are proportionally cut
Subgraph, choose pedestrian head and leg subgraph, assign it is same with former pedestrian image towards label, generate pedestrian head,
Three kinds of leg, whole body training samples.
3. according to claim 1 in low fractional diagram picture pedestrian's direction recognition methods, it is characterised in that the step (2)
In, the method that training dataset is expanded using mirror image processing is as follows：
The image concentrated to data makees the mirror image processing in horizontal direction, i.e., the pixel value of the right and left, is obtained in swap image
The mirror image of original image, to the original image G that size is MXN, pixel values of its mirror image G' in (i, j) position is：
G'[i, j]=G [N-i+1, j]
For the pedestrian image towards label for front or the back side, mirror image is assigned with original image identical towards label, it is right
The label on the mirror image right side or the left side corresponding to towards the pedestrian image that label is the left side or the right side, assign respectively, by institute
There is mirror image to add data set.
4. according to claim 1 in low fractional diagram picture pedestrian's direction recognition methods, it is characterised in that the step (2)
In, the method that training dataset is expanded using multi-scale sampling is as follows：
Multiple dimensioned linear space filtering process, i.e., each pixel in pedestrian image are carried out to the image that data are concentrated
In the contiguous range of (x, y), each pixel and corresponding multiplication in neighborhood, result is added up to obtain point (x, y) place
Response, these coefficients composition matrix be corresponding wave filter, can be obtained by the adjustment to wave filter size and numerical value
To the filter result of image different scale, realize the purpose of multi-scale sampling, by the multi-scale sampling image obtained assign with
Original image identical is towards label, added in data set.
5. according to claim 4 in low fractional diagram picture pedestrian's direction recognition methods, it is characterised in that the step (2)
In, the concrete operations that training dataset is expanded using multi-scale sampling are as follows：
The filter result of image different scale is obtained by the adjustment to wave filter size and numerical value, realizes the mesh of multi-scale sampling
, using motion blur build wave filter f, blurred length len, obscure angle be θ motion blur i.e. represent with level
Angular separation is that len pixel is moved on θ direction, and wave filter f building process is as described below：
The size that (2-1) constructs two-dimensional matrix a f, f should meet that it is len that can accommodate a length just, and slope is tan θ
Line segment l, it is assumed that the size of matrix is axb, then a=len*cos θ+1, b=len*sin θ；
(2-2) calculates the minimum range N_D of the position and line segment l to the position (i, j) in matrix f：
N_D=jcos θ-isin θ
(2-3) calculates the coefficient value at (i, j) place according to minimum range N_D：
F (I, j)=max (1-N_D, 0)；
(2-4) f is normalized：
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
</munder>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
</mrow>
</mfrac>
</mrow>
Using multiple pixel sizes as blurred length, different wave filters is built respectively with multiple angles place is filtered to image
Reason, the multi-scale sampling image obtained is added in data set, all labels are identical with original image.
6. according to claim 1 in low-resolution image pedestrian's direction recognition methods, it is characterised in that the step
(4) in, extraction pedestrian image triple channel feature concrete operations are as follows：
Definition pedestrian head, leg, whole body images are characterized as pedestrian's triple channel feature, by pedestrian's head of the direction to be detected cut
What portion, leg, whole body images difference input step (3) trained is directed to pedestrian head, leg, the convolutional Neural net of whole body images
Network, head is extracted respectively, leg, whole body three passages recognize feature.
7. according to claim 1 in low-resolution image pedestrian's direction recognition methods, it is characterised in that the step
(5) it is as follows to the characteristic weighing formula of different passages in：
F (X)=ω1f(head)+ω2f(body)+ω3f(leg)
Wherein：F (head), f (body), f (leg) be respectively the neural network model corresponding to withing extract pedestrian head, entirely
Body and leg characteristics of image, ωi, i=1,2,3 be weight coefficient, and f (X) is the final feature for differentiating pedestrian's direction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710791630.9A CN107767416B (en) | 2017-09-05 | 2017-09-05 | Method for identifying pedestrian orientation in low-resolution image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710791630.9A CN107767416B (en) | 2017-09-05 | 2017-09-05 | Method for identifying pedestrian orientation in low-resolution image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107767416A true CN107767416A (en) | 2018-03-06 |
| CN107767416B CN107767416B (en) | 2020-05-22 |
Family
ID=61265939
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710791630.9A Expired - Fee Related CN107767416B (en) | 2017-09-05 | 2017-09-05 | Method for identifying pedestrian orientation in low-resolution image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107767416B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108490832A (en) * | 2018-03-27 | 2018-09-04 | 百度在线网络技术（北京）有限公司 | Method and apparatus for sending information |
| CN109711342A (en) * | 2018-12-27 | 2019-05-03 | 河北工程大学 | Face identification method and device |
| CN110309713A (en) * | 2019-05-22 | 2019-10-08 | 深圳壹账通智能科技有限公司 | Expression Recognition model training method, device, equipment and storage medium |
