CN105320940A - Traffic sign detection method based on center-boundary connection model - Google Patents
Traffic sign detection method based on center-boundary connection model Download PDFInfo
- Publication number
- CN105320940A CN105320940A CN201510644535.7A CN201510644535A CN105320940A CN 105320940 A CN105320940 A CN 105320940A CN 201510644535 A CN201510644535 A CN 201510644535A CN 105320940 A CN105320940 A CN 105320940A
- Authority
- CN
- China
- Prior art keywords
- center
- polygon
- point
- frontier point
- connection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/582—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of traffic signs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a traffic sign detection method based on a center-boundary connection model. The method comprises steps of: matching connections in a training sample by using a dynamic programming method, to find a corresponding relationship of each connection of a polygon in all training samples; assuming that a length of each connection in all the training samples meets the Gaussian distribution, and an inclination angle of each connection and a gradient direction of a boundary point meet the Von Mises distribution, estimating connection parameter distribution by using an expectation-maximization method, wherein parameter distribution of each connection in the polygon forms a center-boundary connection model of the polygon; obtaining a potential center and boundary point of the polygon by using a rotational symmetry projection method; and using a projection center as the potential center of the polygon, wherein boundary points corresponding to all the connections in the center-boundary connection model of the polygon form a detected polygon. According to the detection method, a traffic sign can be quickly detected.
Description
Technical field
The invention belongs to Image processing and compute machine vision technique field, particularly a kind of detection method of traffic sign.
Background technology
Road traffic sign detection is an important research direction in computer vision.As road safety affiliated facility, traffic sign plays an important role in specification traffic behavior, instruction condition of road surface, guarantee road effect, guiding pedestrian and safe driving etc.As the gordian technique of intelligent transportation system, road traffic sign detection is automatic Pilot, auxiliary drives and ingredient that the system such as traffic sign maintenance is indispensable.
Road traffic sign detection utilizes the color of traffic sign or shape facility by the traffic sign region detection in image and splits, and detects the precision of the follow-up traffic sign classification of whether accurate directly impact.
Method for traffic sign detection is broadly divided into two kinds: based on the method for color characteristic and the method for Shape-based interpolation feature.Method for traffic sign detection based on color characteristic is that rule of thumb image is divided into traffic sign region and non-traffic sign region by threshold value or other color characteristic in specific color space.The method easy to understand and realization, can get rid of background area rapidly.But colouring information easily affects by extraneous factor, the color degradation of such as traffic sign self, shade and illumination variation all can cause declining to a great extent of accuracy of detection.Therefore, the testing result of color characteristic is adopted separately cannot to meet the requirement of practical application.The method for traffic sign detection of Shape-based interpolation is feature mainly with edge, utilizes traffic sign to be the characteristic of regular polygon, the detection of traffic sign is converted into the detection of regular polygon (as Gonz á lez
garridoMA, LlorcaDF, etal.Automatictrafficsignsandpanelsinspectionsystemusing computervision [J] .IEEETransactionsonIntelligentTransportationSystems, 2011,12 (2): 485-499.).
Generalized Hough Transform (GeneralizedHoughtransformation, GHT) ([2] DaviesER.MachineVision:Theory, Algorithms, Practicalities (ThirdEdition) .SanFrancisco:Elsevier, 2005.387-410) utilize polygonal geometrical property, the test problems of variable space figure is converted into the clustering problem of parameter space, realizes polygonal direct-detection.Be characterized in simple directly, but due to calculated amount large, be generally only applicable to the polygon detecting that the equilateral number of triangle is less.
Barnes etc. propose to utilize radial symmetry transform to detect the method for circle, the method is extended to octagon, rectangle and triangle (BARNESN afterwards, ZELINSKYA.Real-timeradialsymmetryforspeedsigndetection [J] .IntelligentVehiclesSymposium, 2004,566-571).The method adopts symmetrical marginal point to its radial symmetry central projection, can be formed centrally the maximum projection centre of brightness, can determine the particular location of traffic sign thus after all marginal point projections in circular traffic sign.Radial symmetry transform is a kind of univariate conversion, and each marginal point is voted to multiple totalizer separately independent of its neighborhood, which results in higher false alarm rate.
Above method is all the polygon detecting methods based on projection in essence, Barnes etc. propose the regular polygon detection method (BarnesN based on posterior probability, LoyG, ShawD.Theregularpolygondetector [J] .PatternRecognition, 2010,43 (3): 592-602).First the method extracts edge image, then according to polygonal geometric properties, utilize posterior probability to define polygonal probability density function.Compared to projecting method, the method has higher verification and measurement ratio and lower false alarm rate, but the method edge quality requirements is higher, and in the incomplete situation in edge, Detection results declines very fast.
