CN108230389A - Unmanned plane localization method based on color space study - Google Patents
Unmanned plane localization method based on color space study Download PDFInfo
- Publication number
- CN108230389A CN108230389A CN201611153787.0A CN201611153787A CN108230389A CN 108230389 A CN108230389 A CN 108230389A CN 201611153787 A CN201611153787 A CN 201611153787A CN 108230389 A CN108230389 A CN 108230389A
- Authority
- CN
- China
- Prior art keywords
- layers
- image
- cooperative target
- color space
- coordinate
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- 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
- G06T2207/30204—Marker
- G06T2207/30208—Marker matrix
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses the unmanned plane localization methods learnt based on color space, include the following steps, select a cooperative target;Define a depth neuroid;By the value of SGD algorithmic minimizings L, the parameter value of depth neuroid is solved;Take a width test image Zt, by ZtIt is input in input model M, then exports S at this time, then calculate Yt;Obtain coordinate of the central point of cooperative target image under camera coordinates system;It obtains coordinate of the central point of four cooperative target images under camera coordinates system, takes the mean value of coordinate of the central point of four cooperative target images under camera coordinates system, the mean value is then positioning result.The present invention does not need to artificially set threshold value, and can adapt to a large amount of environmental change by abundant sample.It is compared to the method in existing converting colors space, this method can learn to preferably, more complicated, and be nonlinear color space by sample.
Description
Technical field
The present invention relates to unmanned plane localization methods.
Background technology
Existing unmanned plane localization method shoots target by camera, detects the position of target in the picture, so as to
The anti-position for releasing target under unmanned plane coordinate system, wherein target is the cooperative target of known features.The drawback is that utilize figure
As easily being influenced to target detection by environmental change.It is gentle in different illumination such as when cooperative target is red square
It waits under environment, the target image variation that camera is shot is bigger, if with traditional two-value method, needs artificially different
Different threshold values is set, and the setting of threshold value relies on experience, this greatly reduces the reliability of positioning under environment.
Invention content
It is insufficient in existing product the purpose of the present invention is overcoming, a kind of unmanned plane positioning learnt based on color space is provided
Method.
In order to achieve the above object, the present invention is achieved by the following technical solutions:
Based on the unmanned plane localization method of color space study, include the following steps:
(1), a cooperative target is selected, the cooperative target is a square color block Q with color C,
The physics length of side of middle Q is l, will have N web to have the aggregated label of cooperative target image for I, will there is the collection that N web has label image
It closes labeled as Y, marks the region of Regional Representative's color C that pixel value is 1 in image, the pixel value in other regions is 0;
(2), a depth neuroid is defined, by the set I of cooperative target image and with the set Y for marking image
It is all input in depth neuroid, then exports L, S, SI, j, k, the SI, j, kIt is i for coordinate in S in S, one of j, k
Element, S are expression of each pixel in the color space learnt;
(3), by the value of SGD algorithmic minimizings L, the parameter value of depth neuroid is solved, obtains parameter at this time
It is worth corresponding depth neuroid, depth neuroid at this time is labeled as model M;
(4), a width test image Z is takent, by ZtIt is input in model M, then exports S at this time, then calculate Yt, it is describedThe S:,:, kThe matrix that all values of k form, Y are taken for third dimension coordinate in S at this timetEqual to 1
Region is the region Q of the color C of prediction, calculates region Q minimum external square, the pixel coordinate of square center is marked
For (cx, cy), it is by the wide label of the pixel of square;
(5), the internal reference matrix of camera is labeled as K,
Focal lengths of the f for the camera, (u0, v0) be camera imaging central point, then
The central point of cooperative target image is in the homogeneous coordinates of camera imaging plane:
The K-1Inverse matrix for matrix K;
(6), show that coordinate of the central point of cooperative target image under camera coordinates system is:
(7), the cooperative target of four different colours C is chosen, the cooperative target of four different colours is arranged in matrix pattern
Cloth obtains coordinate of the central point of four cooperative target images under camera coordinates system, takes the center of four cooperative target images
The mean value of coordinate of the point under camera coordinates system, the mean value are then positioning result.
