CN103778437B - Construction road pollution detection method based on support vector machine - Google Patents
Construction road pollution detection method based on support vector machine Download PDFInfo
- Publication number
- CN103778437B CN103778437B CN201410026324.2A CN201410026324A CN103778437B CN 103778437 B CN103778437 B CN 103778437B CN 201410026324 A CN201410026324 A CN 201410026324A CN 103778437 B CN103778437 B CN 103778437B
- Authority
- CN
- China
- Prior art keywords
- road
- feature
- matrix
- color
- image
- 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.)
- Active
Links
Abstract
The invention discloses a construction road pollution detection method based on a support vector machine. The method comprises the steps of acquiring road images and conducting characteristic extraction on the road images, selecting the optimum characteristic and a parameter r and training a support vector machine model, and finally conducting prediction and judgment on the cleanness of a construction road. According to the method, by means of the methods of principal component analysis dimensionality reduction and searching for a purest road block, road extraction speed is higher, precision is higher, and then the classification result is more accurate; twelve characteristics are extracted in all, automatic selection of the optimum characteristic and parameter can be achieved for different construction roads, and therefore the method is suitable for most construction roads and high in transportability. The method is high in accuracy and has practical significance for guiding the construction side to clean the road in time so as to avoid accidents caused by road pollution.
Description
Technical field
The present invention relates to image procossing and urban environment Risk Monitoring early warning technology field, more particularly to it is a kind of based on support
The construction road contamination detection method of vector machine.
Background technology
Construction road is due to due to construction and Jing often have the passing through of vehicle, it is easy to be covered with dust all over, serious impact city
Hold the health of city's looks and pedestrian.Construction road is high and uncertain by the frequency of dust pollution, the difficulty for causing manpower to be monitored, because
This needs a kind of construction road contamination detection method of precise and high efficiency.
Current construction road pollution detection is based primarily upon the image partition method in image procossing, the method for image segmentation
Have a lot, such as region growth method, cluster segmentation etc., these methods all can not completely reason part and non-rice habitats part point
Cut open, the speed of segmentation is unhappy, it is impossible to accomplish the real-time detection to construction road；In terms of feature selection, different roads are most
Good feature may be different, and a kind of model is difficult to be suitable for all of construction road.
The content of the invention
Present invention aims at present image segmentation speed is slow and inaccurate, the defect such as model is single, there is provided
Construction road contamination detection method kind based on support vector machine, by principal component analysiss dimensionality reduction and seeking the side for taking most pure road block
Method, extracts the speed of road faster, and precision is bigger, more accurate so as to reach classification results；By extracting 12 features, can be certainly
Dynamic selection optimal characteristics and parameter, so as to be suitable for most of construction road.
The present invention is achieved by the following technical solutions：A kind of construction road pollution detection side based on support vector machine
Method, the method comprises the steps：
Step 1：Collection road image, using the image of collection as training set, and artificial is not divided into training set totally and not
Clean two set.Image positioned at clean set is referred to as positive sample, and positioned at the image totally do not gathered negative sample is referred to as, and trains
The number for concentrating sample is N.
Step 2：Feature extraction is carried out to road image, comprising following sub-step：
(2.1) simple division road and non-rice habitats part：Because photographic head is fixed, every pictures can be unified
Remove non-rice habitats part；The number of road area is represented with n；The matrix A of a m rows n row is defined with A [m, n]；A (y, x) is represented
The value of y rows xth row in matrix A；The row k of A (k) representing matrix A, is a row vector；ATRepresent the transposed matrix of A.
(2.2) color space conversion：Native color spatial transformation to LUV color spaces, with LUVs [n, 3] road picture is recorded
The LUV color value of element, LUVs (k) represents the LUV color value of k-th road pixel point of correspondence.
(2.3) principal component analysiss reduce dimension, and the color space of 3-dimensional is dropped to 2 dimensions；Adopt with the following method：To LUVsT*
LUVs carries out eigenvalue eigendecomposition using SVD methods, that is, have：LUVsT* LUVs=V*2*VT, PC [n, 2] is LUVs*V
First two columns, i.e. the main constituent of road color, with i-th row of PC [i] representing matrix PC.
(2.4) Meanshift cluster segmentations：A matrix S [256,256] is defined first,Wherein 1 { true }=1,1 { false }=0.Do on matrix S
Meanshift is split, and rectangular window of the size for nW*nW is specified first, and nW is determined by the size of detection zone, is not less than
2, method is as follows：
(2.4.1) non-zero entry A is looked in S, if A is not present, step (2.4.4) is jumped to.