Based on above analysis, a kind of method can not detect traffic sign effectively at present, and the present inventor researches and develops this case just under this research background.
Summary of the invention
Object of the present invention, be to provide a kind of method for traffic sign detection based on center-border link model, it can realize the quick detection of traffic sign, effectively can detect and the polygon that visual angle tilts and marginal portion lacks occurs, and time complexity is lower.
In order to reach above-mentioned purpose, solution of the present invention is:
Based on a method for traffic sign detection for center-border link model, comprise the steps:
(1) training sample coupling: by the border uniform sampling of training sample, form some frontier points, this polygon is represented to connecting of frontier point with center, then utilize dynamic programming method the connection in first training sample and other training sample to be mated, find out polygon and be respectively connected to corresponding relation in all training samples;
(2) center-border link model training: suppose that each length be connected in all training samples meets Gaussian distribution, the angle of inclination connected and the gradient direction of frontier point meet VonMises distribution, utilize and expect that maximum solution estimates the parameter distribution connected, in polygon, each parameter distribution connected constitutes this polygonal center-border link model;
(3) polygon center and size estimation in image to be detected: the marginal information extracting each image to be detected, adopt Rotational Symmetry projecting method, symmetrical marginal point is selected to project to its center of rotational symmetry, the point set obtaining projection centre and project to this projection centre, the center potential as polygon and frontier point, calculate polygonal dimensional information according to the distance between center and frontier point;
(4) polygon detecting: take projection centre as the potential center of polygon, each subpoint that is connected to concentrates the frontier point finding the maximum probability that matches, and in polygon center-border link model, all frontier points connecting correspondence constitute the polygon detected.
In above-mentioned steps (1), represent that polygonal method refers to set polygon regarded as center and connect to frontier point, each connection l tlv triple (d with center to connecting of frontier point, θ, α) represent, wherein, d represents that the distance of frontier point is arrived at center, θ represents the angle of connection and horizontal direction, α represents the gradient direction of frontier point, and wherein the span of θ is [0,2 π], the span of α is [0, π].
The detailed content of above-mentioned steps (1) is:
(11) in N number of training sample, random selecting polygon, as first training sample, to its border uniform sampling, obtains n frontier point, to connect represent this training sample with center to the n of frontier point; To all the other sample boundary uniform samplings, obtain m frontier point, training sample is expressed as m the set that connect of center to frontier point, and m>n;
(12) remember that first training sample is P={l
i, 1<i<n, all the other training samples are P'={l'
j, 1<j<m, its center is respectively o and o', the Matching power flow Matrix C=(c of structure n × m
ij), c
ijrepresent l
iand l
j' between diversity factor, be defined as:
c
ij=d(l
i,l
j')=λ
θ*min(|θ
i-θ
j'|,2π-|θ
i-θ
j'|)+
λ
α*min(|α
i-α
j'|,π-|α
i-α
j'|)+λ
d*|d
i-d
j'|
In formula, | θ
i-θ '
j|, | α
i-α '
j| with | d
i-d'
j| represent respectively and connect l
iand l'
jdifference in the horizontal direction in angle, frontier point gradient direction and length; λ
θ, λ
αand λ
dthen represent the weight that this three species diversity is shared in Matching power flow;
(13) the cost matrix C Matching power flow decided between different connection is defined, in polygon P and P', the objective function of matching connection is expressed as: π: { 1,, n} → { 0,1, m}, and 1 ... each value in m} can only be mated once, to represent in second training sample not and be connected l with π (i)=0
ithe option matched, the object representation that solves of polygon matching problem is:
For all π (i)=0, c
i, π (i)=ε, ε are adjustable threshold values, and only have the Matching power flow between two connections to be less than ε, this coupling is just allowed to;
(14) the connection l in first training sample is determined by polygon coupling
icorrespondence in all the other N-1 training sample connects.
In Matching power flow computation process, the establishing method of each parameter is as follows:
A) λ
θ+ λ
α+ λ
d=1, first determine λ
θand λ
dvalue;
B) on training set, adopt the mode of traversal to obtain optimized parameter configuration, interval [0.2,0.8] changes λ
θand λ
dvalue, obtain the matching rate under various parameter combinations, get matching rate the highest time parameter configuration;
C) on [0,1] interval, change the value of ε, obtain corresponding matching rate and wrong matching rate curve, determine the value of ε accordingly.
The detailed content of above-mentioned steps (2) is:
(21) according to the training sample matching result that step (1) obtains, suppose that the length of each connection in all training samples meets Gaussian distribution, the angle of inclination connected and the gradient direction of frontier point meet VonMises distribution, then the model representation of this connection is:
Wherein, u
dand σ
dcentered by the average of Gaussian distribution of the distance d of frontier point and variance, u
θand k
θ, u
αand k
αbe average and the variance of the VonMises distribution of connection and the angle theta of horizontal direction and the gradient direction α of frontier point respectively, wherein, subscript i represents corresponding to i-th connection;
(22) with expecting that maximum solution estimates the parameter u of Gaussian distribution and VonMises distribution
d, σ
d, u
θ, k
θ, u
αand k
α, obtain the parameter distribution connected.
