CN106558051A - A kind of improved method for detecting road from single image - Google Patents
A kind of improved method for detecting road from single image Download PDFInfo
- Publication number
- CN106558051A CN106558051A CN201510622576.6A CN201510622576A CN106558051A CN 106558051 A CN106558051 A CN 106558051A CN 201510622576 A CN201510622576 A CN 201510622576A CN 106558051 A CN106558051 A CN 106558051A
- Authority
- CN
- China
- Prior art keywords
- image
- road
- pixel
- region
- rice habitats
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06T2207/30248—Vehicle exterior or interior
- G06T2207/30252—Vehicle exterior; Vicinity of vehicle
- G06T2207/30256—Lane; Road marking
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a kind of method based on image detection road.The present invention has the feature that analogous color is distributed to detect road end point according to there is pixel on the preferable straight line of orientation consistency and pavement of road in Ordinary Rd image.Detection road end point is that realizing, the region between this two cross linears can cover point road area in the middle part of image, can also cover sometimes part non-rice habitats region in image by meeting the intersection point of the two of above-mentioned condition cross linears.Using this initial segmentation result, Gaussian distribution model is set up for road area the characteristics of relatively concentration for color in road area and textural characteristics distribution, and be directed to and set up nonparametric distribution for non-rice habitats region in image plus be uniformly distributed model the characteristics of color and textural characteristics distribution relatively disperse in non-rice habitats region, the initial road part incidentally detected during end point is detected then can be corrected using a simple likelihood score ratio grader.
Description
Technical field
The present invention relates to it is a kind of it is improved detection road method, more particularly, to one kind it is bottom-up from
Single image detects the full-automatic method of road.
Background technology
Real-time Road is monitored, in the computer vision system such as Intelligent walking robot or automatic driving vehicle
The major issue for facing is exactly the feasible road in detection front and the barrier in detection road etc..
In the case where not having any priori about road to help, from single image detection road it is mathematically
One ill-conditioning problem.
Many counting methods are for the image for obtaining is shot under artificial environment, such as indoor and city street at present
Road image etc., carries out Road Detection.Because these images meet " the Manhattan world " (Manhattan World)
It is assumed that the direction of most straight lines and seat mutually orthogonal in real three-dimensional world coordinate system i.e. in image
Parameter is consistent, thus by these lines according to different directions carry out cluster can help computer estimate three
The spatial structural form of dimension scene.
But the image for obtaining is shot in unartificial environment and the image for obtaining is shot under artificial environment
There is very big otherness, such as suburb, desert and wilderness etc., the majority detected from these images
Straight line is produced by cloud etc. in trees, vegetation, distant place mountain range and sky, and these rectilinear directions with
Mutually orthogonal coordinate system do not correspond in real three-dimensional world, thus these straight lines are for estimating image
Middle road can cause very big interference.
From the point of view of according to the technical literature consulted at present, in the existing lane detection technology based on image
One class be published in in August, 2010 IEEE image procossing periodicals paper (H.Kong, J.Y.Audibert,
and J.Ponce,“General road detection from a single image,”IEEE Trans.On
Image Processing,19(8):2211-2220,2010) have been described.Such technology is empty using one group
Between anisotropic filter array, such as Gabor filter, the grain direction of each pixel in detection image,
Then temporal voting strategy is adopted, the end point that the most pixel of poll is confirmed as road is obtained.In inspection
On the basis of surveying road end point, then estimate edge straight line on both sides of the road.The limitation of this kind of method exists
It is high in voting process computation complexity, it is not suitable for real-time system；If existed in road scene in addition
Some objects with higher edge strength compared with road boundary or track of vehicle, such as billboard,
Or vehicle etc., then voting results occur very big deviation, because the qualification of polling place is typically by pixel
On point, the intensity of anisotropic filter response is determined；The detection method of straight line is based only in addition for bending
The very big road of degree is not simultaneously applied to.
The content of the invention
One aspect of the present invention provides a kind of method based on image detection road, and the method includes following
Step：Input road image；Road area or non-rice habitats region are belonged to based on each pixel in image
The likelihood score anticipation pixel belong to road area or non-rice habitats region；Determine the road image
In skyline；According to the relative position relation amendment pair of each pixel in image and the skyline
Belong to the anticipation in road area or non-rice habitats region in the pixel.
Wherein the anticipation that each pixel in image belongs to road area or non-rice habitats region can be passed through
Following manner is carried out：Segment the image into initial road region and initial non-rice habitats region；Determine image
In each pixel belong to and the likelihood score of road area and belong to the likelihood score in non-rice habitats region；Then will
Each pixel belongs to the likelihood score of road area and belongs to the ratio and one of the likelihood score in non-rice habitats region
Threshold value compares, and whether the pixel belongs to road area according to comparative result anticipation.Wherein initial road
Road region and dividing for initial non-rice habitats region can adopt easy triangle initial road region division
Method or semicircle initial road region division methods.The division in triangle initial road region
Method be by first determining road image midpoint, then by road image midpoint and image lower-left boundary point and
Estimation of the Delta Region that image bottom right boundary point is constituted as initial road region, and in image except
The estimation then as initial non-rice habitats region of other regions outside Delta Region.Semicircle initial road area
The division methods in domain determine a half-circle area being located inside image then with image baseline midpoint as the center of circle
Used as the estimation in initial road region, and other regions in image in addition to half-circle area are then as just
The estimation in beginning non-rice habitats region.
The method can also utilize the straight line information in Ordinary Rd image, with reference to positioned at pixel on straight line
The gradient direction of point answers consistent feature, uncorrelated to road or track of vehicle in removal image
Edge line.Then using least square method these image border fittings are obtained with the side of line correspondence
Journey.And according to road boundary in Ordinary Rd image be from road end point toward image base extend, road
Texture color inside road should have similitude or uniformity, start court from intersection point per two intersecting straight lines
The otherness of pixel texture color into the wrapped up region of line segment of image base, and often intersecting two
The gradient direction of pixel and rectilinear direction homogeneity measure on bar straight line, construct an energy function,
And the road end point for being defined as detecting by the intersection point of the cross linear pair with least energy.For
Cross linear with least energy is to starting to lead between the two lines section of image base from end point
Region can cover point road area in the middle part of image, cover part non-rice habitats region in image, profit sometimes
With this initial rough lane segmentation result, as in Ordinary Rd region, COLOR COMPOSITION THROUGH DISTRIBUTION relatively collects
In, rather than COLOR COMPOSITION THROUGH DISTRIBUTION is relatively disperseed in road area, using different statistical distribution pattern point
Not Wei road area and non-rice habitats region set up COLOR COMPOSITION THROUGH DISTRIBUTION statistical model, it is then simple using one
Likelihood score ratio grader can correct the initial road part incidentally detected during detection end point.
For certain types of road image, the detection method based on end point may not be accurately detected
Initial road region.For example for the image of tee T road, or vehicle is just turning left or is turning right
When in the road image that photographs, it may be possible to road end point cannot be detected.Such
In the case of, then no longer it is suitable for the detection method based on end point.