(2.4.2) in S, it is the barycenter B in nW*nW rectangular windows to calculate the size centered on A.
If (2.4.3) A equals B, central point A values are recorded, all elements centered on A in window all set in S
It is zero, returns to step (2.4.1).If A is not equal to B, B values are assigned to A, return to step (2.4.2).
(2.4.4) K central point Centers [K, 2] is obtained, it is Matrix C enters [K, 2] i-th to define Centers [i]
OK；Define one-dimensional vector Tags [n], wherein Tags [i]=argjMin (| | PC [i]-center [j] | |) (i=0,
1 ... n-1, j=0,1 ... ... k-1), Tags [n] is the class label for recording road pixel；It has been divided into K class, similar unit
Have identical class label..
(2.5) road class is chosen：First vegetarian refreshments of each road sections corresponds to a LUV color and class label, records respectively
In LUVs and Tags, the road waypoint for really arriving need to be found, remove these pixels of car pedestrian, that is, find the class label of real road
tag.Follow two principles：The area of road is big relatively to be unified with color, that is, more than number and the characteristics of color variance is little.To appointing
The class label i of meaning, has：Wherein
The number of class i is calculated, var (i) calculates the variance of class i.
(2.6) most pure path matrix block is found：A two-dimensional matrix IR [iH, iW] is defined, iH represents the height of image,
IW is the width of image, and the y rows X of IR (y, x) representing matrix IR is arranged.IR (y, x)=1 represents that pixel P (x, y) is in step
Rapid 2.1 road sections for retaining and its corresponding class label is tag in step 2.5, IS (y, x)=0 represents other feelings
Condition.Take a matrix window (nH, nW), most pure path matrix window be then in this window IR be 1 number it is most.
(2.7) color and textural characteristics are extracted：Due to the uniformity of road, color can be extracted in the matrix-block for finding
With textural characteristics.The color characteristic of extraction has 3 components of LUV, does average to the value of component respectivelyVarianceSkewnessWherein xiThe value in 3 components in for LUV.Adopt
The textural characteristics extracted with gray scale symbiosis obtain gray level co-occurrence matrixes p [i, j], calculate energy ASM=∑i∑j p(i,j)2, it is right
Than degree CON=∑sn n2(∑|i-j|=nP (i, j)), degree of associationCome to 12 features.
(2.8) scaling of feature：The characteristic of extraction differs greatly, in order to prevent the fluctuation of the leading data of big numerical value,
Fractional value is ignored, and needs the scaling for carrying out data, it is also possible to avoid the difficulty for calculating.Simple linear scale can be adopted,
Eigenvalue is zoomed on [- 1,1].
(2.9) feature extraction terminates.The 12 dimensional feature vector X for extracting are obtained,.
Step 3：Choose optimal feature and parameter r, c：Using support vector machine technology and gaussian kernel function k (x1,x2)=
exp(-r||x1-x2||2).Support vector machine training Data Source be：(X,(i),Y(i)), i=1, wherein 2 ... N, X,(i)It is
The characteristic vector that sample i is extracted through step 2 in training set, Y(i)If=+ 1 sample i is positive sample, Y(i)If=- 1 sample i
It is negative sample.The selection of parameter chooses optimal parameter r, c using the method for trellis traversal.The selection of feature is using such as lower section
Method：The selection of first feature：Each feature is trained, accuracy rate highest feature is chosen.The selection of second feature：
Remaining feature and the combinations of features of first selection are trained, and choose accuracy rate highest feature.Gone down with this, Zhi Daozhun
Till really rate reaches 95%, stop feature selection.
Step 4：Training Support Vector Machines model.The data of training are：(X(i),Y(i)), i=1,2 ... N, the ginseng of training
Number be step 3 choose optimal parameter r, c.Using these data and parameter training supporting vector machine model, and preserve.
Step 5：Prediction judges.The road image that photographic head shoots is extracted, the image to extracting enters according to the method for step 2
Row feature extraction, feature X selected by selecting step 3, the supporting vector machine model preserved using step 4 is predicted to be sentenced
Disconnected, if output y=+1, road is clean, otherwise exports y=-1, represents road unclean.It is clean so as to reach road pavement
The purpose of degree detection.