The present invention depth neuroid be:Input layer -1x1 convolutional layers A-ReLU1 layers -1x1 convolutional layers B -
ReLU2 layers -1x1 volumes base C-softmax layers-crossentropy layers, softmax layers are grader, and described image passes through
Input layer inputs, then exports C1 by 1x1 convolutional layers A, and the C1 passes through 1x1 convolutional layers by ReLU1 layers of output R1, the R1
B exports C2, and the C2 exports C3 by ReLU2 layers of output R2, the R2 by 1x1 convolutional layers C, and C3 is new color space,
The C3 is input to crossentropy layers of output L by softmax layers of output S, the S and the set Y with label image,
Described image is the set I of cooperative target image or test image Zt, ReLU1 layers described, ReLU1 layers 2, ReLU3 layers of definition
All it is f (x)=max (0, x), wherein x is the independent variable of function f.
Softmax layers of definition is:
The θ is softmax layers of classification factor, the θkThe row k of θ is represented, it is describedFor θkTransposition, it is described
C3I, j,:The column vector that all elements of i and j form is taken respectively for the first two coordinate in C3.θ ∈ R of the present invention2×128。
C1 ∈ R of the present inventionH×W×16, R1 ∈ RH×W×16, C2 ∈ RH×W×64, R2 ∈ RH×W×64, C3 ∈ RH×W×128, S ∈ RH×W×2,
The H is the height of image, and W is the width of image, and R is real number.
YI, jIt is i for coordinate in Y, the element of j, SI, j, lFor k
S when=1I, j, kValue, SI, j, 2S during for k=2I, j, kValue.
The convolution kernel of 1x1 convolutional layers A is K1, the K1∈R1×1×3×16, the convolution kernel of the 1x1 convolutional layers B is K2, K2∈
R1×1×16×64, the convolution kernel of the 1x1 convolutional layers C is K3, the K3∈R1×1×64×128。
Beneficial effects of the present invention are as follows:The present invention in a large amount of sample, learns one newly by deep neural network
Color space, while learn a grader so that in the color space learnt, the color characteristics under different scenes
It can be opened by automatic distinguishing.Softmax layers of output in neural network, i.e., each pixel is in the color space learnt
It represents.Compared to the method for more existing used threshold determination, this method does not need to artificially set threshold value, and can be by abundant
Sample adapts to a large amount of environmental change.The method in existing converting colors space is compared to, this method, can by sample
Learn to preferably, more complicated, and be nonlinear color space.
Specific embodiment
Technical scheme of the present invention is described further below:
Based on the unmanned plane localization method of color space study, include the following steps:
(1), a cooperative target is selected, the cooperative target is a square color block Q with color C,
The physics length of side of middle Q is l, will have N web to have the aggregated label of cooperative target image for I, will there is the collection that N web has label image
It closes labeled as Y, marks the region of Regional Representative's color C that pixel value is 1 in image, the pixel value in other regions is 0;
(2), a depth neuroid is defined, by the set I of cooperative target image and with the set for marking image
Y is input in depth neuroid, then exports L, S, SI, j, k, the SI, j, kIt is i for coordinate in S in S, one of j, k
Element, S are expression of each pixel in the color space learnt;
(3), by the value of SGD algorithmic minimizings L, the parameter value of depth neuroid is solved, obtains parameter at this time
It is worth corresponding depth neuroid, depth neuroid at this time is labeled as model M;
(4), a width test image Z is takent, by ZtIt is input in model M, then exports S at this time, then calculate Yt, it is describedThe S:,:, kThe matrix that all values of k form, Y are taken for third dimension coordinate in S at this timetEqual to 1
Region is the region Q of the color C of prediction, calculates region Q minimum external square, the pixel coordinate of square center is marked
For (cx, cy), it is by the wide label of the pixel of square;
(5), the internal reference matrix of camera is labeled as K,
Focal lengths of the f for the camera, (u0, v0) be camera imaging central point, then
The central point of cooperative target image is in the homogeneous coordinates of camera imaging plane:
The K-1Inverse matrix for matrix K;
(6), show that coordinate of the central point of cooperative target image under camera coordinates system is:
(7), the cooperative target of four different colours C is chosen, the cooperative target of four different colours is arranged in matrix pattern
Cloth obtains coordinate of the central point of four cooperative target images under camera coordinates system, takes the center of four cooperative target images
The mean value of coordinate of the point under camera coordinates system, the mean value are then positioning result.