In practical application, the Approach for road detection of the present invention can be realized using following concrete steps：
(1) it is input into a width road image；
(2) set a length threshold T1, angular deviation threshold value T2, and fitting a straight line figure will be used for
As the number N at edge1, the numerical value T that length threshold is gradually reduced3, and meet orientation consistency requirement
The number N of image border2；
(3) using method for detecting image edge obtain in road image the gradient direction of each pixel and
Marginal point；
(4) marginal point isolated in image, connection are removed using grating scanning mode and eight connectivity method
Marginal point, obtains image border；
(5) check whether the length of all image borders that previous step (4) is obtained meets step (2)
The length threshold T for setting in advance1If, in current length threshold value T1Under, can collect for straight line
Fitting N1Bar meets the image border of length requirement, then jump to next step (6), otherwise then gradually
Reduce length threshold T1, every time from current length threshold value T1Reduce T3, until N can be collected1Bar meets length
The image border that degree is required；
(6) N for obtaining1Bar meets the image border of length requirement, and calculating meets length per bar will
The conforming measure value of pixel gradient direction on the image border asked：
Wherein pkAnd pk+1It is two neighbor pixels met on length requirement image border positioned at i-th,WithIt is the gradient direction on the two pixels respectively, symbol | | absolute value, C are asked in expressioni
The conforming measure value of pixel gradient direction on exactly i-th image border for meeting length requirement, this
Individual measure value is less to show that edge gradient orientation consistency is better；
(7) according to formula (1), by N1Bar meets the image border of length requirement according to each calculating
The conforming measure value of pixel gradient direction arranged from small to large, then select from small to large
N2The preferable image border of bar orientation consistency, i.e. their orientation consistency measure value C are smaller；
(8) each figure for meeting length and orientation consistency requirement for obtaining from step (7)
As edge, it is assumed that i-th meets total n on length and orientation consistency requirement image borderiIt is individual, this
A little points constitute following homogeneous coordinates matrix Ai：
Wherein in formula (2), x and y to be represented and meet length and direction positioned at i-th
The space coordinates of pixel on coherence request image border, subscript i represent that to meet length consistent with direction
Property require the sequence number of image border, and subscript then represents the sequence number of the pixel on the image border, model
Enclose from 1 to ni, these matrix As for obtaining are decomposed using singular value decomposition method theni, these matrixes
Number be equal to the number N of the image border for meeting length and orientation consistency requirement2, matrix AiMost
Unit singular vectors corresponding to little singular value are exactly to meet i-th figure of length and orientation consistency requirement
As the straight line parameter vector [a of edge fittingi bi ci], the linear equation is aix+biy+ci=0, change sentence
It is the straight line that these image border fittings are obtained using least square method to talk about；
(9) parameter according to resulting straight line, calculates the intersection point between these straight lines, it is assumed that i-th
The straight line parameter vector obtained with the fitting of j-th strip image border is respectively [ai bi ci] and [aj bj cj],
The intersection point of i-th and j-th strip image border fitting a straight line is obtained by straight line parameter vector cross-products；
(10) for per two intersecting straight lines, starting from their intersection point past along this two straight lines
Below image direction check on this two straight lines the gradient direction of each pixel and place rectilinear direction it
Between difference：
S represents straight line liWith straight line ljIntersection point, whereinRepresent straight line liStart along straight line from intersection point s
Towards that section below image, pkRepresent positioned at line segmentOn a pixel, andRepresent
Gradient direction on the pixel, θ (li) represent straight line liDirection (i.e.), T2
It is exactly the angular deviation threshold value arranged in step (2), # is a counter, calculating is located atOn have many
The gradient direction and rectilinear direction θ (l of few pixeli) between absolute value difference be less than angular deviation threshold value
T2,Represent straight line liStart the length of that section below along straight line towards image from intersection point s,
In formula (3)Value shows more greatly straightwayThe side of the gradient direction and place straight line of upper pixel
To more consistent；
(11) in Ordinary Rd image, road marking line or track of vehicle are all end point courts from afar
To below image, without towards above image, so the scope in image cathetus direction should be
[0 °, 180 °), wherein 180 ° of direction straight lines and 0 ° of direction straight line belong to horizontal line together, straight line vertically downward should
This is 90 °, is considered per two towards the wrapped up intra-zone texture color of intersecting straight lines below image now
Between otherness：
Straight line liWith straight line ljS points are intersected at, whereinWithStraight line l is represented respectivelyiWith straight line ljFrom intersection point s
Start along respective straight line towards that section below image, the image-region between intersecting straight lines is led to
Cross intersection point s, and respectively straight line liWith straight line ljBetween the middle separated time of angle be divided into two parts：Ω1And Ω2,
Formula (4) is exactly the otherness for calculating the texture color between this two parts, # (Ω1) and # (Ω2) represent
Respectively fall in Ω1And Ω2Pixel number in region, and p1And p2Represent in Ω respectively1And Ω2Interior
Pixel, c (p1) and c (p2) color on these aspects is represented, symbol | | | | represent zoning Ω1And Ω2It
Between color mean difference norm；
(12) according to formula (3) and (4), construction intersects at s point straightwaysWithEnergy function：
The energy function per two intersecting straight lines is checked, the intersection point with least energy straight line pair is exactly to detect
Road end point；
(13) above-mentioned two intersecting straight lines for detecting road end point can segment the image into initial road
Road region (being sandwiched in the region between this two intersecting straight lines) and initial non-rice habitats region (are removed in image
Remove the remainder in initial road region)；
(14) for initial road region, calculate the sample mean positioned at picture element point chromatic in this region
The sample covariance matrix Σ of vector value μ (including tri- numerical value of RGB) and color, obtains positioned at this
The Gauss model of picture element point chromatic distribution in region, in image pixel belong to the probability of road area can
To be calculated by following formula：
Wherein x represents the color-values of certain pixel in image, and | Σ | is represented and sought color covariance matrix Σ
Determinant, ΨRExpression belongs to the mark of road, and p (x | ΨR) represent that the pixel belongs to road area
Likelihood score；In another embodiment, the color space of image can be transformed into into Lab from RGB first
Color space, and each pixel neighborhood of a point texture eigenvalue is calculated using a sliding window, then
For initial road region, the mean value for being located at pixel visual signature in the region is calculated (including Lab
Three numerical value and texture), and the covariance matrix Σ of visual signature, at this moment calculate each pixel
In putting the formula (6) for belong to road area probability, x and μ will be 4 n dimensional vector ns, and Σ will be one
4 take advantage of 4 matrixes, constant coefficient(2 π) will be changed into-2；In another embodiment, in formula (6)
Visual feature vector can using 7 dimensional vectors, i.e. each pixel comprising tri- color values of RGB,
Three numerical value of Lab color spaces and textural characteristics (SILTP) numerical value.So x and μ will be
One 7 n dimensional vector n, and Σ will be one 7 and take advantage of 7 matrixes, constant coefficientTo be changed into
(15) for initial non-rice habitats region, all pixels point positioned at this region by initial markers is
Non-rice habitats, these pixels constitute a sample set ΨN={ y1 y2 … ym, wherein yiRepresent pixel
The color-values of point (include tri- numerical value of RGB, 1≤i≤m, m are represented positioned at initial non-rice habitats region picture
The sum of vegetarian refreshments), in image, pixel is belonged to the probability in non-rice habitats region and can be calculated by Density Estimator
Son is obtained：
Wherein H is a symmetric positive definite matrix,Be one it is symmetrical, compact, desired value is zero, covariance
Core distribution function of the matrix for unit matrix, its bandwidth (or variance) value can adopt adaptively selected
Method is obtained, ΨNExpression belongs to off-highroad mark, and p (x | ΨN) represent that this pixel belongs to non-road
The likelihood score in road region；In another embodiment, can be first by the color space of image from RGB
Lab color spaces are transformed into, and it is special each pixel neighborhood of a point texture to be calculated using a sliding window
Value indicative, for initial non-rice habitats region, all pixels point positioned at this region constitutes a sample set
ΨN={ y1 y2 … ym, wherein yiRepresent pixel visual signature (include tri- numerical value of Lab and
Texture eigenvalue) and using the method for cuclear density estimation, according to the visual signature (bag of each pixel
Include tri- numerical value of Lab and texture), each pixel is calculated according to formula (7) belong to non-rice habitats region
Probability；It is similar with step (14), in the calculating to initial non-rice habitats region part, each picture
Vegetarian refreshments can also include tri- color values of RGB, three numerical value of Lab color spaces and a line
Reason feature (SILTP) numerical value, and in view of the dependence between surrounding pixel, it is additionally added space bit confidence
Breath, the i.e. joint of the coordinate (x, y) of each pixel, this view-based access control model feature and spatial positional information
Nonparametric model be avoided that it is isolated consider each pixel, but the neighborhood of each pixel be also contemplated for into
Come；
For initial non-rice habitats region, the road image that initial detecting goes out is not ensured that and real road
Region fits like a glove, in order to strengthen the robustness (Robustness) of the relative extreme case of road model,
The model in non-rice habitats region can be based on nonparametric distributed model, also can be in the base of nonparametric distributed model
Increase by one on plinth and be uniformly distributed model.This is uniformly distributed model can be independently of the position of assessment pixel
And visual characteristic.