The detailed content of above-mentioned steps (3) is:
(31) edge image of image to be detected is asked, the gradient of edge calculation;
(32) by the gradient direction scope of pixel (-π, π] be evenly divided into 36 subsets, the scope of each subset is π/18, finds the point that gradient direction meets Rotational Symmetry condition;
(33) for meeting each point of step (32) conditional to finding its center of rotational symmetry, if p
kand p
tbe two symmetrical marginal points, its center of rotational symmetry p is obtained by following formula:
In formula, (x
k, y
k) be a p
kcoordinate, γ is p
k, p
tline and transverse axis between angle, β is p, p
kline and p
t, p
kthe angle of line, r is p
kor p
tto the distance of center of rotational symmetry p point;
(34) point meeting step (32) conditional projects to its center of rotational symmetry, and the more brightness of point in perspective view of winning the vote is larger;
(35) in projection process, record the position of projection centre and subpoint, utilize following formula to calculate polygonal dimensional information s:
Wherein, Γ (p) represents that the point that the oriented p of institute projects is right, and its quantity is M, || p
kp
t|| represent that point is to p
kand p
tbetween distance.
The detailed content of above-mentioned steps (4) is:
(41) adopt Rotational Symmetry projecting method to form multiple projection centre in image to be detected, several maximum projection centre of getting tickets is as the potential center of polygon, and corresponding projection point set is possible frontier point;
(42) for one of them projection centre o
w, remember that the point set to its projection is E={e
w}={ d
w, θ
w, α
w; 1<w<ne}, wherein ne is to o
wcounting of projection; For each some e in E
w, calculate the probability that it belongs to the connection of i-th, polygon:
(43) point that matching probability is maximum is i-th and connects corresponding frontier point, is designated as e
i *, then the shape that the polygon finally detected is made up of n frontier point, i.e. e (o
w)={ e
1 *, e
2 *..., e
n *,
(44) the center-border link model of circle, octagon, squares and triangles is used respectively, detect the frontier point that a certain projection centre is corresponding, calculate the matching score of often kind of polygon model and frontier point, score the most much higher limit shape is detected polygonal type.
Adopt after such scheme, the present invention utilize that traffic sign is triangular in shape, square, the regular polygonal characteristic such as octagon and circle, road traffic sign detection is converted into the detection of regular polygon, there is following characteristics:
(1) polygonal visual angle is tilted and deformation insensitive:
What the traffic sign under natural scene can be subject to artificial or weather conditions affects generation deformation, visual angle in image acquisition process tilts also can make the traffic sign generation deformation collected, the present invention adopts center to represent polygon to the connection of frontier point, when polygon generation deformation, adjacent order between connection is constant, effectively can detect the traffic sign that deformation and visual angle inclination occur.
(2) can the effective traffic sign that blocks of detecting portion:
Outdoor traffic sign is often subject to blocking of other object and forms incomplete edge, detection method based on projection is strongly depend on the quality at edge, in the present invention, the frontier point mated with it is found in each connection respectively in marginal point, the continuity of edge requires lower, can the polygon of Edge detected excalation, i.e. the traffic sign of partial occlusion.
(3) in testing process, polygonal Shape Classification is completed:
Polygon common in traffic sign has triangle, square, octagon and circle, the present invention connects distributed model for the training of these polygons respectively, for input picture, carry out polygon detecting with this several polygonal distributed model that connects respectively, be with marginal point matching rate the most much higher limit shape in input picture the polygonal type detected.
(4) detection speed is fast, can meet the real-time application demand of road traffic sign detection:
The demand of application in real time should be able to be met for automatic Pilot and the auxiliary method for traffic sign detection driven, the training connecting distributed model in the present invention is that off-line carries out, in testing process, the frontier point mated with it is found in each connection concurrently in edge image, do not interfere with each other, the detection of traffic sign can be completed efficiently.
Accompanying drawing explanation
Fig. 1 is process flow diagram of the present invention;
Fig. 2 is the schematic diagram of the center-border link model of regular polygon;
Fig. 3 is the schematic diagram solving smallest match cost with dynamic programming method;
Fig. 4 is the schematic diagram carrying out octagon coupling with dynamic programming method;
Fig. 5 is the center-border link model schematic diagram of the regular polygon of training gained; Wherein, (a) is octagon, and (b) is circular, and (c) is square, and (d) is triangle;
Fig. 6 is the schematic diagram that Rotational Symmetry projection centre solves;
Fig. 7 is projection result schematic diagram;
Fig. 8 (a) is original image, and (b) is projection centre and projection point set schematic diagram;
Fig. 9 (a) is the result of projection centre and the Point set matching that projects in circular central-contour connection model and Fig. 8, and (b)-(d) is the matching result schematic diagram of center-border link model of octagon, squares and triangles respectively;
Figure 10 is the mark schematic diagram of frontier point on former figure detected;
Wherein (a)-(d) be that speed limit 100 indicates respectively, the testing result of stop sign, Stop and give way mark and children crossing mark.