The invention has the beneficial effects as follows：The present invention is by principal component analysiss dimensionality reduction and seeks the method for taking most pure road block, carries
By way of road speed faster, precision is bigger, so that classification results are more accurate.This method extracts altogether 12 features, different
Construction road can automatically select optimal characteristics and parameter, it is portable strong so as to be suitable for most of construction road.This
The accuracy rate of invention is very high, cleans road in time, prevents because roadway pollution causes the reality of peril generation with work side is instructed
Border meaning.
Description of the drawings
Fig. 1 is construction road contamination detection method flowchart of the present invention；
Fig. 2 is the flow chart that the present invention carries out feature extraction to road picture.
Specific embodiment
As shown in figure 1, the present invention is comprised the steps based on the construction road contamination detection method of support vector machine：
Step 1：Collection road image, using the image of collection as training set, and artificial is not divided into training set totally and not
Clean two set.Image positioned at clean set is referred to as positive sample, and positioned at the image totally do not gathered negative sample is referred to as, and trains
The number for concentrating sample is N.
Step 2：Feature extraction is carried out to road image.Following sub-step is included as shown in Figure 2：
(2.1) simple division road and non-rice habitats part.Because photographic head is fixed, every pictures can be unified
Remove non-rice habitats part, such as greenbelt etc..The number of road area is represented with n；The square of a m rows n row is defined with A [m, n]
Battle array A；A (y, x) represents the value of the y rows xth row in matrix A；The row k of A (k) representing matrix A, is a row vector；ATTable
Show the transposed matrix of A.
(2.2) color space conversion.Native color spatial transformation to LUV color spaces, LUV color model are more met people couple
The sensation of color, makes up the deficiency of RGB color model.The LUV color value of road pixel, LUVs (k) tables are recorded with LUVs [n, 3]
Show the LUV color value of k-th road pixel point of correspondence.
(2.3) principal component analysiss reduce dimension.The color space of 3-dimensional is dropped to 2 dimensions.Adopt with the following method：To LUVsT*
LUVs carries out eigenvalue eigendecomposition using SVD methods, that is, have：LUVsT* LUVs=V*2*VT, PC [n, 2] is LUVs*V
First two columns, i.e. the main constituent of road color, with i-th row of PC [i] representing matrix PC.
(2.4) Meanshift cluster segmentations.A matrix S [256,256] is defined first,Wherein 1 { true }=1,1 { false }=0.Do on matrix S
Meanshift is split, and rectangular window of the size for nW*nW is specified first, and nW is determined by the size of detection zone, is not less than
2, method is as follows：
(2.4.1) non-zero entry A is looked in S, if A is not present, step (2.4.4) is jumped to.
(2.4.2) in S, it is the barycenter B in nW*nW rectangular windows to calculate the size centered on A.
If (2.4.3) A equals B, central point A values are recorded, all elements centered on A in window all set in S
It is zero, returns to step (2.4.1).If A is not equal to B, B values are assigned to A, return to step (2.4.2).
(2.4.4) K central point Centers [K, 2] is obtained, it is Matrix C enters [K, 2] i-th to define Centers [i]
OK；Define one-dimensional vector Tags [n], wherein Tags [i]=argjMin (| | PC [i]-center [j] | |) (i=0,
1 ... n-1, j=0,1 ... ... k-1), Tags [n] is the class label for recording road pixel；It has been divided into K class, similar unit
Have identical class label.
(2.5) road class is chosen.First vegetarian refreshments of each road sections corresponds to a LUV color and class label, records respectively
In LUVs and Tags, the road waypoint for really arriving need to be found, remove these pixels of car pedestrian, that is, find the class label of real road
tag.Follow two principles：The area of road is big relatively to be unified with color, that is, more than number and the characteristics of color variance is little.To appointing
The class label i of meaning, has：Wherein
The number of class i is calculated, var (i) calculates the variance of class i.
(2.6) most pure path matrix block is found.A two-dimensional matrix IR [iH, iW] is defined, iH represents the height of image,
IW is the width of image, and the y rows X of IR (y, x) representing matrix IR is arranged.IR (y, x)=1 represents that pixel P (x, y) is the
Road sections that one step is retained and its corresponding class label is tag in the 5th step, IS (y, x)=0 represents other situations.
Take a matrix window (nH, nW), most pure path matrix window be then in this window IR be 1 number it is most.