It is as follows based on nonparametric distributed model and the formula for being uniformly distributed model：
Wherein, α is hybrid weight, α<<1, γ is one obeys equally distributed stochastic variable.
(16) threshold tau is set0, using a simple likelihood score ratio grader by picture in image
Vegetarian refreshments is divided into and belongs to road and belong to two kinds of marks of non-rice habitats：
T=ln (p (x | ΨR))-ln(p(x|ΨN)) (9)
As general likelihood score adopts logarithmic, in formula (9), ln is represented and is taken natural logrithm, if public
Value t that formula (9) is calculated is more than the threshold tau for arranging0, then judge that current pixel point belongs to road area,
Otherwise then belong to non-rice habitats region.
(17) determine the skyline (or referred to as horizon) in road image, on amendment skyline
The pixel of road area.Because road is typically below skyline, by determining skyline
Position, can correct on skyline by error flag for road pixel pixel.Horizon
Between two regions of the input picture that the height of line can be separated by a hypothesis skyline by detection
Between group, the maximum of non-similarity (interclass dissimilarity) is being estimated.
The detection of skyline can be carried out as feature space based on characteristics of image such as colors.Additionally,
In order to accelerate detection speed, input picture can carry out size compression by sampling.The side of compression of images
Method can also first pass through a low pass filter, then be sampled.
During detection skyline, the group of part is descended to calculating part on which in image per one-row pixels therewith
Between non-similarity, computing formula is as follows：
Wherein,It is the mean vector of the HSV color spaces of upper part,It is bottom
The mean vector of the HSV color spaces for dividing,It is the variance of three elements of part on HSV,It is the variance of three elements of part under HSV.It is non-similar between maximum group in result of calculation
The skyline that the row that degree is located as determines.It is in a relatively small size due to calculating skyline
Carry out on image, finally also need to horizon line position is reverted to by interpolation method and be originally inputted figure
As in.
As it was previously stated, based on road end point being divided into initial road region and initial in said method
Non-rice habitats region step can also be estimated with triangle initial road region or semicircle initial road region
The additive methods such as method replace.
Another aspect of the present invention additionally provides a kind of system based on image detection road, and the system includes
Receive the equipment of road image input；Road area or non-rice habitats are belonged to based on each pixel in image
The likelihood score anticipation pixel in region belongs to the equipment in road area or non-rice habitats region；Determine institute
State the equipment of the skyline in road image；According to the phase of each pixel in image and the skyline
Belong to what the anticipation in road area or non-rice habitats region was modified to position relationship to the pixel
Equipment.Wherein belonging to road area or non-rice habitats region to each pixel in image carries out anticipation and sets
It is standby to include：The equipment in initial road region and initial non-rice habitats region is segmented the image into, really
Determine each pixel in image to belong to the likelihood score of road area and belong to the likelihood score in non-rice habitats region
Equipment, and each pixel is belonged to into the likelihood score of road area and belongs to the likelihood score in non-rice habitats region
Ratio compare to determine whether the pixel belongs to the equipment of road area with a threshold value.
Another aspect of the present invention additionally provides a kind of device based on image detection road, and the device includes
One storage device and a processor.The processor can be used for performing following operation：Input mileage chart
Picture；The likelihood score anticipation picture in road area or non-rice habitats region is belonged to based on each pixel in image
Vegetarian refreshments belongs to road area or non-rice habitats region；Determine the skyline in the road image；According to
The relative position relation amendment of each pixel and the skyline in image belongs to for the pixel
Road region or the anticipation in non-rice habitats region.
Present invention also offers the guider of a kind of vehicle or other movable equipments, the guider
Including the above-mentioned device based on image detection road.
Present invention also offers a kind of computer readable storage devices, the computer readable storage devices bag
Computer code is included, the computer when code is run on a computer, can be caused to perform following operation：
Input road image；The likelihood in road area or non-rice habitats region is belonged to based on each pixel in image
The degree anticipation pixel belongs to road area or non-rice habitats region；Determine the day in the road image
Border line；According to the relative position relation amendment of each pixel in image and the skyline for the picture
Vegetarian refreshments belongs to the anticipation in road area or non-rice habitats region.
The above-mentioned technical proposal of the present invention does not need any priori about road, it is not required that adopt
Use any machine learning method, be it is a kind of it is bottom-up from single image can just detect road it is complete from
Dynamic method.Although highlighted in the application to depend on straightway to detect road end point, it is based on
Initial road segmentation result, can correct the road area of initial segmentation by visual signature statistical analysis
Technology, this method may also be employed the technology of other appropriate initial roads segmentations so as to suitable for towards curved
The very big road image of curvature.And this method can also detect the barrier in road, method
Computation complexity is low, and speed is fast, with the prospect for promoting the use of real-time system.
Description of the drawings
Figure 1A and 1B are the flow charts of the present invention；
Fig. 2A and 2B illustrates the schematic diagram of Road Detection result of the present invention based on parts of images；
Fig. 3 A and 3B are the methods in determination initial road region and initial non-rice habitats region in the present invention
Schematic diagram；
Fig. 4 is the schematic diagram of the device for implementing Approach for road detection of the present invention；
Fig. 5 is a kind of schematic diagram of application system of Approach for road detection of the present invention.
Specific embodiment
Figure 1A is a flow chart of the inventive method.As illustrated, under the operation of this method can pass through
Row step is completing.It is input into a width road image；Rim detection is carried out to road image；It is determined for compliance with
The image border of length requirement and pixel gradient direction condition for consistence；The image border to determined by
Carry out fitting a straight line；The energy of each intersecting straight lines pair is checked, such as by following step (10), (11)
(12) method in；Determine that the intersection point of two minimum straight lines of energy is the end point of road；
It is then determined that the region between two intersecting straight lines with outside is initial road and non-rice habitats region；Respectively
It is that Statistic analysis models are set up in above-mentioned initial road region and initial non-rice habitats region；Each picture is directed to again
The feature of vegetarian refreshments, calculates likelihood score and belong to off-highroad likelihood score which belongs to road；Determine its category
In the likelihood score and a ratio for belonging to off-highroad likelihood score of road, and this ratio is preset with one
Threshold value is compared；If this ratio is more than the predetermined threshold value, judge that current pixel point belongs to road
Region, on the contrary then belong to non-rice habitats region.
A preferred embodiment of the invention, the operation of concrete Road Detection can be by following steps
Suddenly implementing.
(1) it is input into a width road image.
(2) length threshold T is set1, general initial value for input picture catercorner length 1/3rd,
Orientation consistency angular deviation threshold value T is checked during formula (3) is set2For 3 °, fitting a straight line is provided for
Image border number be N1, usually 10 or so；The numerical value that length threshold is reduced every time is set
T3, usually 5 pixels；And meet the image border number N of orientation consistency requirement2, one
As be 5.