Embodiment
Below with reference to accompanying drawing, technical scheme of the present invention is described in detail.
As shown in Figure 1, the invention provides a kind of method for traffic sign detection based on center-border link model, traffic sign is utilized to be the characteristic of regular polygon, the detection of traffic sign is converted into polygonal detection, employing center represents polygon to the connection of frontier point, polygon is regarded as the set that center connects to frontier point, the method comprises training and detects two stages, the parameter distribution of each connection is learnt in the training stage, the frontier point mated with it is found respectively in each connection of detection-phase, complete polygonal detection, mainly comprise following detailed step:
(1) training sample coupling: by the border uniform sampling of first training sample, form some frontier points, this polygon is represented to connecting of frontier point with center, then utilize dynamic programming method the connection in first training sample and other training sample to be mated, find out polygon and be respectively connected to corresponding relation in all training samples.
Polygon is regarded as the set that center connects to frontier point, each connection l can use tlv triple (d, θ, α) to represent, wherein, d represents that the distance of frontier point is arrived at center, and θ represents the angle of connection and horizontal direction, and α represents the gradient direction of frontier point, wherein the span of θ is [0,2 π], the span of α is [0, π].
The particular content of described training sample coupling is:
(11) in N number of training sample (polygon), random selecting polygon is as first training sample, to the border uniform sampling of this training sample, obtain n=1000 frontier point, first training sample is expressed as 1000 set that connect of center to frontier point; To all the other sample boundary uniform samplings, obtain m=1500 frontier point, training sample is expressed as 1500 set that connect of center to frontier point.Wherein, the value of m, n is different, and m>n, by Dynamic Matching, in m frontier point, find out n, connect corresponding with n in first training sample.
(12) defining first training sample is P={l
i, 1<i<n, all the other training samples are P'={l'
j, 1<j<m, its center is respectively o and o', the Matching power flow Matrix C=(c of structure n × m
ij), c
ijrepresent l
iand l
j' between diversity factor, be defined as:
c
ij=d(l
i,l
j')=λ
θ*min(|θ
i-θ
j'|,2π-|θ
i-θ
j'|)+
(1)
λ
α*min(|α
i-α
j'|,π-|α
i-α
j'|)+λ
d*|d
i-d
j'|
In formula, | θ
i-θ '
j|, | α
i-α '
j| with | d
i-d'
j| represent respectively and connect l
iand l'
jdifference in the horizontal direction in angle, frontier point gradient direction and length, λ
θ, λ
αand λ
dthen represent the weight that this three species diversity is shared in Matching power flow, value can adopt 0.5,0.2 and 0.3 respectively.
(13) the cost matrix C Matching power flow decided between different connection is defined, in polygon P and P', the objective function of matching connection can be expressed as: π: { 1,, n} → { 0,1, m}, and 1 ... each value in m} can only be mated once, to represent in second training sample not and be connected l with π (i)=0
ithe option matched, the solving target and can be expressed as of polygon matching problem:
For all π (i)=0, c
i, π (i)=ε, namely when sky mates, Matching power flow is ε, ε is an adjustable threshold value, and only have the Matching power flow between two connections to be less than ε, this coupling is just allowed to, and ε gets 0.6 here.
(14) the connection l in first training sample can be determined by polygon coupling
icorrespondence in all the other N-1 training sample connects.
The problem finding smallest match cost can be exchanged into the problem finding minimal cost path in the digraph of (n+1) shown in Fig. 3 × (m+1) size, in digraph, the Matching power flow of vertical edges representative is ε, the Matching power flow of horizontal sides representative is 0, and the weight on diagonal line is the value c in corresponding cost matrix
ij.
In Matching power flow computation process, the establishing method of each parameter is as follows:
A) λ
θ+ λ
α+ λ
d=1, determine the value of wherein two parameters, then another also decides thereupon, first determines λ
θand λ
dvalue.
B) on training set, adopt the mode of traversal to obtain optimized parameter configuration, interval [0.2,0.8] changes λ
θand λ
dvalue, obtain the matching rate under various parameter combinations, get matching rate the highest time parameter configuration, in the present embodiment, at λ
θ=0.5, λ
α=0.2, λ
dwhen=0.3, matching rate is the highest.