(2.7) color and textural characteristics are extracted.Due to the uniformity of road, color can be extracted in the matrix-block for finding
With textural characteristics.The color characteristic of extraction has 3 components of LUV, does average to the value of component respectivelyVarianceSkewnessxiThe value in 3 components in for LUV.Using ash
The textural characteristics that degree symbiosis is extracted obtain gray level co-occurrence matrixes p [i, j], calculate energy ASM=∑i∑j p(i,j)2, contrast
CON=∑sn n2(∑| i-j |=nP (i, j)), degree of associationCome to 12 features.
(2.8) scaling of feature.The characteristic of extraction differs greatly, in order to prevent the fluctuation of the leading data of big numerical value,
Fractional value is ignored, and needs the scaling for carrying out data, it is also possible to avoid the difficulty for calculating.Simple linear scale can be adopted,
Eigenvalue is zoomed on [- 1,1].
(2.9) feature extraction terminates.The 12 dimensional feature vector X for extracting are obtained,.
Step 3：Choose optimal feature and parameter r, c：Using support vector machine technology and gaussian kernel function k (x1,x2)=
exp(-r||x1-x2||2).Support vector machine training Data Source be：(X,(i),Y(i)), i=1, wherein 2 ... N, X,(i)It is
The characteristic vector that sample i is extracted through step 2 in training set, Y(i)If=+ 1 sample i is positive sample, Y(i)If=- 1 sample i
It is negative sample.The selection of parameter chooses optimal parameter r, c using the method for trellis traversal.The selection of feature is using such as lower section
Method：
The selection of first feature：Each feature is trained, accuracy rate highest feature is chosen.Second feature
Select：Remaining feature and the combinations of features of first selection are trained, and choose accuracy rate highest feature.Gone down with this,
Until rate of accuracy reached is to satisfied, stop feature selection.
Step 4：Training Support Vector Machines model.The data of training are：(X(i),Y(i)), i=1,2 ... N, the ginseng of training
Number be step 3 choose optimal parameter r, c.Using these data and parameter training supporting vector machine model, and preserve.
Step 5：Prediction judges.The road image that photographic head shoots is extracted, the image to extracting enters according to the method for step 2
Row feature extraction, feature X selected by selecting step 3, the supporting vector machine model preserved using step 4 is predicted to be sentenced
Disconnected, if output y=+1, road is clean, otherwise exports y=-1, represents road unclean.It is clean so as to reach road pavement
The purpose of degree detection.
Claims (1)
1. a kind of construction road contamination detection method based on support vector machine, it is characterised in that the method comprises the steps：
Step 1：Collection road image, using the image of collection as training set, and it is artificial training set is divided into it is clean and unclean
Two set；Image positioned at clean set is referred to as positive sample, is referred to as negative sample positioned at the image totally do not gathered, in training set
The number of sample is N；
Step 2：Feature extraction is carried out to road image, comprising following sub-step：
(2.1) simple division road and non-rice habitats part：Because photographic head is fixed, removing that can be unified to every pictures
Non-rice habitats part；The number of road area is represented with n；The matrix A of a m rows n row is defined with A [m, n]；A (y, x) is represented in square
The value of y rows xth row in battle array A；The row k of A (k) representing matrix A, is a row vector；ATRepresent the transposed matrix of A；
(2.2) color space conversion：Native color spatial transformation to LUV color spaces, with LUVs [n, 3] road pixel is recorded
LUV color value, LUVs (k) represents the LUV color value of k-th road pixel point of correspondence；
(2.3) principal component analysiss reduce dimension, and the color space of 3-dimensional is dropped to 2 dimensions；Adopt with the following method：To LUVsT* LUVs is adopted
Eigenvalue eigendecomposition is carried out with SVD methods, that is, is had：LUVsT* LUVs=V*2*VT, PC [n, 2] is front the two of LUVs*V
Row, the i.e. main constituent of road color, with i-th row of PC [i] representing matrix PC；
(2.4) Meanshift cluster segmentations：A matrix S [256,256] is defined first,Wherein 1 { true }=1,1 { false }=0；Do on matrix S
Meanshift is split, and rectangular window of the size for nW*nW is specified first, and nW is determined by the size of detection zone, is not less than
2, method is as follows：
(2.4.1) non-zero entry A is looked in S, if A is not present, step (2.4.4) is jumped to；
(2.4.2) in S, it is the barycenter B in nW*nW rectangular windows to calculate the size centered on A；