(3) using one-dimensional smothing filtering operator [0.037659 0.249153 0.426375 0.249153
0.037659] and one-dimensional differential filtering operator [0.109604 0.276691 0.000000-0.276691
- 0.109604] row and column of image is respectively acting on, the difference numerical in image level direction is obtained；Then
Above-mentioned one-dimensional smothing filtering operator and one-dimensional differential filtering operator are respectively acting on into the columns and rows of image again,
Obtain the difference numerical of image vertical direction；Thus calculate image in each pixel gradient direction and
Amplitude, then using the marginal point in Canny Image Edge-Detection operator detection images.
(4) marginal point isolated in image, connection are removed using grating scanning mode and eight connectivity method
Marginal point, obtains image border, if eight vicinity points of an i.e. pixel have a pixel
Point has been judged as marginal point, and this connects the two marginal points, otherwise then as isolated marginal point,
Subsequent step is not considered.
(5) check whether the length of all image borders that previous step (4) is obtained meets step (2)
The length threshold T for setting in advance1If, in current length threshold value T1Under, can collect for straight line
Fitting N1Bar meets the image border of length requirement, then jump to next step (6), otherwise then gradually
Reduce length threshold T1, every time from current length threshold value T1Reduce T3, until N can be collected1Bar meets length
The image border that degree is required.
(6) N for obtaining1Bar meets the image border of length requirement, and calculating meets length per bar will
The conforming measure value of pixel gradient direction on the image border asked：
Wherein pkAnd pk+1It is two neighbor pixels met on length requirement image border positioned at i-th,WithIt is the gradient direction on the two pixels respectively, symbol | | absolute value, C are asked in expressioni
The conforming measure value of pixel gradient direction on exactly i-th image border for meeting length requirement, this
Individual measure value is less to show that pixel gradient direction uniformity is better on edge.
(7) according to formula (1), by N1Bar meets the image border of length requirement according to each calculating
The conforming measure value of pixel gradient direction arranged from small to large, then select from small to large
N2The preferable image border of bar orientation consistency, i.e. their orientation consistency measure value C are smaller.
(8) each figure for meeting length and orientation consistency requirement for obtaining from step (7)
As edge, it is assumed that i-th meets total n on length and orientation consistency requirement image borderiIt is individual, this
A little points constitute following homogeneous coordinates matrix Ai：
Wherein in formula (2), x and y is represented and length and orientation consistency requirement image side is met positioned at i-th
The space coordinates of pixel on edge, subscript i are represented and meet length and orientation consistency requirement image border
Sequence number, and subscript then represents the sequence number of the pixel on the image border, scope is from 1 to ni, then
These matrix As for obtaining are decomposed using singular value decomposition methodi, the number of these matrixes is equal to satisfaction
The number N of the image border that length and orientation consistency are required2, matrix AiCorresponding to minimum singular value
Unit singular vectors are exactly the straight line for meeting the i-th image border fitting of length and orientation consistency requirement
Parameter vector [ai bi ci], the linear equation is aix+biy+ci=0, in other words it is using a most young waiter in a wineshop or an inn
Method is taken advantage of to obtain the straight line of these image border fittings.
(9) parameter according to resulting straight line, calculates the intersection point between these straight lines, it is assumed that i-th
The straight line parameter vector obtained with the fitting of j-th strip image border is respectively [ai bi ci] and [aj bj cj],
The intersection point of i-th and j-th strip image border fitting a straight line is obtained by straight line parameter vector cross-products.
(10) for per two intersecting straight lines, starting from their intersection point past along this two straight lines
Below image direction check on this two straight lines the gradient direction of each pixel and place rectilinear direction it
Between difference：
S represents straight line liWith straight line ljIntersection point, whereinRepresent straight line liStart along straight line from intersection point s
Towards that section below image, pkRepresent positioned at line segmentOn a pixel, andRepresent
Gradient direction on the pixel, θ (li) represent straight line liDirection (i.e.), T2
It is exactly the angular deviation threshold value arranged in step (2), # is a counter, calculating is located atOn have many
The gradient direction and rectilinear direction θ (l of few pixeli) between absolute value difference be less than angular deviation threshold value
T2,Represent straight line liStart the length of that section below along straight line towards image from intersection point s,
In formula (3)Value shows more greatly straightwayThe side of the gradient direction and place straight line of upper pixel
To more consistent.
(11) in Ordinary Rd image, road marking line or track of vehicle are all end point courts from afar
To below image, without towards above image, so the scope in image cathetus direction should be
[0 °, 180 °), wherein 180 ° of direction straight lines and 0 ° of direction straight line belong to horizontal line together, straight line vertically downward should
This is 90 °, is considered per two towards the wrapped up intra-zone texture color of intersecting straight lines below image now
Between otherness：
Straight line liWith straight line ljS points are intersected at, whereinWithStraight line l is represented respectivelyiWith straight line ljFrom intersection point s
Start along respective straight line towards that section below image, the image-region between intersecting straight lines is led to
Cross intersection point s, and respectively straight line liWith straight line ljBetween the middle separated time of angle be divided into two parts：Ω1And Ω2,
Formula (4) is exactly the otherness for calculating the texture color between this two parts, # (Ω1) and # (Ω2) represent
Respectively fall in Ω1And Ω2Pixel number in region, and p1And p2Represent in Ω respectively1And Ω2Interior
Pixel, c (p1) and c (p2) color on these pixels is represented, symbol | | | | then represent zoning Ω1With
Ω2Between color mean difference norm.
(12) according to formula (3) and (4), construction intersects at s point straightwaysWithEnergy function：
The energy function per two intersecting straight lines is checked, the intersection point with least energy straight line pair is exactly to examine
The road end point for measuring.
(13) above-mentioned two intersecting straight lines for detecting road end point can segment the image into initial road
Road region (from the region wrapped up by two lines section that road end point starts to lead to below image) and just
Beginning non-rice habitats region (removes the remainder in initial road region) in image.
(14) for initial road region, calculate the average value mu positioned at picture element point chromatic in this region
(including tri- numerical value of RGB) and the covariance matrix Σ of color, obtains positioned at pixel in this region
Point color Gaussian distribution model, in image pixel belong to road area likelihood score can pass through under
Formula is calculated：
Wherein x represents the color-values of certain pixel, and | Σ | represents the ranks for seeking color covariance matrix Σ
Formula, ΨRExpression belongs to the mark of road, and p (x | ΨR) represent that the pixel belongs to the likelihood of road area
Degree.
In another embodiment, the color space of image can be transformed into into Lab colors from RGB first
Color space, and each pixel neighborhood of a point texture eigenvalue (calculating line is calculated using a sliding window
Reason characteristic value can be according to being published in paper S.Liao, et in IEEECVPR2010 international conferences
al.,“Modeling pixel process with scale invariant local patterns for background
The local ternary masterplate SILTP of the Scale invariant described on subtraction in complex scenes "
(Scale Invariant Local Ternary Pattern), as the SILTP of each pixel is adjacent according to one
Domain is calculated a string of binary numbers, can be converted into a decimal value according to certain space order,
Thus the textural characteristics of each pixel can be represented with a decimal value).Then it is directed to initial road
Road region, calculates the mean value for being located at pixel visual signature in the region (including tri- numerical value of Lab
And texture), and the covariance matrix Σ of visual signature, at this moment calculate each pixel and belong to
In the formula (6) of road area probability, x and μ will be 4 n dimensional vector ns, and Σ will be one 4 and take advantage of 4 squares
Battle array, constant coefficient(2 π) will be changed into-2。
In another embodiment, the visual feature vector in formula (6) can adopt 7 dimensional vectors,
It is that each pixel includes tri- color values of RGB, three numerical value of Lab color spaces and one
Textural characteristics (SILTP) numerical value.So x and μ will be 7 n dimensional vector ns, and Σ will be one 7 and take advantage of 7
Matrix, constant coefficientTo be changed into
(15) for initial non-rice habitats region, all pixels point positioned at this region by initial markers is
Non-rice habitats, these pixels constitute a sample set ΨN={ y1 y2 … ym, wherein yiRepresent pixel
The color-values of point (include tri- numerical value of RGB, 1≤i≤m, m are the pictures positioned at initial non-rice habitats region
Vegetarian refreshments sum), a kind of nonparametric COLOR COMPOSITION THROUGH DISTRIBUTION model in simplest method construct non-rice habitats region is exactly
Non-rice habitats regional color histogram is calculated, the frequency that picture element point chromatic occurs in histogram is then looked at
As the pixel belong to non-rice habitats region likelihood score p (x | ΨN)。
Wherein H is a symmetric positive definite matrix,Be one it is symmetrical, compact, desired value is zero, covariance
Core distribution function of the matrix for unit matrix, its bandwidth (or variance) value can adopt adaptively selected
Method is obtained, ΨNExpression belongs to off-highroad mark, and p (x | ΨN) represent that this pixel belongs to non-road
The likelihood score in road region.
