CN102831420B - Circular traffic sign positioning method based on color information and randomized circle detection - Google Patents

Abstract

A circular traffic sign positioning method based on color information and randomized circle detection includes steps of firstly, transferring an image of an RGB (red, green and blue) color space into an HSV (hue, saturation and value) space and precutting the image according to color features of a circular traffic sign; secondly, carrying out circle detection in an edge image by details of selecting a candidate circle from a randomized sampling accumulation process under a constraint condition and then checking whether proper circles exist or not and determining accurate quantity and accurate positions of the proper circles according to accumulative evidence, and accordingly positioning the circular traffic sign. The circular traffic sign positioning method based on color information and randomized circuit detection has the advantages of high noise resistance, accuracy and real-time performance under different illumination conditions in complicated natural scenes.

Description

Based on the circular traffic sign localization method of colouring information and random loop truss

Technical field

The present invention relates to intelligent transportation field, especially a kind of circular traffic sign localization method.

Background technology

In vehicle-mounted backup system, identify the traffic sign in road how fast, accurately and efficiently and it fed back to driver or control system, for ensure driver safety, avoid traffic accident there is very important Research Significance.Realize the final of traffic sign accurately to identify, first to accomplish be exactly from natural scene image detection and location to sign image.Traffic sign is set to special CF usually, can make a distinction to greatest extent, be convenient to driver and grasp road information in time with nature and artificial background.

Current traffic sign localization method is divided into three major types: based on the localization method of colouring information, Shape-based interpolation feature and Color and shape.The people such as Escalera propose direct colored threshold segmentation method, and the characteristic information in RGB color space to traffic sign is split, and then determining that target area determines whether by profile analysis is traffic sign.Cyganek designs two look-up tables at HIS color space and carries out color segmentation to red and blueness, and recycling genetic algorithm locates traffic sign in designated space search.To wait people quietly to propose first to adopt color segmentation and not color decomposition technique to isolate circular traffic sign, and remove part background area, then adopt non-linear least square curve fitting technique accurately to extract circular traffic sign.

Because Hough transform can the geometric configuration such as detection of straight lines, circle, it is utilized to be very direct effective method to detect circular traffic sign.Circle Hough Transform (CHT) is the circle detection method be most widely used at present, the maximum feature of the method is that reliability is high, under the state of noise, distortion, subregion incompleteness, still can obtain desirable result, but the shortcoming of the method is that calculated amount and memory space are large.As the people such as Garcia-Garrido adopt Hough transform to detect traffic sign, but this algorithm is only suitable for detecting mark in the region preset, and calculated amount is larger.The people such as Xu propose random Hough transformation (RHT), carry out many-to-one mapping, decrease the expense of calculated amount and internal memory after edge point stochastic sampling.But the method for RHT is still the accumulation based on carrying out parameter space, the problem of consumption storage space consuming time fundamentally cannot be solved.The people such as Chen propose random loop truss algorithm (RCD).RCD method is derived by RHT thought, but is not the thought based on Hough transform.It has employed the structure of hypothesis-checking to detect possible circle.Owing to not relating to the accumulation of parameter space, and the calculating of hypothesis circle directly carries out in original image space, so efficiency and accuracy all increase compared with the method based on RHT; Compared to Hough transform and RHT conversion, RCD required storage is considerably less, has real-time speed and good noise immunity.

Summary of the invention

Consider that driver or pedestrian are to the carelessness of prohibitory sign, Warning Mark and speed(-)limit sign etc., cause the generation of a lot of traffic hazard.And for traffic sign location technology exist noiseproof feature poor, real-time is poor, the deficiency that the Position location accuracy that brings of the reason such as different illumination conditions and distortion is poor, the invention provides that a kind of noiseproof feature is strong, real-time is better, accuracy is good based on the circular traffic sign localization method of colouring information and random loop truss.