| CN110580427A (en) * | 2018-06-08 | 2019-12-17 | 杭州海康威视数字技术股份有限公司 | face detection method, device and equipment |
| CN110749317A (en) * | 2018-07-24 | 2020-02-04 | 东旭科技集团有限公司 | Navigation assisting method and device, street lamp and electronic equipment |
| CN110766093A (en) * | 2019-10-31 | 2020-02-07 | 智慧视通(杭州)科技发展有限公司 | Video target re-identification method based on multi-frame feature fusion |
| WO2021098346A1 (en) * | 2019-11-20 | 2021-05-27 | 上海商汤智能科技有限公司 | Body orientation detection method and apparatus, electronic device, and computer storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105335716A (en) * | 2015-10-29 | 2016-02-17 | 北京工业大学 | Improved UDN joint-feature extraction-based pedestrian detection method |
| CN105787478A (en) * | 2016-04-14 | 2016-07-20 | 中南大学 | Face direction change recognition method based on neural network and sensitivity parameter |
| CN106127173A (en) * | 2016-06-30 | 2016-11-16 | 北京小白世纪网络科技有限公司 | A kind of human body attribute recognition approach based on degree of depth study |
| CN106709449A (en) * | 2016-12-22 | 2017-05-24 | 深圳市深网视界科技有限公司 | Pedestrian re-recognition method and system based on deep learning and reinforcement learning |
| CN106778902A (en) * | 2017-01-03 | 2017-05-31 | 河北工业大学 | Milk cow individual discrimination method based on depth convolutional neural networks |
-
2017
- 2017-09-05 CN CN201710791630.9A patent/CN107767416B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105335716A (en) * | 2015-10-29 | 2016-02-17 | 北京工业大学 | Improved UDN joint-feature extraction-based pedestrian detection method |
| CN105787478A (en) * | 2016-04-14 | 2016-07-20 | 中南大学 | Face direction change recognition method based on neural network and sensitivity parameter |
| CN106127173A (en) * | 2016-06-30 | 2016-11-16 | 北京小白世纪网络科技有限公司 | A kind of human body attribute recognition approach based on degree of depth study |
| CN106709449A (en) * | 2016-12-22 | 2017-05-24 | 深圳市深网视界科技有限公司 | Pedestrian re-recognition method and system based on deep learning and reinforcement learning |
| CN106778902A (en) * | 2017-01-03 | 2017-05-31 | 河北工业大学 | Milk cow individual discrimination method based on depth convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| 朱龙: ""基于多部位联合的行人检测方法研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108490832A (en) * | 2018-03-27 | 2018-09-04 | 百度在线网络技术（北京）有限公司 | Method and apparatus for sending information |
| CN110580427A (en) * | 2018-06-08 | 2019-12-17 | 杭州海康威视数字技术股份有限公司 | face detection method, device and equipment |
| CN110749317A (en) * | 2018-07-24 | 2020-02-04 | 东旭科技集团有限公司 | Navigation assisting method and device, street lamp and electronic equipment |
| CN109711342A (en) * | 2018-12-27 | 2019-05-03 | 河北工程大学 | Face identification method and device |
| CN110309713A (en) * | 2019-05-22 | 2019-10-08 | 深圳壹账通智能科技有限公司 | Expression Recognition model training method, device, equipment and storage medium |
| WO2020233368A1 (en) * | 2019-05-22 | 2020-11-26 | 深圳壹账通智能科技有限公司 | Expression recognition model training method and apparatus, and device and storage medium |
| CN110766093A (en) * | 2019-10-31 | 2020-02-07 | 智慧视通(杭州)科技发展有限公司 | Video target re-identification method based on multi-frame feature fusion |
| WO2021098346A1 (en) * | 2019-11-20 | 2021-05-27 | 上海商汤智能科技有限公司 | Body orientation detection method and apparatus, electronic device, and computer storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107767416B (en) | 2020-05-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tao et al. | Smoke detection based on deep convolutional neural networks | |
| CN109344736B (en) | Static image crowd counting method based on joint learning | |
| Yu et al. | Deep learning in remote sensing scene classification: a data augmentation enhanced convolutional neural network framework | |
| Hou et al. | Change detection based on deep features and low rank | |
| CN107767416A (en) | The recognition methods of pedestrian's direction in a kind of low-resolution image | |
| Garcia-Garcia et al. | A review on deep learning techniques applied to semantic segmentation | |
| CN105574550B (en) | A kind of vehicle identification method and device | |
| CN105488517B (en) | A kind of vehicle brand type identifier method based on deep learning | |
| CN108921822A (en) | Image object method of counting based on convolutional neural networks | |
| CN104616316B (en) | Personage's Activity recognition method based on threshold matrix and Fusion Features vision word | |
| CN108875600A (en) | A kind of information of vehicles detection and tracking method, apparatus and computer storage medium based on YOLO | |
| CN106682697A (en) | End-to-end object detection method based on convolutional neural network | |
| CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN108090403A (en) | A kind of face dynamic identifying method and system based on 3D convolutional neural networks | |
| CN109886066A (en) | Fast target detection method based on the fusion of multiple dimensioned and multilayer feature | |
| CN107016357A (en) | A kind of video pedestrian detection method based on time-domain convolutional neural networks | |
| Wang et al. | FE-YOLOv5: Feature enhancement network based on YOLOv5 for small object detection | |
| CN102831427B (en) | Texture feature extraction method fused with visual significance and gray level co-occurrence matrix (GLCM) | |
| CN105894045A (en) | Vehicle type recognition method with deep network model based on spatial pyramid pooling | |
| CN110765833A (en) | Crowd density estimation method based on deep learning | |
| CN108734210A (en) | A kind of method for checking object based on cross-module state multi-scale feature fusion | |
| CN105205449A (en) | Sign language recognition method based on deep learning | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| Shao et al. | Label embedded dictionary learning for image classification |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200522 |