In another embodiment, the color space of image can be transformed into into Lab colors from RGB first empty
Between, and each pixel neighborhood of a point texture eigenvalue is calculated using a sliding window, for initial non-
Road area, all pixels point positioned at this region constitute a sample set ΨN={ y1 y2 … ym,
Wherein yiRepresent the visual signature (including tri- numerical value of Lab and texture eigenvalue) of pixel and adopt core
The method of density estimation, (includes tri- numerical value of Lab and texture according to the visual signature of each pixel
Numerical value), the nonparametric visual signature distributed model in non-rice habitats region is constructed, gaussian kernel function can be adopted
Method of estimation, calculates the probability that each pixel belongs to non-rice habitats region.Estimate in gaussian kernel function in addition
The variance yields of adaptively selected each pixel kernel function in meter method, can be adopted, one group is such as given
The variance yields of kernel function, selects optimum variance yields according to maximum-likelihood criterion on each pixel.
It is similar with step (14), in the calculating to initial non-rice habitats region part, each pixel
Tri- color values of RGB, three numerical value of Lab color spaces and a textural characteristics can be included
(SILTP) numerical value, and in view of the dependence between surrounding pixel, spatial positional information is additionally added, i.e.,
The joint nonparametric of the coordinate (x, y) of each pixel, this view-based access control model feature and spatial positional information
Model is avoided that isolated each pixel of consideration, but the neighborhood of each pixel is also contemplated for into.
For initial non-rice habitats region, the road image that initial detecting goes out is not ensured that and real road
Region fits like a glove, in order to strengthen the robustness (Robustness) of the relative extreme case of road model,
The model in non-rice habitats region can be based on nonparametric distributed model, also can be in the base of nonparametric distributed model
Increase by one on plinth and be uniformly distributed model.This is uniformly distributed model can be independently of the position of assessment pixel
And visual characteristic.
It is as follows based on nonparametric distributed model and the formula for being uniformly distributed model：
Wherein, α is hybrid weight, α<<1, γ is one obeys equally distributed stochastic variable.
(16) threshold tau is set0=0.5, using a simple likelihood score ratio grader by image
Pixel is divided into and belongs to road and belong to two kinds of marks of non-rice habitats：
T=ln (p (x | ΨR))-ln(p(x|ΨN)) (9)
As general likelihood score adopts logarithmic, in formula (9), ln is represented and is taken natural logrithm, if public
More than the threshold value 0.5 for arranging, the t values that formula (9) is calculated then judge that current pixel point belongs to road area,
Otherwise then belong to non-rice habitats region.
(17) it is in order to improve accuracy of this method to road area detection in road image, of the invention
Additionally use the bearing calibration to the pre- judgement of road area as shown in Figure 1B.According to Figure 1B,
The present invention determines the skyline (or referred to as horizon) in road image first, and then corrects positioned at horizon
Road pixel on line.Because road is typically below skyline, by determining skyline
Position, can correct on skyline by error flag for road pixel pixel.
And the determination of horizon line position, can adopt with the following method.First, based on color, texture,
The characteristics of image such as gray scale estimate the position at skyline place, and the skyline estimated will be input into road image point
It is divided into first area and second area；Then calculate non-similarity between the group between first and second region；
The skyline estimated is moved up and down, non-similarity between the group of change can be drawn, and passed through between determination group
The maximum of non-similarity, it may be determined that the accurate height of skyline.
As it was previously stated, the detection of skyline can be carried out based on characteristics of image such as color, texture, gray scales,
Preferably mode is being detected using features such as colors as feature space.Additionally, in order to accelerate
Detection speed, input picture can be compressed by sampling.The method of compression of images can also be first
By a low pass filter, then sampled.
During the skyline that correction is estimated, to part bottom therewith is calculated on which in image per one-row pixels
Non- similarity between the group divided, can adopt and formula is calculated as below：
Wherein,It is the mean vector of the HSV color spaces of upper part.Under being
The mean vector of partial HSV color spaces.It is three elements of part on HSV
Variance,It is the variance of three elements of part under HSV.In result of calculation between maximum group
The skyline that the row that non-similarity is located as determines.
Above-mentioned one kind preferably specific embodiment for the present invention.Those skilled in the art can manage
Solution, is to realize the purpose of the present invention, and not above-mentioned all steps are all necessary, and multiple steps
Between order flexibly can also exchange.For example, the detection of image border length and according to length and pixel
Uniformity etc. of point gradient direction is to the selection of image border not necessarily step.And in each step
The selection of circular and parameter is also not necessarily in strict accordance with the mode described in above-described embodiment
Come carry out.The place to go of the marginal point for example isolated in image not necessarily adopts grating scanning mode and eight
Connection method, it would however also be possible to employ other conventional determination methods；The energy of intersecting straight lines pair not necessarily base
Determine in the otherness in relevant range between color, it is also possible to based on other visual signatures, such as ash
Degree, texture etc., calculate.Orientation consistency angular deviation threshold value T is checked in such as formula (3) again2
Not necessarily be set to 3 °, it would however also be possible to employ 2 °, 4 ° it is equivalent；It is provided for the image border of fitting a straight line
Number is N1Also it is not necessarily intended to select as 10, it is also possible to select the numerical value such as 15,20；Length threshold is each
The numerical value T of reduction3Can also unrestricted choice, rather than be necessary for 5 pixels；Meet orientation consistency
The image border number N of requirement2Could be arranged to the bar number outside 5.
By being processed with the method for the present invention to thousands of width road images have been downloaded from internet,
And the road area in each image is marked by hand, determining this method carries out the accuracy rate of Road Detection,
The technical indicator such as recall rate and effective road markings rate.In these images, condition of road surface difference is very big, has
Substantially the urban road of road surface identification, the only desert of track of vehicle and snow road, highway,
The very dark road of road ambient lighting or road at night time etc..The present invention is carried out into pixel school using skyline
The corresponding skill of the method proposed in the paper that technical indicator and Kong of method part et al. are delivered before just
Art index is done one and is compared, it is possible to find accuracy rate and valid trace road sign of the method for the present invention in Road Detection
All increase in knowledge rate, as a result as shown in table 1.
The quantitative comparison of 1. 3 kinds of method testing results of table
Fig. 2A schematically illustrates said determination based on the parts of images in above-mentioned thousands of width road images
Method and result.From left to right, the image positioned at first row is input picture.Secondary series is that correspondence is defeated
Enter the road area of the manual mark of image.3rd row are according to method described in Kong et al. papers
The road for detecting, wherein dark point is the initial end point obtained according to the method, light color point is root
According to the revised end point of the method, and two lines section represents two borders for detecting road respectively.