If (2.4.3) A equals B, central point A values are recorded, all elements centered on A in window are all set to zero in S,
Return to step (2.4.1)；If A is not equal to B, B values are assigned to A, return to step (2.4.2)；
(2.4.4) K central point Centers [K, 2] is obtained, it is Matrix C enters [K, 2] the i-th row to define Centers [i]；It is fixed
Adopted one-dimensional vector Tags [n], wherein Tags [i]=argjMin (| | PC [i]-center [j] | |) (i=0,1 ... n-
1, j=0,1 ... ... k-1), Tags [n] is the class label for recording road pixel；K class is divided into, has had phase with dvielement
Same class label；
(2.5) road class is chosen：First vegetarian refreshments of each road sections corresponds to a LUV color and class label, is separately recorded in
In LUVs and Tags, the road waypoint for really arriving need to be found, remove these pixels of car pedestrian, that is, find the class label of real road
tag；Follow two principles：The area of road is big relatively to be unified with color, that is, more than number and the characteristics of color variance is little；It is right
Arbitrary class label i, has：WhereinMeter
The number of class i is calculated, var (i) calculates the variance of class i；
(2.6) most pure path matrix block is found：A two-dimensional matrix IR [iH, iW] is defined, iH represents the height of image, and iW is
The width of image, the y rows X row of IR (y, x) representing matrix IR；IR (y, x)=1 represents that pixel P (x, y) is in step
2.1 road sections for retaining and its corresponding class label is tag in step 2.5, IS (y, x)=0 represents other situations；
Take a matrix window (nH, nW), most pure path matrix window be then in this window IR be 1 number it is most；
(2.7) color and textural characteristics are extracted：Due to the uniformity of road, color and stricture of vagina can be extracted in the matrix-block for finding
Reason feature；The color characteristic of extraction has 3 components of LUV, does average to the value of component respectivelyVarianceSkewnessxiThe value in 3 components in for LUV；Using ash
The textural characteristics that degree symbiosis is extracted obtain gray level co-occurrence matrixes p [i, j], calculate energy ASM=∑i∑jp(i,j)2, contrast
CON=∑snn2(∑| i-j |=nP (i, j)), degree of associationCome to 12 features；
(2.8) scaling of feature：The characteristic of extraction differs greatly, in order to prevent the fluctuation of the leading data of big numerical value, decimal
Value is ignored, and needs the scaling for carrying out data, it is also possible to avoid the difficulty for calculating；Simple linear scale can be adopted, feature
Value is zoomed on [- 1,1]；
(2.9) feature extraction terminates；Obtain the 12 dimensional feature vector X' for extracting；
Step 3：Choose optimal feature and parameter r, c：Using support vector machine technology and gaussian kernel function k (x1,x2)=exp
(-r||x1-x2||2) support vector machine training Data Source be：(X'(i),Y(i)), i=1, wherein 2 ... N, X'(i)It is training
The characteristic vector for concentrating sample i to be extracted through step 2, Y(i)If=+ 1 sample i is positive sample, Y(i)If=- 1 sample i is negative
Sample；The selection of parameter chooses optimal parameter r, c using the method for trellis traversal；The selection of feature is adopted with the following method：The
The selection of one feature：Each feature is trained, accuracy rate highest feature is chosen；The selection of second feature：It is remaining
Feature and first selection combinations of features be trained, choose accuracy rate highest feature；Gone down with this, until accuracy rate
Till reaching satisfaction, stop feature selection；
Step 4：Training Support Vector Machines model：The data of training are：(X(i),Y(i)), i=1,2 ... N, the parameter of training is
Optimal parameter r that step 3 is chosen, c；Using these data and parameter training supporting vector machine model, and preserve；
Step 5：Prediction judges：The road image that photographic head shoots is extracted, the image to extracting carries out spy according to the method for step 2
Extraction is levied, feature X selected by selecting step 3, the supporting vector machine model preserved using step 4 is predicted judgement,
If output y=+1, road is clean, otherwise exports y=-1, represents road unclean；So as to reach road pavement cleanliness factor
The purpose of detection.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410026324.2A CN103778437B (en) | 2014-01-21 | 2014-01-21 | Construction road pollution detection method based on support vector machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410026324.2A CN103778437B (en) | 2014-01-21 | 2014-01-21 | Construction road pollution detection method based on support vector machine |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103778437A CN103778437A (en) | 2014-05-07 |
| CN103778437B true CN103778437B (en) | 2017-04-12 |
Family
ID=50570650