The present invention depth neuroid be:Input layer -1x1 convolutional layers A-ReLU1 layers -1x1 convolutional layers B -
ReLU2 layers -1x1 volumes base C-softmax layers-crossentropy layers, softmax layers are grader, and described image passes through
Input layer inputs, then exports C1 by 1x1 convolutional layers A, and the C1 passes through 1x1 convolutional layers by ReLU1 layers of output R1, the R1
B exports C2, and the C2 exports C3 by ReLU2 layers of output R2, the R2 by 1x1 convolutional layers C, and C3 is new color space,
The C3 is input to crossentropy layers of output L by softmax layers of output S, the S and the set Y with label image,
Described image is the set I of cooperative target image or test image Zt。
Softmax layers of definition is:
The θ is softmax layers of classification factor, the θkThe row k of θ is represented, it is describedFor θkTransposition, it is described
C3I, j,:The column vector that all elements of i and j form is taken respectively for the first two coordinate in C3.θ ∈ R of the present invention2×128。
C1 ∈ R of the present inventionH×W×16, R1 ∈ RH×W×16, C2 ∈ RH×W×64, R2 ∈ RH×W×64, C3 ∈ RH×W×128, S ∈ RH×W×2,
The H is the height of image, and W is the width of image, and R is real number.
YI, jIt is i for coordinate in Y, the element of j, SI, j, 1For k
S when=1I, j, kValue, SI, j, 2S during for k=2I, j, kValue.
The convolution kernel of 1x1 convolutional layers A is K1, the K1∈R1×1×3×16, the convolution kernel of the 1x1 convolutional layers B is K2, K2∈
R1×1×16×64, the convolution kernel of the 1x1 convolutional layers C is K3, the K3∈R1×1×64×128.By the value of SGD algorithmic minimizings L,
The parameter value of depth neuroid is solved, the parameter value of solution refers to solving K at this time1Value, K2Value, K3Value, θ values, from
And obtain input model M.
The present invention in a large amount of sample, learns a new color space, while learn one by deep neural network
A grader so that in the color space learnt, the color characteristics under different scenes can be opened by automatic distinguishing.Nerve
Softmax layers of output in network, i.e., expression of each pixel in the color space learnt.Compared to more existing used threshold
It is worth the method for judgement, this method does not need to artificially set threshold value, and can adapt to a large amount of environment by abundant sample and become
Change, positioning is relatively reliable.The method in existing converting colors space is compared to, this method can learn by sample to more
It is suitable, more complicated, and be nonlinear color space.
By the method for deep learning, from sample learning to a more preferably color space, in this color space,
Can the easier color characteristics distinguished under varying environment, increase unmanned plane positioning robustness under various circumstances.
It should be noted that listed above is only a kind of specific embodiment of the invention.It is clear that the invention is not restricted to
Upper embodiment, can also be there are many deforming, in short, those of ordinary skill in the art can directly lead from present disclosure
All deformations for going out or associating, are considered as protection scope of the present invention.
Claims (7)
1. the unmanned plane localization method based on color space study, which is characterized in that include the following steps:
(1), a cooperative target is selected, the cooperative target is a square color block Q with color C, wherein Q's
The physics length of side is l, will have N web to have the aggregated label of cooperative target image for I, will there is the set mark that N web has label image
Y is denoted as, marks the region of Regional Representative's color C that pixel value is 1 in image, the pixel value in other regions is 0;
(2), define a depth neuroid, by the set I of cooperative target image and with mark image set Y it is all defeated
Enter into depth neuroid, then export L, S, SI, j, k, the SI, j, kIt is i for coordinate in S, an element of j, k, institute
S is stated as expression of each pixel in the color space learnt;
(3), by the value of SGD algorithmic minimizings L, the parameter value of depth neuroid is solved, obtains parameter value pair at this time
The depth neuroid answered, depth neuroid at this time are labeled as model M;
(4), a width test image Z is takent, by ZtIt is input in model M, then exports S at this time, then calculate Yt, it is describedThe S:,:, kThe matrix that all values of k form, Y are taken for third dimension coordinate in S at this timetEqual to 1
Region is the region Q of the color C of prediction, calculates region Q minimum external square, the pixel coordinate of square center is marked
For (cx, cy), it is by the wide label of the pixel of square;
(5), the internal reference matrix of camera is labeled as K,
Focal lengths of the f for the camera, (u0, v0) be camera imaging central point, then cooperate
The central point of target image is in the homogeneous coordinates of camera imaging plane:
The K-1Inverse matrix for matrix K;
(6), show that coordinate of the central point of cooperative target image under camera coordinates system is:
(7), the cooperative target of four different colours C is chosen, the cooperative target of four different colours is arranged in matrix pattern, obtained
Go out coordinate of the central point of four cooperative target images under camera coordinates system, take the central point of four cooperative target images in phase
The mean value of coordinate under machine coordinate system, the mean value are then positioning result.