C) on [0,1] interval, change the value of ε, obtain corresponding matching rate and wrong matching rate curve, determine the value of ε accordingly.
Fig. 4 shows use (d, θ, α) and represents connection, utilizes dynamic programming method to carry out the result of octagon coupling.
(2) center-border link model training: suppose that each length be connected in all training samples meets Gaussian distribution, the angle of inclination connected and the gradient direction of frontier point meet VonMises distribution (Feng meter Sai Si distribution), utilize and expect that maximum solution estimates the parameter distribution connected, in polygon, each parameter distribution connected constitutes this polygonal center-border link model.The detailed content of this step is:
(21) according to the training sample matching result that step (1) obtains, suppose that the length of each connection in all training samples meets Gaussian distribution, the angle of inclination connected and the gradient direction of frontier point meet VonMises distribution, then i-th model connected can be expressed as:
Wherein, u
dand σ
dcentered by the average of Gaussian distribution of the distance d of frontier point and variance, u
θand k
θ, u
αand k
αbe average and the variance of the VonMises distribution of connection and the angle theta of horizontal direction and the gradient direction α of frontier point respectively, wherein, subscript i represents corresponding to i-th connection.
(22) with expecting that maximum solution estimates the parameter u of Gaussian distribution and VonMises distribution
d, σ
d, u
θ, k
θ, u
αand k
α, the parameter distribution connected can be obtained.
The parameter distribution of all connections constitutes polygonal center-border link model, and in the traffic sign of training gained, common polygonal center-border link model as shown in Figure 5.Wherein, the length of each connection and width are by parameter
with
provide, the direction of connection is by parameter
determine, the direction of each connection terminal represents the gradient direction of frontier point, by parameter
determine
(3) polygon center and size estimation in image to be detected: the marginal information extracting each image to be detected, adopt Rotational Symmetry projecting method, symmetrical marginal point is selected to project to its center of rotational symmetry, the maximum projection centre of brightness can be formed centrally in polygon after projection, obtain the point set projected to this projection centre simultaneously, respectively using this projection centre and point set as the potential center of polygon and frontier point, calculate polygonal dimensional information according to the distance between frontier point.The detailed content of this step is:
(31) ask the edge image of image to be detected, the gradient of edge calculation, ignore the marginal point that amplitude is less than average gradient 0.1.
(32) by the gradient direction scope of pixel (-π, π] be evenly divided into 36 subsets, the scope of each subset is π/18, finds the marginal point that gradient direction meets Rotational Symmetry condition.
Circular, square and octagon all meets the Rotational Symmetry that angle is π, and the marginal point therefore finding gradient direction phase difference of pi projects to its center of rotational symmetry, and namely for a marginal point, its corresponding point are differing in the subset of 18 with it.
Triangle meets the Rotational Symmetry that angle is 2 π/3, and therefore find the marginal point that gradient direction differs 2 π/3 and project to its center of rotational symmetry, namely for a marginal point, its corresponding point are differing in the subset of 12 with it.
(33) for meeting each point of step (32) conditional to finding its center of rotational symmetry, as shown in Figure 6, p
kand p
tbe two symmetrical marginal points, its center of rotational symmetry p can be obtained by following formula:
In formula, (x
k, y
k) be a p
kcoordinate, γ is p
k, p
tline and transverse axis between angle, β is p, p
kline and p
t, p
kthe angle of line, r is p
kor p
tto the distance of center of rotational symmetry p point.
(34) point meeting step (32) conditional projects to its center of rotational symmetry, and the more brightness of point in perspective view of winning the vote is larger, and in polygonal, be formed centrally the maximum subpoint of brightness, projection result as shown in Figure 7.
(35) as shown in Figure 8, record the position of projection centre and subpoint in projection process, utilize following formula to calculate polygonal dimensional information s, wherein Γ (p) represents that the point that the oriented p of institute projects is right, and its quantity is M, || p
kp
t|| represent that point is to p
kand p
tbetween distance.
(4) polygon detecting: using projection centre as the potential center of polygon, subpoint is concentrated and is comprised this polygonal frontier point, each in polygon is connected to the frontier point a little concentrated and find the maximum probability that matches, and allly in polygon center-border link model connects corresponding frontier point and constitutes the polygon detected.The detailed content of this step is:
(41) Rotational Symmetry projecting method can form multiple projection centre in image to be detected, and several maximum projection centre of getting tickets is as the potential center of polygon, and corresponding projection point set is possible frontier point.
(42) for one of them projection centre o
w, remember that the point set to its projection is E={e
w}={ d
w, θ
w, α
w; 1<w<ne}, wherein ne is to o
wcounting of projection.For each some e in E
w, calculate the probability that it belongs to the connection of i-th, polygon.
(43) point that matching probability is maximum is i-th and connects corresponding frontier point, is designated as e
i *.The shape that the polygon then finally detected is made up of n frontier point, i.e. e (o
w)={ e
1 *, e
2 *..., e
n *.
(44) use the center-border link model of circle, octagon, squares and triangles respectively, detect the frontier point that a certain projection centre o is corresponding.Calculate the matching score of often kind of polygon model and frontier point, score the most much higher limit shape is detected polygonal type.
Fig. 9 (a) carries out the result of mating for the Polygonal Boundary in all connections in circular central-contour connection model and Fig. 8, its matching score is 0.9613, Fig. 9 (b)-(d) is respectively the matching result of octagon, squares and triangles, its matching score is respectively 0.7942,0.4208 and 0.1103.
Marked on former figure by the frontier point detected, as shown in Figure 10, conveniently show, former figure processes, and reduces the brightness of image.
Above embodiment is only and technological thought of the present invention is described, can not limit protection scope of the present invention with this, and every technological thought proposed according to the present invention, any change that technical scheme basis is done, all falls within scope.
Claims (7)
1., based on a method for traffic sign detection for center-border link model, it is characterized in that comprising the steps:
(1) training sample coupling: by the border uniform sampling of training sample, form some frontier points, this polygon is represented to connecting of frontier point with center, then utilize dynamic programming method the connection in first training sample and other training sample to be mated, find out polygon and be respectively connected to corresponding relation in all training samples;
(2) center-border link model training: suppose that each length be connected in all training samples meets Gaussian distribution, the angle of inclination connected and the gradient direction of frontier point meet VonMises distribution, utilize and expect that maximum solution estimates the parameter distribution connected, in polygon, each parameter distribution connected constitutes this polygonal center-border link model;
(3) polygon center and size estimation in image to be detected: the marginal information extracting each image to be detected, adopt Rotational Symmetry projecting method, symmetrical marginal point is selected to project to its center of rotational symmetry, the point set obtaining projection centre and project to this projection centre, the center potential as polygon and frontier point, calculate polygonal dimensional information according to the distance between center and frontier point;
(4) polygon detecting: take projection centre as the potential center of polygon, each subpoint that is connected to concentrates the frontier point finding the maximum probability that matches, and in polygon center-border link model, all frontier points connecting correspondence constitute the polygon detected.
2. a kind of method for traffic sign detection based on center-border link model as claimed in claim 1, it is characterized in that: in described step (1), represent that polygonal method refers to set polygon regarded as center and connect to frontier point with center to connecting of frontier point, each connection l tlv triple (d, θ, α) represent, wherein, d represents that the distance of frontier point is arrived at center, and θ represents the angle of connection and horizontal direction, and α represents the gradient direction of frontier point, wherein the span of θ is [0,2 π], the span of α is [0, π].
3. a kind of method for traffic sign detection based on center-border link model as claimed in claim 2, is characterized in that the detailed content of described step (1) is:
(11) in N number of training sample, random selecting polygon, as first training sample, to its border uniform sampling, obtains n frontier point, to connect represent this training sample with center to the n of frontier point; To all the other sample boundary uniform samplings, obtain m frontier point, training sample is expressed as m the set that connect of center to frontier point, and m>n;
(12) remember that first training sample is P={l
i, 1<i<n, all the other training samples are P'={l'
j, 1<j<m, its center is respectively o and o', the Matching power flow Matrix C=(c of structure n × m
ij), c
ijrepresent l
iand l
j' between diversity factor, be defined as:
c
ij=d(l
i,l′
j)=λ
θ*min(|θ
i-θ′
j|,2π-|θ
i-θ′
j|)+λ
α*min(|α
i-α′
j|,π-|α
i-α′
j|)+λ
d*|d
i-d′
j|
In formula, | θ
i-θ '
j|, | α
i-α '
j| with | d
i-d'
j| represent respectively and connect l
iand l'
jdifference in the horizontal direction in angle, frontier point gradient direction and length; λ
θ, λ
αand λ
dthen represent the weight that this three species diversity is shared in Matching power flow;
(13) the cost matrix C Matching power flow decided between different connection is defined, in polygon P and P', the objective function of matching connection is expressed as: π: { 1,, n} → { 0,1, m}, and 1 ... each value in m} can only be mated once, to represent in second training sample not and be connected l with π (i)=0
ithe option matched, the object representation that solves of polygon matching problem is:
For all π (i)=0, c
i, π (i)=ε, ε are adjustable threshold values, and only have the Matching power flow between two connections to be less than ε, this coupling is just allowed to;
(14) the connection l in first training sample is determined by polygon coupling
icorrespondence in all the other N-1 training sample connects.