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410026324.2A Active CN103778437B (en) | 2014-01-21 | 2014-01-21 | Construction road pollution detection method based on support vector machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103778437B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107545254A (en) * | 2017-09-14 | 2018-01-05 | 湖南酷陆网络科技有限公司 | The detection method and device of road surface cleanliness factor |
| CN110688945A (en) * | 2019-09-26 | 2020-01-14 | 成都睿云物联科技有限公司 | Cleanliness detection method and device, computer equipment and storage medium |
| CN111242223B (en) * | 2020-01-15 | 2020-11-13 | 中国科学院地理科学与资源研究所 | Street space quality evaluation method based on streetscape image multi-feature fusion |
| CN116884226B (en) * | 2023-09-07 | 2023-11-21 | 山东金宇信息科技集团有限公司 | Ecological monitoring and early warning method, equipment and medium for road maintenance |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102270301A (en) * | 2011-06-07 | 2011-12-07 | 南京理工大学 | Method for detecting unstructured road boundary by combining support vector machine (SVM) and laser radar |
| CN103077501A (en) * | 2012-12-30 | 2013-05-01 | 信帧电子技术（北京）有限公司 | Road target marking method and device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP5083658B2 (en) * | 2008-03-26 | 2012-11-28 | 本田技研工業株式会社 | Vehicle lane recognition device, vehicle, and vehicle lane recognition program |
-
2014
- 2014-01-21 CN CN201410026324.2A patent/CN103778437B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102270301A (en) * | 2011-06-07 | 2011-12-07 | 南京理工大学 | Method for detecting unstructured road boundary by combining support vector machine (SVM) and laser radar |
| CN103077501A (en) * | 2012-12-30 | 2013-05-01 | 信帧电子技术（北京）有限公司 | Road target marking method and device |
Non-Patent Citations (3)
| Title |
|---|
| "Detection algorithm for road surface condition using wavelet packet transform and SVM";Hun-Jun Yang .etc;《Frontiers of Computer VIsion (FCV)》;20130201;全文 * |
| "Multi-Bernoulli based track-before-detect with road constraints";Wong Jefrey .etc;《Information Fusion(FUSION),2012 15th International Conference on》;20120731;全文 * |
| "利用双窗口检测和粒子滤波提取高分辨率SAR图像道路中心点";程江华，等;《电子与信息学报》;20120531;第34卷(第5期);1109-1114 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103778437A (en) | 2014-05-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | Megadepth: Learning single-view depth prediction from internet photos | |
| Xu et al. | Segment as points for efficient online multi-object tracking and segmentation | |
| CN105022982B (en) | Hand motion recognition method and apparatus | |
| Wang et al. | Dairy goat detection based on Faster R-CNN from surveillance video | |
| CN104240264B (en) | The height detection method and device of a kind of moving object | |
| CN104866616B (en) | Monitor video Target Searching Method | |
| CN103440640B (en) | A kind of video scene cluster and browsing method | |
| CN103778437B (en) | Construction road pollution detection method based on support vector machine | |
| CN107679503A (en) | A kind of crowd's counting algorithm based on deep learning | |
| CN110532970B (en) | Age and gender attribute analysis method, system, equipment and medium for 2D images of human faces | |
| CN103530638B (en) | Method for pedestrian matching under multi-cam | |
| CN101833664A (en) | Video image character detecting method based on sparse expression | |
| CN104715251B (en) | A kind of well-marked target detection method based on histogram linear fit | |
| CN103336835B (en) | Image retrieval method based on weight color-sift characteristic dictionary | |
| Cui et al. | Convolutional neural network for recognizing highway traffic congestion | |
| CN112766123B (en) | Crowd counting method and system based on criss-cross attention network | |
| CN108198202A (en) | A kind of video content detection method based on light stream and neural network | |
| CN113139489A (en) | Crowd counting method and system based on background extraction and multi-scale fusion network | |
| CN111242096B (en) | People number gradient-based people group distinguishing method | |
| CN109034258A (en) | Weakly supervised object detection method based on certain objects pixel gradient figure | |
| Hsu et al. | Industrial smoke detection and visualization | |
| CN112069985A (en) | High-resolution field image rice ear detection and counting method based on deep learning | |
| CN115393587A (en) | Expressway asphalt pavement disease sensing method based on fusion convolutional neural network | |
| Marvaniya et al. | Small, sparse, but substantial: techniques for segmenting small agricultural fields using sparse ground data | |
| CN106611030A (en) | Object similarity comparison method and object search method based on video, and object similarity comparison system and object search system based on video |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |