CN105320940B - A kind of method for traffic sign detection based on center-border link model - Google Patents
A kind of method for traffic sign detection based on center-border link model Download PDFInfo
- Publication number
- CN105320940B CN105320940B CN201510644535.7A CN201510644535A CN105320940B CN 105320940 B CN105320940 B CN 105320940B CN 201510644535 A CN201510644535 A CN 201510644535A CN 105320940 B CN105320940 B CN 105320940B
- Authority
- CN
- China
- Prior art keywords
- polygon
- center
- point
- connection
- boundary point
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
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 kind of method for traffic sign detection based on center-border link model, and step is: being matched using dynamic programming method to the connection in training sample, finds out the corresponding relationship that polygon is respectively connected in all training samples;Assuming that the length being each connected in all training samples meets Gaussian Profile, the tilt angle of connection and the gradient direction of boundary point meet Von Mises distribution, using the parameter distribution of desired maximum solution estimation connection, the parameter distribution respectively connected in polygon constitutes the center-border link model of the polygon;The potential center of polygon and boundary point are obtained using rotational symmetry projecting method;Using projection centre as the potential center of polygon, the corresponding boundary point of all connections constitutes the polygon detected in polygon center-border link model.Such detection method can realize the quick detection of traffic sign.
Description
Technical field
The invention belongs to image procossing and technical field of computer vision, in particular to a kind of detection side of traffic sign
Method.
Background technique
Road traffic sign detection is an important research direction in computer vision.As road safety affiliated facility, hand over
Logical mark specification traffic behavior, instruction condition of road surface, ensure road effect, guidance pedestrian and in terms of play weight
It acts on.As the key technology of intelligent transportation system, road traffic sign detection is automatic Pilot, auxiliary drives and traffic sign dimension
The indispensable component parts of systems such as shield.
Road traffic sign detection is color or shape feature using traffic sign by the traffic sign region detection in image
And split, detection directly affects the precision of subsequent traffic sign classification whether accurate.
Method for traffic sign detection is broadly divided into two kinds: the method based on color characteristic and the side based on shape feature
Method.Method for traffic sign detection based on color characteristic is that rule of thumb threshold value or other colors are special in specific color space
Sign divides the image into traffic sign region and non-traffic sign region.This method should be readily appreciated that and realize, can rapidly exclude
Background area.However, colouring information is easy to be influenced by extraneous factor, such as color degradation, shade and the light of traffic sign itself
It will cause declining to a great extent for detection accuracy according to variation.Therefore, reality is individually unable to satisfy using the testing result of color characteristic
The requirement of application.Method for traffic sign detection based on shape is in regular polygon using traffic sign mostly characterized by edge
Characteristic, convert the detection of traffic sign to detection (such as Gonz á lez of regular polygonGarrido M A,Llorca
D F,et al.Automatic traffic signs and panels inspection system using computer
vision[J].IEEE Transactions on Intelligent Transportation Systems,2011,12(2):
485-499.)。
Generalized Hough Transform (Generalized Hough transformation, GHT) ([2] Davies E
R.Machine Vision:Theory,Algorithms,Practicalities(Third Edition).San
Francisco:Elsevier, 2005.387-410) using the geometrical property of polygon, by the test problems of variable space figure
It is converted into the clustering problem of parameter space, realizes the direct detection of polygon.Its main feature is that it is simple direct, but due to calculation amount
Greatly, generally it is only applicable to the less polygon detecting of the number of edges such as triangle.
The method that the propositions such as Barnes detect circle using radial symmetry transform, the method is expanded to octagon, square later
Shape and triangle (BARNES N, ZELINSKY A.Real-time radial symmetry for speed sign
detection[J].Intelligent Vehicles Symposium,2004,566-571).This method is using symmetrical
Marginal point to its radial symmetric central projection, can be bright in being centrally formed for circular traffic sign after the projection of all marginal points
Maximum projection centre is spent, thus can determine the specific location of traffic sign.Radial symmetry transform is a kind of univariate conversion,
Each marginal point is individually voted to multiple accumulators independently of its neighborhood, and which results in higher false alarm rates.
Above method is substantially all based on the polygon detecting method of projection, and Barnes etc. is proposed based on posterior probability
Regular polygon detection method (Barnes N, Loy G, Shaw D.The regular polygon detector [J]
.Pattern Recognition,2010,43(3):592-602).This method extracts edge image first, then according to polygon
The geometrical characteristic of shape, the probability density function that polygon is defined using posterior probability.Compared to projecting method, this method has higher
Verification and measurement ratio and lower false alarm rate, however this method is more demanding to edge quality, detects in the incomplete situation in edge
Effect decline is very fast.
Based on the above analysis, effectively traffic sign can be detected there is no a kind of method at present, the present inventor
This case is exactly researched and developed under this research background.
Summary of the invention
The purpose of the present invention is to provide a kind of method for traffic sign detection based on center-border link model, can
The quick detection for realizing traffic sign can be detected effectively and the polygon that visual angle inclination is lacked with marginal portion occurs, and when
Between complexity it is lower.
In order to achieve the above objectives, solution of the invention is:
A kind of method for traffic sign detection based on center-border link model, includes the following steps:
(1) training sample matches: by the boundary uniform sampling of training sample, several boundary points is formed, with center to boundary
The connection of point indicates the polygon, then utilizes dynamic programming method by the company in first training sample and other training samples
Capable matching is tapped into, the corresponding relationship that polygon is respectively connected in all training samples is found out;
(2) center-border link model training: assuming that the length being each connected in all training samples meets Gauss point
Cloth, the tilt angle of connection and the gradient direction of boundary point meet Von Mises distribution, estimate to connect using desired maximum solution
The parameter distribution connect, the parameter distribution respectively connected in polygon constitute the center-border link model of the polygon;
(3) polygon center and size estimation in image to be detected: extracting the marginal information of each image to be detected, uses
Rotational symmetry projecting method, select symmetrical marginal point to its center of rotational symmetry project, obtain projection centre and to this
The point set of projection centre projection is counted as the potential center of polygon and boundary point according to the distance between center and boundary point
Calculate the dimensional information of polygon;
(4) polygon detecting: using projection centre as the potential center of polygon, be each connected to subpoint concentrate find with
The maximum boundary point of matching probability, the corresponding boundary point of all connections constitutes inspection in polygon center-border link model
The polygon measured.
In above-mentioned steps (1), referred in regarding polygon as with the method that the connection at center to boundary point indicates polygon
The set that the heart is connected to boundary point, each connection l are indicated with triple (d, θ, α), wherein d indicate center to boundary point away from
From θ indicates that the angle of connection and horizontal direction, α indicate the gradient direction of boundary point, and wherein the value range of θ is [0,2 π], α
Value range be [0, π].
The detailed content of above-mentioned steps (1) is:
(11) in N number of training sample, a polygon is randomly selected as first training sample, it is uniform to its boundary
Sampling, obtains n boundary point, indicates the training sample with the n connection at center to boundary point;Remaining sample boundary is uniformly adopted
Sample obtains m boundary point, and the set of m connection of boundary point, and m > n are arrived centered on training sample is indicated;
(12) remember that first training sample is P={ li, 1 < i < n, remaining training sample is P'={ l'j, 1 < j < m, wherein
The heart is respectively o and o', constructs matching cost Matrix C=(c of n × mij), cijIndicate liAnd lj' between diversity factor, is defined as:
cij=d (li,lj')=λθ*min(|θi-θj'|,2π-|θi-θj'|)+
λα*min(|αi-αj'|,π-|αi-αj'|)+λd*|di-dj'|
In formula, | θi-θ'j|、|αi-α'j| and | di-d'j| respectively indicate connection liAnd l'jAngle, boundary point in the horizontal direction
Difference on gradient direction and length;λθ, λαAnd λdThen indicate these three differences weight shared in matching cost;
(13) it defines cost matrix C and is used to determine the matching cost between different connections, matching connection in polygon P and P'
Objective function indicate are as follows: π: { 1 ..., n } → { 0,1 ..., m }, and each value in { 1 ..., m } can only match once, use π
(i)=0 indicate in second training sample not with connect liThe option to match, the solution target of polygon matching problem
It indicates are as follows:
For all π (i)=0, ci,π(i)=ε, ε are an adjustable threshold values, and only there are two the matchings between connection
Cost is less than ε, and this matching is just allowed to;
(14) the connection l determined in first training sample is matched by polygoniIn remaining N-1 training sample
It is correspondingly connected with.