2. the unmanned plane localization method according to claim 1 based on color space study, which is characterized in that the depth god
It is through metanetwork:Input layer-B-ReLU2 layers -1x1 volumes base C of 1x1 convolutional layers A-ReLU1 layers -1x1 convolutional layers -
Softmax layers-crossentropy layers, described softmax layers is grader, and described image is inputted by input layer, then is passed through
1x1 convolutional layers A exports C1, and the C1 exports C2 by ReLU1 layers of output R1, the R1 by 1x1 convolutional layers B, and the C2 leads to
It crosses ReLU2 layers of output R2, the R2 and C3 is exported by 1x1 convolutional layers C, the C3 is new color space, and the C3 passes through
Softmax layers of output S, the S and the set Y with label image are input to crossentropy layers of output L, and described image is
The set I of cooperative target image or test image Zt, ReLU1 layers described, ReLU2 layers, ReLU3 layers of definition be f (x)=max
(0, x), wherein x are the independents variable of function f.
3. the unmanned plane localization method according to claim 2 based on color space study, which is characterized in that described
Softmax layers of definition is:
The θ is softmax layers of classification factor, the θkThe row k of θ is represented, it is describedFor θkTransposition, the C3I, j,:
The column vector that all elements of i and j form is taken respectively for the first two coordinate in C3.
4. the unmanned plane localization method according to claim 3 based on color space study, which is characterized in that the θ ∈ R2 ×128。
5. the unmanned plane localization method according to claim 2 based on color space study, which is characterized in that the C1 ∈ RH ×W×16, R1 ∈ RH×W×16, C2 ∈ RH×W×64, R2 ∈ RH×W×64, C3 ∈ RH×W×128, S ∈ RH×W×2, height of the H for image, W
For the width of image, R is real number.
6. the unmanned plane localization method according to claim 2 based on color space study, which is characterized in that describedThe YI, jIt is i for coordinate in Y, the element of j, the SI, j, 1
S during for k=1I, j, kValue, the SI, j, 2S during for k=2I, j, kValue.
7. the unmanned plane localization method according to claim 2 based on color space study, which is characterized in that described 1x1 volumes
The convolution kernel of lamination A is K1, the K1∈R1×1×3×16, the convolution kernel of the 1x1 convolutional layers B is K2, the K2∈R1×1×16×64,
The convolution kernel of the 1x1 convolutional layers C is K3, the K3∈R1×1×64×128。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611153787.0A CN108230389B (en) | 2016-12-14 | 2016-12-14 | Unmanned aerial vehicle positioning method based on color space learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611153787.0A CN108230389B (en) | 2016-12-14 | 2016-12-14 | Unmanned aerial vehicle positioning method based on color space learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108230389A true CN108230389A (en) | 2018-06-29 |
CN108230389B CN108230389B (en) | 2021-07-13 |
Family
ID=62638966
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611153787.0A Active CN108230389B (en) | 2016-12-14 | 2016-12-14 | Unmanned aerial vehicle positioning method based on color space learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108230389B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102967305A (en) * | 2012-10-26 | 2013-03-13 | 南京信息工程大学 | Multi-rotor unmanned aerial vehicle pose acquisition method based on markers in shape of large and small square |
CN103218607A (en) * | 2013-04-11 | 2013-07-24 | 北京航空航天大学 | Cooperative target designing and locating method for unmanned aerial vehicle autonomous landing |
CN104392234A (en) * | 2014-11-27 | 2015-03-04 | 中国人民解放军国防科学技术大学 | Image fast Fourier transformation (FFT) symbol information based unmanned aerial vehicle autonomous landing target detection method |
CN106054931A (en) * | 2016-07-29 | 2016-10-26 | 北方工业大学 | Unmanned aerial vehicle fixed-point flight control system based on visual positioning |