4. a kind of method for traffic sign detection based on center-border link model as claimed in claim 3, it is characterized in that: in Matching power flow computation process, the establishing method of each parameter is as follows:
A) λ
θ+ λ
α+ λ
d=1, first determine λ
θand λ
dvalue;
B) on training set, adopt the mode of traversal to obtain optimized parameter configuration, interval [0.2,0.8] changes λ
θand λ
dvalue, obtain the matching rate under various parameter combinations, get matching rate the highest time parameter configuration;
C) on [0,1] interval, change the value of ε, obtain corresponding matching rate and wrong matching rate curve, determine the value of ε accordingly.
5. a kind of method for traffic sign detection based on center-border link model as claimed in claim 1, is characterized in that the detailed content of described step (2) is:
(21) according to the training sample matching result that step (1) obtains, suppose that the length of each connection in all training samples meets Gaussian distribution, the angle of inclination connected and the gradient direction of frontier point meet VonMises distribution, then the model representation of this connection is:
Wherein, u
dand σ
dcentered by the average of Gaussian distribution of the distance d of frontier point and variance, u
θand k
θ, u
αand k
αbe average and the variance of the VonMises distribution of connection and the angle theta of horizontal direction and the gradient direction α of frontier point respectively, wherein, subscript i represents corresponding to i-th connection;
(22) with expecting that maximum solution estimates the parameter u of Gaussian distribution and VonMises distribution
d, σ
d, u
θ, k
θ, u
αand k
α, obtain the parameter distribution connected.
6. a kind of method for traffic sign detection based on center-border link model as claimed in claim 1, is characterized in that the detailed content of described step (3) is:
(31) edge image of image to be detected is asked, the gradient of edge calculation;
(32) by the gradient direction scope of pixel (-π, π] be evenly divided into 36 subsets, the scope of each subset is π/18, finds the point that gradient direction meets Rotational Symmetry condition;
(33) for meeting each point of step (32) conditional to finding its center of rotational symmetry, if p
kand p
tbe two symmetrical marginal points, its center of rotational symmetry p is obtained by following formula:
In formula, (x
k, y
k) be a p
kcoordinate, γ is p
k, p
tline and transverse axis between angle, β is p, p
kline and p
t, p
kthe angle of line, r is p
kor p
tto the distance of center of rotational symmetry p point;
(34) point meeting step (32) conditional projects to its center of rotational symmetry, and the more brightness of point in perspective view of winning the vote is larger;
(35) in projection process, record the position of projection centre and subpoint, utilize following formula to calculate polygonal dimensional information s:
Wherein, Γ (p) represents that the point that the oriented p of institute projects is right, and its quantity is M, ‖ p
kp
t‖ represents a little to p
kand p
tbetween distance.
7. a kind of method for traffic sign detection based on center-border link model as claimed in claim 1, is characterized in that the detailed content of described step (4) is:
(41) adopt Rotational Symmetry projecting method to form multiple projection centre in image to be detected, several maximum projection centre of getting tickets is as the potential center of polygon, and corresponding projection point set is possible frontier point;
(42) for one of them projection centre o
w, remember that the point set to its projection is E={e
w}={ d
w, θ
w, α
w; 1<w<ne}, wherein ne is to o
wcounting of projection; For each some e in E
w, calculate the probability that it belongs to the connection of i-th, polygon:
(43) point that matching probability is maximum is i-th and connects corresponding frontier point, is designated as
the shape that the polygon then finally detected is made up of n frontier point, namely
(44) the center-border link model of circle, octagon, squares and triangles is used respectively, detect the frontier point that a certain projection centre is corresponding, calculate the matching score of often kind of polygon model and frontier point, score the most much higher limit shape is detected polygonal type.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510644535.7A CN105320940B (en) | 2015-10-08 | 2015-10-08 | A kind of method for traffic sign detection based on center-border link model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510644535.7A CN105320940B (en) | 2015-10-08 | 2015-10-08 | A kind of method for traffic sign detection based on center-border link model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105320940A true CN105320940A (en) | 2016-02-10 |
CN105320940B CN105320940B (en) | 2019-02-12 |
Family
ID=55248295
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510644535.7A Active CN105320940B (en) | 2015-10-08 | 2015-10-08 | A kind of method for traffic sign detection based on center-border link model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105320940B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107633502A (en) * | 2017-07-27 | 2018-01-26 | 西北工业大学 | A kind of target center recognition methods of peg-in-hole assembly automatic centering |
CN109389024A (en) * | 2018-01-30 | 2019-02-26 | 长城汽车股份有限公司 | Method, apparatus, storage medium and vehicle based on image identification road cone |
WO2021013227A1 (en) * | 2019-07-25 | 2021-01-28 | 长城汽车股份有限公司 | Image processing method and apparatus for target detection |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102074005A (en) * | 2010-12-30 | 2011-05-25 | 杭州电子科技大学 | Interest-region-oriented stereo matching method |
CN104657734A (en) * | 2015-03-05 | 2015-05-27 | 东方网力科技股份有限公司 | Traffic sign detecting method and device |
-
2015
- 2015-10-08 CN CN201510644535.7A patent/CN105320940B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102074005A (en) * | 2010-12-30 | 2011-05-25 | 杭州电子科技大学 | Interest-region-oriented stereo matching method |
CN104657734A (en) * | 2015-03-05 | 2015-05-27 | 东方网力科技股份有限公司 | Traffic sign detecting method and device |
Non-Patent Citations (2)
Title |
---|
DAN XU 等: "Real time road sign detection based on ratational center voting and shape analysis", 《MECHATRONICS AND AUTOMATION(ICMA),2012 INTERNATIONAL CONFERENCE ON.IEEE》 * |
李伦波: "自然场景下交通标志的检测与分类算法研究", 《中国博士学位论文全文数据库信息科技辑》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107633502A (en) * | 2017-07-27 | 2018-01-26 | 西北工业大学 | A kind of target center recognition methods of peg-in-hole assembly automatic centering |
CN107633502B (en) * | 2017-07-27 | 2020-09-29 | 西北工业大学 | Target center identification method for automatic centering of shaft hole assembly |
CN109389024A (en) * | 2018-01-30 | 2019-02-26 | 长城汽车股份有限公司 | Method, apparatus, storage medium and vehicle based on image identification road cone |
CN109389024B (en) * | 2018-01-30 | 2020-09-04 | 长城汽车股份有限公司 | Method and device for identifying road cone based on image, storage medium and vehicle |
US11256929B2 (en) | 2018-01-30 | 2022-02-22 | Great Wall Motor Company Limited | Image-based road cone recognition method and apparatus, storage medium, and vehicle |
WO2021013227A1 (en) * | 2019-07-25 | 2021-01-28 | 长城汽车股份有限公司 | Image processing method and apparatus for target detection |
Also Published As
Publication number | Publication date |
---|---|
CN105320940B (en) | 2019-02-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103727930B (en) | A kind of laser range finder based on edge matching and camera relative pose scaling method | |
CN105023008B (en) | The pedestrian of view-based access control model conspicuousness and multiple features recognition methods again | |
Kong et al. | Vanishing point detection for road detection | |
CN101814144B (en) | Water-free bridge target identification method in remote sensing image | |
CN108038846A (en) | Transmission line equipment image defect detection method and system based on multilayer convolutional neural networks | |
CN110379004B (en) | Method for classifying ground features and extracting single objects of oblique photography results | |
CN103164858A (en) | Adhered crowd segmenting and tracking methods based on superpixel and graph model | |
CN102831427B (en) | Texture feature extraction method fused with visual significance and gray level co-occurrence matrix (GLCM) | |
CN103996195A (en) | Image saliency detection method | |
CN104463866A (en) | Local shape matching method based on outline random sampling | |
Durand-Petiteville et al. | Tree detection with low-cost three-dimensional sensors for autonomous navigation in orchards | |
CN104463856A (en) | Outdoor scene three-dimensional point cloud data ground extraction method based on normal vector ball | |
Huang et al. | Rule-based roof plane detection and segmentation from laser point clouds | |
CN106127205A (en) | A kind of recognition methods of the digital instrument image being applicable to indoor track machine people | |
CN103886589A (en) | Goal-oriented automatic high-precision edge extraction method | |
CN102867195B (en) | Method for detecting and identifying a plurality of types of objects in remote sensing image | |
CN111145228A (en) | Heterogeneous image registration method based on local contour point and shape feature fusion | |
CN103164693B (en) | A kind of monitor video pedestrian detection matching process | |
CN102609938A (en) | Method for detecting vanishing points of road based on single image | |
Ding et al. | Efficient vanishing point detection method in complex urban road environments | |
CN105469111A (en) | Small sample set object classification method on basis of improved MFA and transfer learning | |
CN104063711A (en) | Corridor vanishing point rapid detection algorithm based on K-means method | |
CN105279769A (en) | Hierarchical particle filtering tracking method combined with multiple features | |
CN103400399A (en) | Spatial moment based line structured light center extraction method | |
CN105320940A (en) | Traffic sign detection method based on center-boundary connection model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
CB03 | Change of inventor or designer information | ||
CB03 | Change of inventor or designer information |
Inventor after: Xu Dan Inventor after: Hu Chunlong Inventor after: Liu Qinghua Inventor after: Zhu Xia Inventor after: Zuo Xin Inventor before: Xu Dan Inventor before: Liu Qinghua Inventor before: Zhu Xia Inventor before: Zuo Xin |
|
GR01 | Patent grant | ||
GR01 | Patent grant |