In matching cost calculating process, the setting method of each parameter is as follows:
a)λθ+λα+λd=1, it is first determined λθAnd λdValue;
B) optimized parameter configuration is obtained by the way of traversal on training set, changes λ on section [0.2,0.8]θAnd λd
Value, obtain various parameters combination under matching rate, take parameter configuration when matching rate highest;
C) value for changing ε on [0,1] section obtains corresponding matching rate and wrong matching rate curve, determines therefrom that ε's
Value.
The detailed content of above-mentioned steps (2) is:
(21) the training sample matching result obtained according to step (1), it is assumed that the length of each connection in all training samples
Degree meets Gaussian Profile, the tilt angle of connection and the gradient direction of boundary point meet Von Mises distribution, then the connection
Model is expressed as:
Wherein, udAnd σdCentered on to boundary point distance d Gaussian Profile mean value and variance, uθAnd kθ、uαAnd kαPoint
It is not the mean value and variance for connecting the Von Mises distribution of the gradient direction α with the angle theta of horizontal direction and boundary point,
In, subscript i indicates to correspond to i-th of connection;
(22) with the parameter u of desired maximum solution estimation Gaussian Profile and Von Mises distributiond、σd、uθ、kθ、uαAnd kα,
The parameter distribution connected.
The detailed content of above-mentioned steps (3) is:
(31) edge image for seeking image to be detected calculates the gradient at edge;
(32) by the gradient direction range of pixel (- π, π] be uniformly divided into 36 subsets, the range of each subset is π/18,
Find the point that gradient direction meets rotational symmetry condition;
(33) its center of rotational symmetry is found for meeting each pair of point of step (32) conditional, if pkAnd ptIt is two
Symmetrical marginal point, center of rotational symmetry p are obtained by following formula:
In formula, (xk,yk) it is point pkCoordinate, γ is pk, ptLine and trunnion axis between angle, β is p, pkLine
With pt, pkThe angle of line, r are pkOr ptTo the distance of center of rotational symmetry p point;
(34) point for meeting step (32) conditional is projected to its center of rotational symmetry, wins the vote more points in perspective view
Brightness it is bigger;
(35) position that projection centre and subpoint are recorded in projection process calculates the scale of polygon using following formula
Information s:
Wherein, Γ (p) indicates the point pair that the oriented p of institute is projected, quantity M, | | pkpt| | indicate point to pkAnd ptBetween
Distance.
The detailed content of above-mentioned steps (4) is:
(41) multiple projection centres are formed in image to be detected using rotational symmetry projecting method, several most throwings of getting tickets
As the potential center of polygon, the corresponding point set that projects is possible boundary point at shadow center;
(42) for one of projection centre ow, remember that the point set projected to it is E={ ew}={ dw,θw,αw;1<w<
Ne }, wherein ne is to owThe points of projection;For each of E point ew, calculate it and belong to the general of i-th of polygon connection
Rate:
(43) the maximum point of matching probability is the corresponding boundary point of i-th of connection, is denoted as ei *, then what is eventually detected is more
The shape that side shape is made of n boundary point, i.e. e (ow)={ e1 *,e2 *,…,en *,
(44) respectively with circle, octagon, squares and triangles center-border link model, detect a certain projection
The corresponding boundary point in center calculates the matching score of every kind of polygon model and boundary point, and the polygon of highest scoring is quilt
Detect the type of polygon.
After adopting the above scheme, the present invention utilizes the rules such as traffic sign triangular in shape, square, octagon and circle
The characteristic of polygon converts road traffic sign detection to the detection of regular polygon, has the following characteristics that
(1) insensitive for the visual angle inclination and deformation of polygon:
Traffic sign under natural scene will receive artificial or weather conditions influences, and deformation occurs, in image acquisition process
Visual angle inclination can also make collected traffic sign deformation occurs, the present invention is polygon using the connection expression at center to boundary point
Shape, when deformation occurs for polygon, the adjacent sequence between connection is constant, can effectively detect that deformation occurs and visual angle is inclined
Traffic sign.
It (2) can the effective traffic sign that blocks of detection part:
Outdoor traffic sign is often blocked by other objects and forms incomplete edge, the detection based on projection
Method is strongly depend on the quality at edge, respectively connects in the present invention and finds matched boundary point in marginal point respectively, right
The continuity requirement at edge is lower, is able to detect the polygon of marginal portion missing, the i.e. traffic sign of partial occlusion.
(3) Shape Classification of polygon is completed in the detection process:
Common polygon has triangle, square, octagon and circle in traffic sign, and the present invention is directed to these respectively
Polygon training connection distributed model carries out polygon input picture with the connection distributed model of these types of polygon respectively
Shape detection is the type of the polygon detected with the highest polygon of marginal point matching rate in input picture.
(4) detection speed is fast, can satisfy the real-time application demand of road traffic sign detection:
The method for traffic sign detection driven for automatic Pilot and auxiliary should be able to meet the needs of applying in real time, the present invention
The training of middle connection distributed model carries out offline, in the detection process each connection concurrently find in edge image and its
Matched boundary point, does not interfere with each other, and can be efficiently completed the detection of traffic sign.
Detailed description of the invention
Fig. 1 is flow chart of the invention;
Fig. 2 is the schematic diagram of the center-border link model of regular polygon;
Fig. 3 is the schematic diagram that smallest match cost is solved with dynamic programming method;
Fig. 4 is to carry out the matched schematic diagram of octagon with dynamic programming method;
Fig. 5 is the center-border link model schematic diagram of trained resulting regular polygon;Wherein, (a) is octagon,
(b) it is circle, is (c) square, is (d) 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, is (b) projection centre and projection point set schematic diagram;
Fig. 9 (a) be in circular central-contour connection model and Fig. 8 projection centre and project point set it is matched as a result, (b)-
(d) be respectively octagon, squares and triangles center-border link model matching result schematic diagram;
Figure 10 is the mark schematic diagram of the boundary point that detects in original image;
Wherein (a)-(d) be respectively 100 mark of speed limit, stop sign, stop sign and children crossing mark inspection
Survey result.
Specific embodiment
Below with reference to attached drawing, technical solution of the present invention is described in detail.
As shown in Figure 1, the present invention provides a kind of method for traffic sign detection based on center-border link model, utilize
Traffic sign is in the characteristic of regular polygon, the detection of traffic sign is converted to the detection of polygon, using center to boundary
The connection of point indicates polygon, regards polygon as set that center is connected to boundary point, this method includes training and detection two
A stage learns the parameter distribution of each connection in the training stage, finds respectively in each connection of detection-phase matched
Boundary point completes the detection of polygon, mainly includes following detailed step:
(1) training sample matches: by the boundary uniform sampling of first training sample, forming several boundary points, uses center
Connection to boundary point indicates the polygon, then utilizes dynamic programming method by first training sample and other training samples
In connection matched, find out the corresponding relationship that polygon is respectively connected in all training samples.
Regarding polygon as set that center is connected to boundary point, each connection l can be indicated with triple (d, θ, α),
In, d indicates center to the distance of boundary point, and θ indicates to connect and the angle of horizontal direction, α indicate the gradient direction of boundary point,
The value range of middle θ is [0,2 π], and the value range of α is [0, π].
The matched particular content of training sample is:
(11) in N number of training sample (polygon), a polygon is randomly selected as first training sample, to this
The boundary uniform sampling of training sample, obtains n=1000 boundary point, will arrive boundary point centered on first training sample expression
1000 connection set;To remaining sample boundary uniform sampling, m=1500 boundary point is obtained, training sample is indicated
Centered on to 1500 of boundary point connections set.Wherein, the value of m, n are different, and m > n, by Dynamic Matching, at m
It is a that n is found out in boundary point, it is corresponding with n connection in first training sample.
(12) defining first training sample is P={ li, 1 < i < n, remaining training sample is P'={ l'j, 1 < j < m,
Center is respectively o and o', constructs matching cost Matrix C=(c of n × mij), cijIndicate liAnd lj' between diversity factor, definition
Are as follows:
cij=d (li,lj')=λθ*min(|θi-θj'|,2π-|θi-θj'|)+
(1)
λα*min(|αi-αj'|,π-|αi-αj'|)+λd*|di-dj'|
In formula, | θi-θ'j|、|αi-α'j| and | di-d'j| respectively indicate connection liAnd l'jAngle, boundary point in the horizontal direction
Difference on gradient direction and length, λθ, λαAnd λdThen indicate these three differences weight shared in matching cost, value
0.5,0.2 and 0.3 can be respectively adopted.
(13) it defines cost matrix C and is used to determine the matching cost between different connections, matching connection in polygon P and P'
Objective function can indicate are as follows: π: { 1 ..., n } → { 0,1 ..., m }, and each value in { 1 ..., m } can only match once,
With π (i)=0 indicate in second training sample not with connect liThe option to match, the solution of polygon matching problem
Target may be expressed as:
For all π (i)=0, ci,π(i)=ε, i.e., in empty matched situation, matching cost ε, ε are one adjustable
The threshold value of section, only there are two the matching costs between connection to be less than ε, and this matching is just allowed to, and ε takes 0.6 here.
(14) the connection l in first training sample can be determined by polygon matchingiIn remaining N-1 training sample
In be correspondingly connected with.
The problem of finding smallest match cost can be exchanged into seeks in the digraph of (n+1) shown in Fig. 3 × (m+1) size
The problem of looking for minimal cost path, for the matching cost that vertical edges represent in digraph as ε, the matching cost that horizontal sides represent is 0,
Weight on diagonal line is the value c in corresponding cost matrixij。
In matching cost calculating process, the setting method of each parameter is as follows:
a)λθ+λα+λd=1, it is determined that the value of two of them parameter, then another is also decided therewith, it is first determined
λθAnd λdValue.
B) optimized parameter configuration is obtained by the way of traversal on training set, changes λ on section [0.2,0.8]θAnd λd
Value, obtain various parameters combination under matching rate, parameter configuration when matching rate highest is taken, in the present embodiment, in λθ
=0.5, λα=0.2, λdMatching rate highest when=0.3.
C) value for changing ε on [0,1] section obtains corresponding matching rate and wrong matching rate curve, determines therefrom that ε's
Value.
Fig. 4 shows that use (d, θ, α) indicates connection, carries out the matched result of octagon using dynamic programming method.
(2) center-border link model training: assuming that the length being each connected in all training samples meets Gauss point
Cloth, the tilt angle of connection and the gradient direction of boundary point meet Von Mises distribution (von mises distribution), utilize expectation
Maximum solution estimates that the parameter distribution of connection, the parameter distribution respectively connected in polygon constitute the center-border of the polygon
Link model.The detailed content of the step is:
(21) the training sample matching result obtained according to step (1), it is assumed that the length of each connection in all training samples
Degree meets Gaussian Profile, and the tilt angle of connection and the gradient direction of boundary point meet Von Mises distribution, then i-th of company
The model connect may be expressed as:
Wherein, udAnd σdCentered on to boundary point distance d Gaussian Profile mean value and variance, uθAnd kθ、uαAnd kαPoint
It is not the mean value and variance for connecting the Von Mises distribution of the gradient direction α with the angle theta of horizontal direction and boundary point,
In, subscript i indicates to correspond to i-th of connection.
(22) with the parameter u of desired maximum solution estimation Gaussian Profile and Von Mises distributiond、σd、uθ、kθ、uαAnd kα,
The parameter distribution of connection can be obtained.
The parameter distribution of all connections constitutes the center-border link model of polygon, the resulting traffic sign of training
In common polygon center-border link model it is as shown in Figure 5.Wherein, the length and width of each connection is by parameterWithIt provides, the direction of connection is by parameterIt determines, the direction of each connection terminal represents the gradient direction of boundary point, by joining
NumberIt determines
(3) polygon center and size estimation in image to be detected: extracting the marginal information of each image to be detected, uses
Rotational symmetry projecting method selects symmetrical marginal point to project to its center of rotational symmetry, can be in polygon after projection
Be centrally formed the maximum projection centre of brightness, while obtaining the point set projected to the projection centre, respectively by the projection centre and
Point set calculates the dimensional information of polygon according to the distance between boundary point as the potential center of polygon and boundary point.It should
The detailed content of step is:
(31) edge image for seeking image to be detected calculates the gradient at edge, ignores the side that amplitude is less than average gradient 0.1
Edge point.
(32) by the gradient direction range of pixel (- π, π] be uniformly divided into 36 subsets, the range of each subset is π/18,
Find the marginal point that gradient direction meets rotational symmetry condition.
Round, square and octagon are all satisfied the rotational symmetry that angle is π, therefore find the side of gradient direction phase difference of pi
Edge point is projected to its center of rotational symmetry, i.e., for a marginal point, corresponding points are in the subset for differing from it by 18.
Triangle meets the rotational symmetry that angle is 2 π/3, therefore the marginal point for finding 2 π/3 of gradient direction difference is revolved to it
Turn symmetrical centre projection, i.e., for a marginal point, corresponding points are in the subset for differing from it by 12.
(33) its center of rotational symmetry is found for meeting each pair of point of step (32) conditional, as shown in fig. 6, pkWith
ptIt is two symmetrical marginal points, center of rotational symmetry p can be obtained by following formula:
In formula, (xk,yk) it is point pkCoordinate, γ is pk, ptLine and trunnion axis between angle, β is p, pkLine
With pt, pkThe angle of line, r are pkOr ptTo the distance of center of rotational symmetry p point.
(34) point for meeting step (32) conditional is projected to its center of rotational symmetry, wins the vote more points in perspective view
Brightness it is bigger, be centrally formed the maximum subpoint of brightness in polygon, projection result is as shown in Figure 7.
(35) it as shown in figure 8, recording the position of projection centre and subpoint in projection process, is calculated using following formula more
The dimensional information s of side shape, wherein Γ (p) indicates the point pair that the oriented p of institute is projected, quantity M, | | pkpt| | indicate point to pkWith
ptThe distance between.
(4) polygon detecting: using projection centre as the potential center of polygon, subpoint is concentrated comprising the polygon
Boundary point, each of polygon are connected to the boundary point for concentrating the matching maximum probability of searching, polygon center-border
The corresponding boundary point of all connections constitutes the polygon detected in link model.The detailed content of the step is:
(41) rotational symmetry projecting method can form multiple projection centres in image to be detected, several most projections of getting tickets
As the potential center of polygon, the corresponding point set that projects is possible boundary point at center.
(42) for one of projection centre ow, remember that the point set projected to it is E={ ew}={ dw,θw,αw;1<w<
Ne }, wherein ne is to owThe points of projection.For each of E point ew, calculate it and belong to the general of i-th of polygon connection
Rate.
(43) the maximum point of matching probability is the corresponding boundary point of i-th of connection, is denoted as ei *.What is then eventually detected is more
The shape that side shape is made of n boundary point, i.e. e (ow)={ e1 *,e2 *,…,en *}。
(44) respectively with circle, octagon, squares and triangles center-border link model, detect a certain projection
The corresponding boundary point of center o.The matching score of every kind of polygon model and boundary point is calculated, the polygon of highest scoring is quilt
Detect the type of polygon.
Fig. 9 (a) is that the Polygonal Boundary progress in circular central-contour connection model in all connections and Fig. 8 is matched
As a result, its matching score is the matching result that 0.9613, Fig. 9 (b)-(d) is respectively octagon, squares and triangles,
It is respectively 0.7942,0.4208 and 0.1103 with score.
The boundary point that will test out is labeled in original image, as shown in Figure 10, is shown for convenience, original image has all done place
Reason, reduces the brightness of image.
The above examples only illustrate the technical idea of the present invention, and this does not limit the scope of protection of the present invention, all
According to the technical idea provided by the invention, any changes made on the basis of the technical scheme each falls within the scope of the present invention
Within.
Claims (5)
1. a kind of method for traffic sign detection based on center-border link model, it is characterised in that include the following steps:
(1) training sample matches: by the boundary uniform sampling of training sample, forming several boundary points, arrives boundary point with center
Connection indicates polygon, is then carried out the connection in first training sample and other training samples using dynamic programming method
Matching, finds out the corresponding relationship that polygon is respectively connected in all training samples;
(2) center-border link model training: assuming that the length being each connected in all training samples meets Gaussian Profile,
The tilt angle of connection and the gradient direction of boundary point meet Von Mises distribution, estimate to connect using desired maximum solution
Parameter distribution, the parameter distribution respectively connected in polygon constitutes the center-border link model of the polygon;
(3) polygon center and size estimation in image to be detected: the marginal information of each image to be detected is extracted, using rotation
Symmetrical projecting method selects symmetrical marginal point to project to its center of rotational symmetry, obtains projection centre and to the projection
The point set of central projection calculates more as the potential center of polygon and boundary point according to the distance between center and boundary point
The dimensional information of side shape;
(4) it polygon detecting: using projection centre as the potential center of polygon, is each connected to subpoint and concentrates and find therewith
Boundary point with maximum probability, the corresponding boundary point of all connections, which constitutes, in polygon center-border link model detects
Polygon.
2. a kind of method for traffic sign detection based on center-border link model as described in claim 1, feature exist
In: in the step (1), is referred to the method that the connection at center to boundary point indicates polygon and regard polygon as center to side
The set of boundary's point connection, each connection l are indicated with triple (d, θ, α), wherein d indicates distance of the center to boundary point, θ table
Show that the angle of connection and horizontal direction, α indicate the gradient direction of boundary point, wherein the value range of θ is [0,2 π], the value of α
Range is [0, π].
3. a kind of method for traffic sign detection based on center-border link model as described in claim 1, it is characterised in that
The detailed content of the step (2) is:
(21) the training sample matching result obtained according to step (1), it is assumed that the length of each connection in all training samples accords with
Gaussian Profile is closed, the tilt angle of connection and the gradient direction of boundary point meet Von Mises distribution, then the model of the connection
It indicates are as follows:
Wherein, udAnd σdCentered on to boundary point distance d Gaussian Profile mean value and variance, uθAnd kθ、uαAnd kαIt is respectively
The mean value and variance of the Von Mises distribution of the gradient direction α of the angle theta and boundary point of connection and horizontal direction, wherein on
Marking i indicates to correspond to i-th of connection;
(22) with the parameter u of desired maximum solution estimation Gaussian Profile and Von Mises distributiond、σd、uθ、kθ、uαAnd kα, obtain
The parameter distribution of connection.
4. a kind of method for traffic sign detection based on center-border link model as described in claim 1, it is characterised in that
The detailed content of the step (3) is:
(31) edge image for seeking image to be detected calculates the gradient at edge;
(32) by the gradient direction range of pixel (- π, π] be uniformly divided into 36 subsets, the range of each subset is π/18, is found
Gradient direction meets the point of rotational symmetry condition;
(33) its center of rotational symmetry is found for meeting each pair of point of step (32) conditional, if pkAnd ptIt is two symmetrical
Marginal point, center of rotational symmetry p obtains by following formula:
In formula, (xk,yk) it is point pkCoordinate, γ is pk, ptLine and trunnion axis between angle, β is p, pkLine and pt,
pkThe angle of line, r are pkOr ptTo the distance of center of rotational symmetry p point;
(34) point for meeting step (32) conditional is projected to its center of rotational symmetry, and more point of winning the vote is bright in perspective view
It spends bigger;
(35) position that projection centre and subpoint are recorded in projection process calculates the dimensional information of polygon using following formula
S:
Wherein, Γ (p) indicates the point pair that the oriented p of institute is projected, quantity M, | | pkpt| | indicate point to pkAnd ptThe distance between.
5. a kind of method for traffic sign detection based on center-border link model as described in claim 1, it is characterised in that
The detailed content of the step (4) is:
(41) multiple projection centres are formed in image to be detected using rotational symmetry projecting method, in several most projections of getting tickets
For the heart as the potential center of polygon, the corresponding point set that projects is possible boundary point;
(42) for one of projection centre ow, remember that the point set projected to it is E={ ew}={ dw,θw,αw;1 < w <
Ne }, wherein ne is to owThe points of projection;For each of E point ew, calculate it and belong to the general of i-th of polygon connection
Rate:
Wherein, s indicates the dimensional information of polygon, liIndicate i-th of connection of polygon;udAnd σdCentered on arrive boundary point
The mean value and variance of the Gaussian Profile of distance d, uθAnd kθ、uαAnd kαIt is the angle theta and boundary point of connection and horizontal direction respectively
Gradient direction α Von Mises distribution mean value and variance, wherein subscript i indicate correspond to i-th connection;dw、θw、αw
For projection centre owTo point ewConnection triple indicate;
(43) the maximum point of matching probability is the corresponding boundary point of i-th of connection, is denoted as ei *, then the polygon that eventually detects
The shape being made of n boundary point, i.e. e (ow)={ e1 *,e2 *,…,en *,
(44) respectively with circle, octagon, squares and triangles center-border link model, detect a certain projection centre
Corresponding boundary point, calculates the matching score of every kind of polygon model and boundary point, and the polygon of highest scoring is to be detected
The type of polygon.
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 CN105320940A (en) | 2016-02-10 |
CN105320940B true 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) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
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 |
CN111539907B (en) * | 2019-07-25 | 2023-09-12 | 毫末智行科技有限公司 | Image processing method and device 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 |
---|
Real time road sign detection based on ratational center voting and shape analysis;Dan Xu 等;《Mechatronics and Automation(ICMA),2012 International Conference on.IEEE》;20121231;全文 |
自然场景下交通标志的检测与分类算法研究;李伦波;《中国博士学位论文全文数据库信息科技辑》;20100215(第201002期);37-116 |
Also Published As
Publication number | Publication date |
---|---|
CN105320940A (en) | 2016-02-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Chen et al. | Sports camera calibration via synthetic data | |
Boltes et al. | Automatic extraction of pedestrian trajectories from video recordings | |
Lin et al. | A fast, complete, point cloud based loop closure for LiDAR odometry and mapping | |
Coughlan et al. | Manhattan world: Compass direction from a single image by bayesian inference | |
CN106340044B (en) | Join automatic calibration method and caliberating device outside video camera | |
CN110175576A (en) | A kind of driving vehicle visible detection method of combination laser point cloud data | |
CN101887586B (en) | Self-adaptive angular-point detection method based on image contour sharpness | |
Zhou et al. | LIDAR and vision-based real-time traffic sign detection and recognition algorithm for intelligent vehicle | |
CN104850850B (en) | A kind of binocular stereo vision image characteristic extracting method of combination shape and color | |
CN103868510B (en) | A kind of quickly whole day star chart autonomous fixed star recognition methods | |
CN107564062A (en) | Pose method for detecting abnormality and device | |
CN103473571B (en) | Human detection method | |
CN107680137B (en) | A kind of polyphaser calibration and parameter optimization method based on sphere projection common tangent | |
CN104463866A (en) | Local shape matching method based on outline random sampling | |
CN105320940B (en) | A kind of method for traffic sign detection based on center-border link model | |
Moghadam et al. | Road direction detection based on vanishing-point tracking | |
CN103512558B (en) | A kind of conical target binocular video pose measuring method and target pattern | |
CN110379004B (en) | Method for classifying ground features and extracting single objects of oblique photography results | |
CN106127205A (en) | A kind of recognition methods of the digital instrument image being applicable to indoor track machine people | |
CN103295239A (en) | Laser-point cloud data automatic registration method based on plane base images | |
CN109685855A (en) | A kind of camera calibration optimization method under road cloud monitor supervision platform | |
CN109544599A (en) | A kind of three-dimensional point cloud method for registering based on the estimation of camera pose | |
CN104504675B (en) | A kind of active vision localization method | |
Krajník et al. | Image features and seasons revisited | |
CN105809678B (en) | A kind of line segment feature global registration method between two views under short base line condition |
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 |
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 |
|
CB03 | Change of inventor or designer information | ||
GR01 | Patent grant | ||
GR01 | Patent grant |