4th row are the end points that the method according to the invention is detected, and wherein midium line segment is for calculating two
The difference of the two subdivision texture colors separated by intermediate angle line in the wrapped up region of bar intersecting straight lines,
It is some of the construction per two intersecting straight lines energy functions.In figure, out conductor section is exactly have minimum energy
Two intersecting straight lines of value, so the region on the outside of this between two lines section detects
Initial road region, in image, other parts are then non-rice habitats region；According to initial road region and just
Beginning non-rice habitats region, is respectively adopted Gaussian statistics distributed model and nonparametric statistics distributed model Lai Wei roads
Statistical model is set up in road region and non-rice habitats region, it is not necessary to by any extraneous knowledge.Rightmost
Row are Road Detection results.
As previously mentioned, for certain types of road image, the detection method based on end point may not
Initial road region can be accurately detected.For example for the image of tee T road, or vehicle is just
In the road image for turning left or photographing when turning right, it may be possible to road cannot be detected and disappeared
Lose point.In such cases, then the detection method based on end point is no longer suitable for, and is used following
Initial road area determination method or other suitable methods shown in Fig. 3 A and 3B is substituted.
And by the road image com-parison and analysis in some road image databases, it can be seen that this
Bright employing is carried out in road area after pixel bearing calibration based on skyline, and its Road Detection is accurate
Degree and accuracy have and are obviously improved for comparing existing method.The road image data of actual test
Storehouse includes Sowerby databases (comprising 104 width British Countryside path images), CamVid databases
(path image shot in comprising more than 700 width running cars) and KITTI databases (include without mark
The multilane urban road UMM of the urban road UU of knowledge, the urban road UM of tape identification and tape identification,
Actual test wherein 289 width road images).Table 2 lists the method for the present invention and three kinds of known methods
Path image test result to three image data bases.In table, numerical value is that three kinds of method tests draw
Effective road result exponent, the denominator numeral in its bracket represent detected total number of images, and divide
Sub-figure represents that TIP'2010 methods can detect the picture number of figure road.
The quantitative comparison of 2. present invention of table and three kinds of known method testing results
In addition, compare it is known based on the lane detection technology of image for, the method for the present invention may be used also
More accurately to detect the feature in road, such as track boundary etc..Fig. 2 B show the present invention
Contrast of the method with several known Approach for road detection to feature recognition ability in road.Precedent reference
In detection method embodiment described in Fig. 1, step 1 describes initial in input picture for determining to 13
Road area and a kind of method in initial non-rice habitats region, step 14 to 16 describe a kind of amendment just
The method of the estimation of beginning road area, step 17 are come to judging in image based on a skyline for determining
The method that pixel in road area is corrected.Initial road region and initial non-rice habitats region
It is not always necessary that being determined with method shown in Fig. 1.In another embodiment, initial road region and
Initial non-rice habitats region can be determined with basic rough estimate.For example, as shown in Figure 3A, can will be defeated
Enter the midpoint of image, and the triangle shadow region that image lower-left boundary point and image bottom right boundary point are constituted
Estimation of the domain as initial road region, and other regions then conduct in image in addition to Delta Region
The estimation in initial non-rice habitats region.Fig. 3 B show another method for determining initial road region,
The method determines a half-circle area being located inside image as initial with image baseline midpoint as the center of circle
The estimation of road area, and other regions in image in addition to half-circle area are then as initial non-rice habitats
The estimation in region.
Experimental result shows, for particular kind of road image, for example non-structural road image (
Be exactly no clear and definite road boundary in image, and the situation of road marking line), using so it is coarse just
Beginning segmentation result, Road Detection result may be specifically referred to better than the Detection results based on end point
Table 1.
With respect to the known method based on image detection road, computational methods of the present invention are easy, road inspection
Degree of testing the speed has and greatly improves.In addition, can be save using the method that initial road region is estimated first will
Estimate the calculating section of lane boundary line, accelerate detection speed.And it is right before the calculating of Road Detection
Input picture is compressed as previously described, can further speed up Road Detection speed.Table 3 shows
The method of the present invention is contrasted with the detection speed of Approach for road detection known to three kinds.
The comparison of 3. present invention of table and three kinds of known method detection speeds
| TIP'2010[7] | ITS'2011[13] | PFSS‘2006[10] | The inventive method | |
| Sowerby(64 96) | 4～5s/ frames | ～0.25s/ frames | ～3s/ frames | ～0.1s/ frames |
| CamVid[28](240 320) | 9～10m/ frames | ～8s/ frames | ～45s/ frames | ～2s/ frames |
| KITTI[29](375 1242) | 9～10m/ frame * | 165s/ frames | 4～5m/ frames | ～9s/ frames |
Preceding method can be applicable to the Road Detection based on still image, be equally applicable for based on Dynamic Graph
Road Detection on picture or video.When being applied to based on Road Detection on dynamic image or video,
The method includes with reference to the relativity of time domain before and after dynamic image or video sequence between frame, further adopts
With filtering method, such as kalman filter method realizes the Road Detection of dynamic image sequence.
It is aforementioned that for video sequence, the first frame can estimate road area using above-mentioned simple method,
Due to there is very big relativity of time domain before and after video between frame, subsequent frame can be with the detection of former frame
As a result as the estimation in initial road region, then using proposed likelihood score ratio grader with regard to energy
Amendment road area, reaches time dependent correction result.
Fig. 4 shows a kind of device for performing the inventive method.The device includes that an image/video is gathered
Equipment, the equipment can adopt digital photographing/video camera, or other similar devices, to obtain mileage chart
Picture and/or video.The image/video collecting device is connected with a processor by a receiving device.This connects
Receiving unit can by wired or wireless data transfer receive come from image/video collecting device image and
/ or video, and by its transmission processor.The receiving device can also receive order, such as user instruction,
To control or adjust the execution of image Approach for road detection.Processor can be received and be gathered from image/video
The image and/or video of equipment, and aforementioned Approach for road detection is performed under the control of dependent instruction.Inspection
After having surveyed, processor can send operational order based on testing result, such as the operation of control vehicle or machine
The instruction of people's walking.Or testing result is only sent to kinetic control system by processor, a such as vehicle
Automatically/DAS (Driver Assistant System) or robot ambulation control system, by the kinetic control system based on detection knot
The control moved by fruit.Image Road Detection result and/or input picture can be connected to place by one
The display device of reason device is shown.
Fig. 5 is a kind of schematic diagram of application apparatus of Approach for road detection of the present invention.This method can be used to
The road that automatic detection driving vehicle is located.Graphic display unit includes a vehicle-mounted pick-up equipment, the car
Carry picture pick-up device and may be installed position on or near vehicle front windshield, and can be to vehicle front road
Imaged.The image for photographing can be by wired or wireless linkup transmit to on-board processing and storage
Equipment, and can be displayed on Vehicular display device.On-board processing equipment, it is possible to use this method is based on institute
The image for receiving carries out road Identification, and sends control to vehicle control apparatus according to road Identification result
Instruction.
Above-mentioned vehicle-mounted pick-up equipment may be mounted at vehicle front, it is also possible to installed in rear view of vehicle or its
His position, instructs or controls to provide to reversing or other driver behaviors.The vehicle-mounted pick-up equipment can be with
With some special camera functions, such as night vision function so that the device of the present invention lighting conditions compared with
Driver behavior is provided in poor environment and instruct or control.
The method of the present invention can be used for other any rail-frees motions outside Vehicular automatic driving or
In the environment of driving.For example, this method can apply to can walking robot road Identification, toy
Identification of road etc. in vehicle automatic running.
" road " use herein, is not limited in the road travelled by person-carrying vehicle, also including which
Passage in the passage of his type, such as supermarket between shelf, the tool for being available for toy car to travel in room
There is the region on border etc..