-
2016
- 2016-12-14 CN CN201611153787.0A patent/CN108230389B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102967305A (en) * | 2012-10-26 | 2013-03-13 | 南京信息工程大学 | Multi-rotor unmanned aerial vehicle pose acquisition method based on markers in shape of large and small square |
CN103218607A (en) * | 2013-04-11 | 2013-07-24 | 北京航空航天大学 | Cooperative target designing and locating method for unmanned aerial vehicle autonomous landing |
CN104392234A (en) * | 2014-11-27 | 2015-03-04 | 中国人民解放军国防科学技术大学 | Image fast Fourier transformation (FFT) symbol information based unmanned aerial vehicle autonomous landing target detection method |
CN106054931A (en) * | 2016-07-29 | 2016-10-26 | 北方工业大学 | Unmanned aerial vehicle fixed-point flight control system based on visual positioning |
Non-Patent Citations (4)
Title |
---|
ALI PIRSHAYAN 等: "Target Localization Using Cooperative Unmanned Aerial Vehicles", 《ADVANCES IN COMPUTER SCIENCE: AN INTERNATIONAL JOURNAL》 * |
刘全波 等: "基于视觉的无人机自动着陆定位算法", 《电子产品世界》 * |
王保平 等: "基于局部颜色_空间特征的图像检索方法研究", 《计算机应用与软件》 * |
陈盛福: "基于视觉的无人机自动着陆相关技术研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑(月刊)》 * |
Also Published As
Publication number | Publication date |
---|---|
CN108230389B (en) | 2021-07-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106326937B (en) | Crowd density distribution estimation method based on convolutional neural networks | |
CN105631880B (en) | Lane line dividing method and device | |
CN106127702B (en) | A kind of image defogging method based on deep learning | |
CN107507172A (en) | Merge the extra high voltage line insulator chain deep learning recognition methods of infrared visible ray | |
CN101394573B (en) | Panoramagram generation method and system based on characteristic matching | |
CN110188835B (en) | Data-enhanced pedestrian re-identification method based on generative confrontation network model | |
CN108509920B (en) | CNN-based face recognition method for multi-patch multi-channel joint feature selection learning | |
CN107993260A (en) | A kind of light field image depth estimation method based on mixed type convolutional neural networks | |
CN108108746A (en) | License plate character recognition method based on Caffe deep learning frames | |
CN107424142A (en) | A kind of weld joint recognition method based on saliency detection | |
CN103679677B (en) | A kind of bimodulus image decision level fusion tracking updating mutually based on model | |
CN108319973A (en) | Detection method for citrus fruits on tree | |
CN108614994A (en) | A kind of Human Head Region Image Segment extracting method and device based on deep learning | |
CN103426182A (en) | Electronic image stabilization method based on visual attention mechanism | |
CN105719247A (en) | Characteristic learning-based single image defogging method | |
CN110378247A (en) | Virtual objects recognition methods and device, storage medium and electronic device | |
CN109045676B (en) | Chinese chess recognition learning algorithm and robot intelligent system and method based on algorithm | |
CN109191429B (en) | 3D printing nozzle detection method based on machine vision | |
CN107067015A (en) | A kind of vehicle checking method and device based on multiple features deep learning | |
CN109754440A (en) | A kind of shadow region detection method based on full convolutional network and average drifting | |
CN105825204A (en) | Method for identifying power equipment meter at night through intelligent camera | |
CN106991428A (en) | Insect image-recognizing method based on adaptive pool model | |
CN110298893A (en) | A kind of pedestrian wears the generation method and device of color identification model clothes | |
CN107862333A (en) | A kind of method of the judgment object combustion zone under complex environment | |
CN104751487B (en) | A kind of moving target detecting method based on the plane discoloration frames of colored RGB tri- difference |
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 |