The embodiment introduced referring to the drawings is directed to use with the defined stream of computing device computer program
Journey.So the present invention extends to the computer program that can implement the present invention, one kind is especially attached to
The computer program of carrier.
This computer program can be source code, object code or between source code and target generation
A kind of code between code, such as code of part compilation, or other are any suitable for implementing this
The code of bright institute's definitim flow.
Aforementioned bearer can be it is any can be with the carrier of storage program.For example this carrier can be comprising storage
Medium, including read-only storage, such as read-only optical disc or semiconductor ROM, or magnetic storage
Medium, such as floppy disk or hard disk, or optical storage media.Further, this medium can be non-physical
Medium, for example, can pass through the electric signal or optical signal of cable or optical cable or other modes transmission.This
Medium could be for the integrated circuit for inscribing the computer program, and the integrated circuit is made for
Run aforementioned related procedure.The present invention can be realized by software, it is also possible to by hardware device or
Software is implemented in combination in hardware.
Claims (27)
1. a kind of method based on image detection road, it is characterised in that the method comprises the following steps：
Input road image；
The likelihood score anticipation picture in road area or non-rice habitats region is belonged to based on each pixel in image
Vegetarian refreshments belongs to road area or non-rice habitats region；
Road area is belonged to each pixel according to a visual signature amendment of the input road image
Or the anticipation in non-rice habitats region.
2. the method based on image detection road according to claim 1, wherein described vision
The skyline being characterized as in the road image, the amendment include：
Determine the skyline in the road image；
According to the relative position relation amendment of each pixel in image and the skyline for the pixel
Point belongs to the anticipation in road area or non-rice habitats region.
3. the method based on image detection road according to claim 1 and 2, it is characterised in that
The wherein described likelihood score for belonging to road area or non-rice habitats region to each pixel in image it is pre-
Sentence including：
By be input into Road image segmentation into initial road region and initial non-rice habitats region；
In determining image, each pixel belongs to the likelihood score of road area and belongs to non-rice habitats region seemingly
So spend；
Each pixel is belonged to the likelihood score of road area and the ratio of the likelihood score for belonging to non-rice habitats region
Value compares with a threshold value with anticipation whether the pixel belongs to road area.
4. the method based on image detection road according to Claims 2 or 3, it is characterised in that
The step of skyline in the wherein described detection road image, includes：
One skyline is estimated based on characteristics of image, this is estimated skyline and the road image is divided into first
Region and second area；
The height of the estimated skyline of correction, so that between the group of the first area and second area
Non- similarity is maximized.
5. the method based on image detection road according to claim 4, it is characterised in that the party
Method includes estimating the skyline based on the color character of image.
6. the method based on image detection road according to aforementioned any claim, its feature exist
In before road area or the anticipation of non-rice habitats region is belonged to each pixel, to the aqueduct
Road image is compressed.
7. the method based on image detection road according to claim 6, it is characterised in that wherein
Described image compression is carried out first LPF and then by way of sampling.
8. the method based on image detection road according to aforementioned any claim, its feature exist
Input Road image segmentation is included into initial road region and initial non-rice habitats region in described：
Image lower-left boundary point, bottom right boundary point and image midpoint are connected and composed into a Delta Region two-by-two,
And determine that this Delta Region is initial road region；And
Region beyond above-mentioned Delta Region in image is defined as into initial non-rice habitats region.
9. the method based on image detection road according to any claim in claim 1-7,
It is characterized in that described wrap input Road image segmentation into initial road region and initial non-rice habitats region
Include：
Determine the gradient direction and marginal point of pixel in road image；
Connect marginal point to obtain image border；
It is determined that meeting the image border of pre-provisioning request in resulting image border；
The straight line of these image border fittings is obtained for the above-mentioned image border for meeting pre-provisioning request；
It is determined that the intersection point between resulting straight line；
Construction intersects at the energy function of the straight line pair of each intersection point and energy is determined always according to each straight line
The intersection point of line pair is the road end point for detecting；And
By two straight lines of above-mentioned determination road end point segment the image into the initial road region and
The initial non-rice habitats region.
10. the method based on image detection road according to aforementioned any claim, its feature
It is that each pixel belongs to and the likelihood score of road area and belongs to non-rice habitats region in the determination image
Likelihood score include：
Determine the statistical distribution pattern of the particular visual characteristics of pixel in initial road region, and according to
During on the statistical distribution pattern and pixel, particular visual characteristics calculate image, each pixel belongs to road
The likelihood score in region；And
It is determined that in initial non-rice habitats region the particular visual characteristics of pixel statistical distribution pattern, and root
Determine that each pixel belongs to non-in image according to particular visual characteristics on the statistical distribution pattern and pixel
The likelihood score of road area.
11. methods based on image detection road according to claim 10, it is characterised in that institute
Stating particular visual characteristics includes color character.
12. methods based on image detection road according to claim 10, it is characterised in that institute
Stating particular visual characteristics also includes textural characteristics.
13. methods based on image detection road according to claim 11 or 12, its feature
It is：
The color character and/or the statistical distribution pattern of textural characteristics of pixel in the initial road region
For Gaussian distribution model；And
The color character and/or the statistical distribution mould of textural characteristics of pixel in the initial non-rice habitats region
Type is nonparametric statistics distributed model.
14. methods based on image detection road according to claim 11 or 12, its feature
It is the color character and/or the statistical distribution mould of textural characteristics of pixel in the initial non-rice habitats region
Type adds for nonparametric statistics distributed model and is uniformly distributed model.
15. methods based on image detection road according to claim 14, it is characterised in that institute
State Model Independent is uniformly distributed in the position of assessment pixel.
16. methods based on image detection road according to claim 14, it is characterised in that institute
State Model Independent is uniformly distributed in visual characteristic.
17. according to any claim in claim 9 to 16 based on image detection road
Method, it is characterised in that the connection marginal point is included with obtaining image border：
The marginal point isolated in image is removed using grating scanning mode and eight connectivity method, connection is remaining
Marginal point is obtaining image border.
18. according to any claim in claim 9 to 17 based on image detection road
Method, it is characterised in that meet the image border of pre-provisioning request in the image border obtained by the determination
Including：
Detect whether the image border obtained by connection marginal point meets predetermined according to a pre-set length threshold
Length requirement；And
The image border quantity for meeting predetermined length requirement as obtained by is less than a predetermined number, by one
Predetermined decrement value gradually reduces length threshold, until that collects predetermined quantity meets predetermined length requirement
Image border.
19. methods based on image detection road according to claim 18, it is characterised in that institute
The image border for meeting pre-provisioning request in stating the image border obtained by determining includes：
Calculating meets the conforming measure value of pixel gradient direction on the image border of length requirement per bar；
And
The pixel of predetermined quantity is selected from the image border for meet length requirement according to gained measure value
The preferable image border of gradient direction uniformity.
20. sides based on image detection road according to any claim in claim 9-19
Method, it is characterised in that the image border for meeting pre-provisioning request obtains the straight of image border fitting
Line includes the straight line that these image border fittings are obtained using least square method.
21. sides based on image detection road according to any claim in claim 9-20
Method, it is characterised in that the construction intersects at the energy function of the straight line pair of each intersection point to be included：
For per two intersecting straight lines, starting along this two straight lines toward image base from its intersection point
Direction determines the difference on this two straight lines between the gradient direction and place rectilinear direction of each pixel
It is different；
It is determined that the intra-zone texture wrapped up towards the intersecting straight lines of image base per two and color it
Between otherness；And
Between gradient direction and place rectilinear direction according to each pixel on above-mentioned two intersecting straight lines
Difference and the otherness structure between the intra-zone texture that wrapped up of above-mentioned two intersecting straight lines and color
Make the energy function for intersecting at each intersection point straight line pair.
22. methods based on image detection road according to aforementioned any claim, its feature
It is the input picture for dynamic image, the method is included with reference to before and after dynamic image sequence between frame
Relativity of time domain, the Road Detection of dynamic image sequence is further realized using filtering method.
23. methods based on image detection road according to claim 22, it is characterised in that after
In continuous two field picture, initial road region and initial non-rice habitats region are with the image detection result of adjacent former frame
Based on determine.
24. a kind of devices based on image detection road, it is characterised in that the device includes：
Divide the image into into the equipment in initial road region and initial non-rice habitats region；
The likelihood score anticipation picture in road area or non-rice habitats region is belonged to based on each pixel in image
Vegetarian refreshments belongs to the equipment in road area or non-rice habitats region；Determine setting for the skyline in road image
It is standby；And
According to the relative position relation amendment of each pixel in image and the skyline for the pixel
Point belongs to the anticipation in road area or non-rice habitats region.
25. a kind of devices based on image detection road, it is characterised in that the device includes：
One storage device；And
One processor, is connected with the storage device, arbitrary in requiring 1 to 23 for perform claim
Method described in claim.
The guider of a kind of 26. vehicles or other movable equipments, it is characterised in that the guider
Including the device based on image detection road according to claim 24 or 25.
A kind of 27. computer readable storage devices, it is characterised in that the computer readable storage devices bag
Computer code is included, can be caused when the code is run on a computer arbitrary in claim 1 to 23
The execution of claim methods described.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510622576.6A CN106558051A (en) | 2015-09-25 | 2015-09-25 | A kind of improved method for detecting road from single image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510622576.6A CN106558051A (en) | 2015-09-25 | 2015-09-25 | A kind of improved method for detecting road from single image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106558051A true CN106558051A (en) | 2017-04-05 |
Family
ID=58414578
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510622576.6A Pending CN106558051A (en) | 2015-09-25 | 2015-09-25 | A kind of improved method for detecting road from single image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106558051A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108052904A (en) * | 2017-12-13 | 2018-05-18 | 辽宁工业大学 | The acquisition methods and device of lane line |
| CN108399614A (en) * | 2018-01-17 | 2018-08-14 | 华南理工大学 | It is a kind of based on the fabric defect detection method without sampling small echo and Gumbel distribution |
| CN108427935A (en) * | 2018-03-28 | 2018-08-21 | 天津市测绘院 | Streetscape compares the generation method and device of image |
| CN108510740A (en) * | 2018-05-04 | 2018-09-07 | 百度在线网络技术（北京）有限公司 | Report the method for digging and device of road conditions by mistake |
| CN108960055A (en) * | 2018-05-30 | 2018-12-07 | 广西大学 | A kind of method for detecting lane lines based on local line's stage mode feature |
| CN110059629A (en) * | 2019-04-19 | 2019-07-26 | 北京理工大学 | A kind of the road surface area recognizing method and system of surface mine road |
| CN110832568A (en) * | 2017-07-05 | 2020-02-21 | 歌乐株式会社 | Vehicle environment recognition device |
| CN111223114A (en) * | 2020-01-09 | 2020-06-02 | 北京达佳互联信息技术有限公司 | Image area segmentation method and device and electronic equipment |
| CN112417943A (en) * | 2019-08-21 | 2021-02-26 | 美光科技公司 | Advanced Driver Assistance System (ADAS) operation with algorithmic skyline detection |
| CN114820629A (en) * | 2022-07-01 | 2022-07-29 | 山东意吉希精密制造有限公司 | Welding identification method for automobile parts |
-
2015
- 2015-09-25 CN CN201510622576.6A patent/CN106558051A/en active Pending
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110832568A (en) * | 2017-07-05 | 2020-02-21 | 歌乐株式会社 | Vehicle environment recognition device |
| CN110832568B (en) * | 2017-07-05 | 2022-03-25 | 歌乐株式会社 | Vehicle environment recognition device |
| CN108052904A (en) * | 2017-12-13 | 2018-05-18 | 辽宁工业大学 | The acquisition methods and device of lane line |
| CN108052904B (en) * | 2017-12-13 | 2021-11-30 | 辽宁工业大学 | Method and device for acquiring lane line |
| CN108399614A (en) * | 2018-01-17 | 2018-08-14 | 华南理工大学 | It is a kind of based on the fabric defect detection method without sampling small echo and Gumbel distribution |
| CN108427935A (en) * | 2018-03-28 | 2018-08-21 | 天津市测绘院 | Streetscape compares the generation method and device of image |
| CN108510740A (en) * | 2018-05-04 | 2018-09-07 | 百度在线网络技术（北京）有限公司 | Report the method for digging and device of road conditions by mistake |
| CN108960055B (en) * | 2018-05-30 | 2021-06-08 | 广西大学 | Lane line detection method based on local line segment mode characteristics |
| CN108960055A (en) * | 2018-05-30 | 2018-12-07 | 广西大学 | A kind of method for detecting lane lines based on local line's stage mode feature |
| CN110059629A (en) * | 2019-04-19 | 2019-07-26 | 北京理工大学 | A kind of the road surface area recognizing method and system of surface mine road |
| CN112417943A (en) * | 2019-08-21 | 2021-02-26 | 美光科技公司 | Advanced Driver Assistance System (ADAS) operation with algorithmic skyline detection |
| CN111223114A (en) * | 2020-01-09 | 2020-06-02 | 北京达佳互联信息技术有限公司 | Image area segmentation method and device and electronic equipment |
| CN114820629A (en) * | 2022-07-01 | 2022-07-29 | 山东意吉希精密制造有限公司 | Welding identification method for automobile parts |
| CN114820629B (en) * | 2022-07-01 | 2022-09-02 | 山东意吉希精密制造有限公司 | Welding identification method for automobile parts |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106558051A (en) | A kind of improved method for detecting road from single image | |
| CN109165549B (en) | Road identification obtaining method based on three-dimensional point cloud data, terminal equipment and device | |
| CN111539273B (en) | Traffic video background modeling method and system | |
| CN110415277B (en) | Multi-target tracking method, system and device based on optical flow and Kalman filtering | |
| CN105260699B (en) | A kind of processing method and processing device of lane line data | |
| CN108986465B (en) | Method, system and terminal equipment for detecting traffic flow | |
| CN104331873A (en) | Method for detecting road from single image | |
| WO2019175286A1 (en) | Image annotation | |
| CN103164711B (en) | The method of region based on pixel and support vector machine artificial abortion's density Estimation | |
| CN108052904B (en) | Method and device for acquiring lane line | |
| CN105608417B (en) | Traffic lights detection method and device | |
| CN109791598A (en) | The image processing method of land mark and land mark detection system for identification | |
| CN109284674A (en) | A kind of method and device of determining lane line | |
| CN104766058A (en) | Method and device for obtaining lane line | |
| CN108280450A (en) | A kind of express highway pavement detection method based on lane line | |
| Hinz | Detection and counting of cars in aerial images | |
| CN108171695A (en) | A kind of express highway pavement detection method based on image procossing | |
| CN106780727B (en) | Vehicle head detection model reconstruction method and device | |
| CN113034586B (en) | Road inclination angle detection method and detection system | |
| CN110222658A (en) | The acquisition methods and device of road vanishing point position | |
| CN112070756A (en) | Three-dimensional road surface disease measuring method based on unmanned aerial vehicle oblique photography | |
| CN102831419B (en) | Method for detecting and blurring plate number in street view image rapidly | |
| CN105631849B (en) | The change detecting method and device of target polygon | |
| JP2020160840A (en) | Road surface defect detecting apparatus, road surface defect detecting method, road surface defect detecting program | |
| KR102416714B1 (en) | System and method for city-scale tree mapping using 3-channel images and multiple deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170405 |