The technical solution adopted for the present invention to solve the technical problems is:

Based on a circular traffic sign localization method for colouring information and random loop truss, described localization method comprises the following steps:

1) coloured image of acquisition camera acquisition, described multicolour pattern is that rgb format stores, and the image of RGB color space is transformed into HSV space, and sets the H of traffic sign redness and traffic sign blueness, S components range, the color characteristic according to circular traffic sign carries out pre-segmentation;

2) in edge image, carry out loop truss, comprise following process:

2.1) edge extracting: medium filtering is carried out to pre-segmentation result, and adopt Sobel to carry out rim detection, obtain the bianry image comprising object edge information, i.e. edge point set;

2.2) constraint condition of stochastic sampling is set: point-to-point transmission is maximum, minor increment T _dmin, T _dmax; Sampled probability estimates maximum sampling number F; Edge point set allows minimum points N _min;

2.3) Stochastic choice 4 v are concentrated at described marginal point _i, v _j, v _k, v _l;

If d _{i → j}, d _{j → k}, d _{k → l}, d _{l → i}> T _dmin; d _{i → j}, d _{j → k}, d _{k → i}< T _dmax, judge v _i, v _j, v _k3 on possible candidate circle; Otherwise enter 2.7);

2.4) v is calculated _ldistance d between point to possibility candidate circle _{l → ijk}if, d _{l → ijk}be less than T _d, T is set _d∈ (1,3) then confirms as candidate's circle C _ijk; Otherwise enter 2.7);

2.5) ovality is set $O_r=\frac{\mathrm{\&Delta;r}}{r}\&times;100\%;$

Wherein, r is candidate's radius of a circle, and Δ r is radial misalignment value;

And set up annulus D: $\left\{D\right|{(r_{\mathrm{ijk}}-\frac{\mathrm{\&Delta;r}}{2})}^2\&le;{a_{\mathrm{ijk}}}^2+{b_{\mathrm{ijk}}}^2\&le;{(r_{\mathrm{ijk}}+\frac{\mathrm{\&Delta;r}}{2})}^2\};$

Wherein, (a _ijk, b _ijk) be the center of circle that candidate justifies; r _ijkit is candidate's radius of a circle;

2.6) the number count of accumulative annulus D inward flange point; If count>=2 π r _ijkt _r, T _rfor the integrity degree parameter threshold of circle, then confirm C _ijkbe a proper circle, and point on circle is concentrated removal from marginal point, sample counter zero setting, otherwise enter 2.7);

2.7) accumulative stochastic sampling number of times f, if f is less than the maximum sampling number F of permission, and the residue edge amount of counting is greater than N _min, then step 2.3 is turned back to), otherwise location is terminated.

Further, described step 2.2) in, estimate that the process of maximum sampling number F is as follows according to sampled probability:

If round A, B, C that image has three distributions of differing in size different, suppose circle is counted as N ₁, N ₂, N ₃, noise is counted as N ₄, then the probability while of stochastic sampling 4 on circle A is

$P_A=\frac{N_1(N_1-1)(N_1-2)(N_1-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_1}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4.---\left(1\right)$

In like manner, have

$P_B=\frac{N_2(N_2-1)(N_2-2)(N_2-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_2}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4,---\left(2\right)$

$P_C=\frac{N_3(N_3-1)(N_3-2)(N_3-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_3}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4.---\left(3\right)$

The Effective Probability that sampling should be carried out is

$P=P_A+P_B+P_C\&ap;\frac{{\mathrm{\&Sigma;}}_{i=1}^4{N_i}^4}{{\left({\mathrm{\&Sigma;}}_{i=1}^4{N}_i\right)}^4}.---\left(4\right)$

If counting of three circles is all equal to noise spot N, then draw if there be S circle in figure, so the probability of an efficiently sampling is estimate the minimum sampling number determining candidate's circle suppose to have on same circle and imitate sampling n time, so just have 4n to put concyclic; When n>=2, this candidate circle is that true round possibility is just very large; Extrapolating the minimum sampling number confirming as true circle when n=2 is can arrange according to picture noise degree and detect that the permission sampling number of a circle is F=2K ~ 8K.

Further again, described step 2.2) and 2.3) in, three non-colinear marginal points can determine a circle, be too near to and can't detect true circle in order to avoid 3 middle any two points, by image size setting threshold value T _dmin, image size is rows × columns; The size proportion accounting for original image according to the circle of reality detection is arranged $T_{d\max }=\{1,\frac{1}{2},\frac{1}{4},\frac{1}{8}\}\min (\mathrm{rows},\mathrm{columns}),$ Image size is rows × columns.

Technical conceive of the present invention is: for avoiding the factor such as illumination condition and Changes in weather on the impact of traffic sign tone, first the image of RGB color space is transformed into HSV space, the color characteristic according to circular traffic sign carries out pre-segmentation; Then in edge image, carry out loop truss, first add up process, to extract candidate's circle from stochastic sampling, re-use evidence and add up checking and whether there is proper circle and determine its accurate number and position, thus positioning round traffic sign.Finally by adopting the different complex background pictures of actual photographed to verify, the results show validity of algorithm herein.

Beneficial effect of the present invention is mainly manifested in: noiseproof feature is strong, accuracy is good, real-time is better.

Accompanying drawing explanation

Fig. 1 is the schematic diagram that stochastic sampling 4 determines 4 circles.

Fig. 2 is at concyclic 4 and determines candidate's circle C _ijkschematic diagram.

Fig. 3 is the schematic diagram of the character of circle string.

Fig. 4 detects circle and true round deviation schematic diagram.

Fig. 5 is the schematic diagram of the circle utilizing ovality adjustment to detect.

Fig. 6 is the localization method process flow diagram of circular traffic sign.

Fig. 7 is the simple process figure of circle detection.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1 ~ Fig. 7, a kind of circular traffic sign localization method based on colouring information and random loop truss, comprises following process:

Edge extracting: using shape to detect a slow main cause of traffic sign is because it must calculate each pixel in whole image.After above-mentioned color segmentation, a large amount of interference regions is removed.In order to reduce computation complexity further, carry out medium filtering herein to the result of color segmentation, its output is g (x, y)=med { f (x-k, y-l), (k, l) ∈ W }, wherein, f (x, y), g (x, y) be respectively original image and the rear image of process, W is 3 × 3 templates.It is very effective that median filtering method offsets particle-removing noise, can be used to again Protect edge information information.Then we adopt Sobel boundary operator to carry out rim detection.Finally, the bianry image that the quantity of information of object edge is little is obtained comprising.

Random loop truss: RCD basic thought is: stochastic sampling 4 marginal points, uses a distance criterion to judge whether to there is candidate's circle.Candidate's circle judges needs four times, as shown in Figure 1, then justifies candidate and carries out evidence and add up to determine whether proper circle.

If V is image border point set, stochastic sampling four point, wherein front three-point (x _m, y _m), m=i, j, k, institute determines circle C _ijkthe center of circle be (a _ijk, b _ijk), radius is r _ijk, then they meet

${(x_m-a_{\mathrm{ijk}})}^2+{(y_m-b_{\mathrm{ijk}})}^2=r_{\mathrm{ijk}}^2,m=i,j,k.---\left(5\right)$

Above formula can be rewritten as

${2x}_m{a}_{\mathrm{ijk}}+{2y}_m{b}_{\mathrm{ijk}}+d_{\mathrm{ijk}}=x_m^2+y_m^2,---\left(6\right)$

Wherein, three non-colinear marginal points can determine a circle.Be too near in order to avoid 3 middle any two points and can't detect true circle, by image size setting threshold value T _dmin, general T _dmin=a ₀min (rows, columns), a ₀∈ (1/60,1/40), image size is rows × columns.The distance between 3 is only had to be greater than T _dmin, this 3 determined Circle Parameters could be calculated.By (x _m, y _m), m=i, j, k, can obtain three equations after substituting into formula (6), simultaneous equations can try to achieve Circle Parameters a _ijk, b _ijk, r _ijk.The calculated amount that simultaneous three equation of a circles solve Circle Parameters is obviously larger, and here, we adopt the center of circle to cross the character of the perpendicular bisector of string to try to achieve Circle Parameters, as shown in Figure 3.

If marginal point v _i=(x _i, y _i), v _j=(x _j, y _j), v _k=(x _k, y _k) concyclic, the determined perpendicular bisector equation of simultaneous two string is

$\left\{\begin{array}{c}y-\frac{y_i+y_j}{2}=k_{\mathrm{ij}}(x-\frac{x_i+x_j}{2})\\ y-\frac{y_i+y_k}{2}=k_{\mathrm{jk}}(x-\frac{x_j+x_k}{2})\end{array}\right.,---\left(7\right)$

Wherein, if there is k _ij≈ k _jkor y _j=y _ior y _k=y _jspecial circumstances, then select else 3 calculate.

From formula 6) simultaneous equations can try to achieve the center of circle (a _ijk, b _ijk), radius $r_{\mathrm{ijk}}=\sqrt{{(x_i-a_{\mathrm{ijk}})}^2+{(y_i-b_{\mathrm{ijk}})}^2},$ 4th v _l=(x _l, y _l) to circle C _ijkdistance be

$d_{l\&RightArrow;\mathrm{ijk}}=|\sqrt{{(x_l-a_{\mathrm{ijk}})}^2+{(y_l-b_{\mathrm{ijk}})}^2}-r_{\mathrm{ijk}}|.$

Ideally, if sample four concyclic, then four points are all zero to the distance of circle contour, but consider the quantization error of digital picture, a therefore given threshold value T _d, generally T is set _d∈ (1,3).As long as d _{i → jkl}, d _{j → ikl}, d _{k → ijl}, d _{l → ijk}in any one is less than T _d, so just can determine candidate's circle.

After determining candidate's circle, to the distance d of each some p calculation level in the point set V of image border to candidate's circle C _{p → C}, p ∈ V, the counter count of initialization simultaneously.If d _{p → C}< T _d, then count=count+1.The integrity degree parameter threshold of a circle is set according to the incomplete degree of circle if count > 2 π is r _ijkt _r, then think that candidate's circle is proper circle, and from the point set V of image border, remove the upper point of circle.

The probability analysis of random loop truss: for stochastic sampling four point, just effective when the point only sampled is on same circle, so necessarily there is a large amount of Null Spot groups.The factor affecting algorithm speed comprise find candidate's circle before calculate the number of times of Circle Parameters and candidate's circle evidence cumulative frequency.The setting of sampling number also determines to be detected as power.

If round A, B, C that image has three distributions of differing in size different, suppose circle is counted as N ₁, N ₂, N ₃, noise is counted as N ₄, then the probability while of stochastic sampling 4 on circle A is

$P_A=\frac{N_1(N_1-1)(N_1-2)(N_1-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_1}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4.---\left(8\right)$

In like manner, have

$P_B=\frac{N_2(N_2-1)(N_2-2)(N_2-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_2}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4,---\left(9\right)$

$P_C=\frac{N_3(N_3-1)(N_3-2)(N_3-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_3}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4.---\left(10\right)$

The Effective Probability that sampling should be carried out is

$P=P_A+P_B+P_C\&ap;\frac{{\mathrm{\&Sigma;}}_{i=1}^4{N_i}^4}{{\left({\mathrm{\&Sigma;}}_{i=1}^4{N}_i\right)}^4}.---\left(11\right)$

If counting of three circles is all equal to noise spot N, draw if there be S circle in figure, so the probability of an efficiently sampling is the minimum sampling number 1/P determining candidate's circle can be estimated, suppose to have on same circle and imitate sampling n time, so just have 4n to put concyclic; When n>=2, this candidate circle is that true round possibility is just very large; Extrapolating the minimum sampling number confirming as true circle when n=2 is can arrange according to picture noise degree and detect that the permission sampling number of a circle is F=2K ~ 8K.

Improvement to random circle detection method: random circle detection method occurs good testing result when circle contour is better under process certain noise, but when occurring that distortion, disappearance and noise are too large, detection speed and precision just decline to some extent.We are from sampling constraints condition and arrange distance proportion threshold value two aspect and improve algorithm.

Adopt the random circle detection method of distance restraint: it seems from above-mentioned probability analysis, the probability of the likelihood ratio efficiently sampling of invalid sampling is much bigger.The accumulation of Null Spot group means that Circle Parameters calculation times increases and increases with candidate's circle evidence cumulative frequency, and the probability so reducing invalid sampling just can the detection speed of boosting algorithm.

We propose a kind of simple sampled distance constraint thought: although sampling mechanism is random equiprobability (being uniformly distributed), the distribution of the marginal point in image is not uniform.For a circle, the point on circle is all concentrate on a specific region.Suppose that first sampling is a bit and on circle, to stochastic sampling three point, if this distance of 3 o'clock to first o'clock is all less than threshold value T _dmax, ideal situation is set to maximum diameter of a circle r _max, just calculate this 4 point; Otherwise do not calculate.Which reduces the number of times and candidate's circle checking number of times that also calculate Circle Parameters at other 3 o'clock to first o'clock when distance is greater than this threshold value.By the priori in test and actual testing process, this range constraint mode is easy to realize.The size proportion that the circle that we can detect according to reality accounts for original image is arranged $T_{d\max }=\{1,\frac{1}{2},\frac{1}{4},\frac{1}{8}\}\min (\mathrm{rwos},\mathrm{columns}),$ Image size is rows × columns.Such as during road traffic sign detection, the size of circular traffic sign is relevant with the angle and distance of observer, can set threshold value directly max-thresholds can be given when knowing diameter of a circle scope.

According to the method for distance restraint, random selecting 4 point, suppose that be point 1 on circle at first, latter 3 is point 2,3,4, if meet the distance d of point 1 _{i → 1}< T _dmax, i=2,3,4, just think that this group is effective, candidate's circle deterministic process can be entered; Otherwise, be designated as Null Spot group, sample next time.

We are by 4 width sizes the trial image that 282 × 282 images emulate as MATLAB, and host CPU frequency is 1.66GHz.We get maximal distance threshold T _dmaxfor 1/4 of picture size.Table 1 gives and adopts RCD method the Circle Parameters calculation times N required for all circles successfully to be detected to every width image _pand meet the 4th the candidate's circle number N on the determined circle of front three-point _c.The identical data adopting the RCD method of distance restraint to detect is provided in table 2.

Table 1.RCD detects the N of circle _p, N _c

	a	b	c	d
					N P	22,404	7,793	13,254	8,874
N C	1,093	925	726	529

Round N is detected under table 2. distance constraints _p, N _c

	a	b	c	d
					N P	9，374	2,492	9,118	4,847
N C	304	282	339	187

From table 1, can find out in 2, the parameter calculation times under distance constraints and candidate's circle number obviously reduce.Which reduces the calculated amount of Circle Parameters and the checking of candidate's circle.

The optimum configurations analysis of random loop truss: adopt distance threshold T in the proof procedure of candidate's circle _dwith the integrity degree parameter threshold T of circle _r.If in distance threshold, be then designated as the available point on circle.If the integrity degree parameter of circle reaches threshold value, then using the parameter of the parameter of this candidate circle as the circle detected.These two threshold values of conbined usage can make there is a deviation between detected parameters and actual parameter, as shown in Figure 4.

At candidate's circle Qualify Phase, need service range threshold value and proportion threshold value to confirm true circle.But there is size and the different circle of concyclic ratio in practice.Candidate's circle set a distance really, i.e. the 4th the distance T to candidate's circle _dbeing a fixed value with checking distance, is irrational to the round checking varied in size like this.One is used to evaluate circular parameter, relative elliptical degree

$O_r=\frac{\mathrm{\&Delta;r}}{r}\&times;100\%,---\left(12\right)$

Wherein, r is candidate's radius of a circle, and Δ r is radial misalignment value.As long as a given like this O _r, just can define a circle ring area

$\left\{D\right|{(r_{\mathrm{ijk}}-\frac{\mathrm{\&Delta;r}}{2})}^2\&le;a^2+b^2\&le;{(r+\frac{\mathrm{\&Delta;r}}{2})}^2\},---\left(13\right)$

Wherein, (a, b) is the center of circle of candidate's circle.As long as marginal point just thinks that in this annulus this point is an available point, which improves the probability of concyclic ring.O can be set before detection _rdefine and need detected circle contour, this parameter and circle contour size have nothing to do.

In addition, the visual angle for the shooting of reality is different, and geometric warping or deformation can occur toroidal, in ellipticity, as shown in Figure 5.

As long as marginal point is in annulus, just think concyclic ring, just can as the point of the true circle of checking.Rational Δ r value is set, just can revises at the enterolithic circle of annulus D, detect that radius is the circle of r.At true circle Qualify Phase, we just can verify all marginal points, and directly accumulation calculates the marginal point in annulus, sees that whether meeting concyclic proportion threshold value confirms as true circle.

In the loop truss algorithm IRCD improved, f is sample counter, and F is the maximum sampling number detecting that a circle allows, | V| is the size of residue edge point set, N _minfor judging the threshold value counted in the minimum edge of a circle whether needed for proper circle, T _dmaxand T _dminbe respectively maximum, minor increment that point-to-point transmission allows, T _rfor the integrity degree parameter threshold of circle.

RH T, the Performance comparision of RCD, IRCD: use RHT respectively, RCD and IRCD detects above-mentioned 4 width images, and provides the time of three kinds of methods detections in table 3.As can be seen from Table 3 IRCD detection speed comparatively first two method double left and right.From testing result, IRCD method can avoid RCD detection institute to occur offset issue.

Contrast detection time of table 3. three kinds of methods

	RHT(s)	RCD(s)	IRCD(s)
				Coin	1.257	0.9962	0.5452
Planet	0.983	0.7783	0.4926
				Gobang	0.891	0.5856	0.2647
Cake	0.834	0.5398	0.2490

The detection method of circular traffic sign describes: color segmentation and SHAPE DETECTION are the important methods of two kinds in road traffic sign detection.Gray level image is generally adopted to the method for SHAPE DETECTION, its significant drawbacks is that computation complexity is high; For coloured image generally by means of color segmentation, need a suitable wave filter to be filtered out from complex background by colored traffic sign, the easy distortion of colouring information, be subject to the interference of surroundings, Algorithm robustness is poor.

The detection method of analysis integrated CF:

(1) color of traffic sign is strict design, is also one of key character, can detects roughly potential traffic sign region, remove most of jamming pattern by color segmentation technology;

(2) cause the computation complexity of SHAPE DETECTION high to whole colour or gray level image in pixel level row operation.If convert it into the bianry image that information content is substantially constant, denoising, computation complexity will reduce greatly.Consider that the circular edge feature of circular traffic sign does loop truss.

The color segmentation of HSV color space: the coloured image being arranged on the video camera acquisition on motor vehicle stores by rgb format.For RGB color model, R, G, B component has the correlativity of height, is vulnerable to the impact of body surface reflectivity, intensity of illumination etc., therefore RGB color space the Accurate Segmentation of Traffic Sign Images under being not suitable for nature complex scene.Therefore, be necessary closely-related for component rgb space to be transformed into the substantially incoherent chrominance space of its component.We select to carry out color segmentation closer to the hsv color space of human experience and perception, and it has three components: H represents tone, and S represents saturation degree, and V represents lightness.

According to the color assignment [9] of the traffic sign of China, Li Lunbo [1] is according to the color analysis combined 3000 width natural scene Traffic Sign Images under various weather condition, and determine the color threshold of HSV color space, concrete threshold value is as shown in table 4.Threshold value in table has been standardized interval to [0,1] all.H _r, S _rand H _b, S _bbe respectively the component value of traffic sign redness and traffic sign blueness.

The color classification threshold value of table 4.HSV color space

Component/color	R	B	Y
				H	H R< 0.025 or H R＞0.80	0.51＜H B＜0.67	0.118＜H Y＜0.60
S	S R＞0.50	S B＞0.50	S Y＞0.60

Existing circular traffic sign mainly comprises this three class of prohibitory sign, speed(-)limit sign and Warning Mark.Its color characteristic is in table 5.

The color characteristic of table 5. circular traffic sign

Type of sign	Prohibitory sign	Speed(-)limit sign	Warning Mark
				Color characteristic	Red edge	Red edge	Blue background

For real scene shooting traffic RGB image, first color segmentation algorithm is transformed into HSV space, and is normalized to [0,1].And associative list 4,5, set scope that is red and each component of blueness through great many of experiments.

Utilize the color characteristic of the distinctive red edge of circular traffic sign and blue background can remove the background area larger with target area color distinction, but can not ensure to be exactly pure sign image, also may comprise other background areas consistent or close with flag colors.Further consider that the differences in shape of circular traffic sign and jamming pattern will use circle detection method, thus accurate positioning round traffic sign region.

Detection based on IRCD is implemented: for the quick and precisely location of traffic sign, mainly combine the detection method of CF herein.We are according to the IRCD of above-mentioned analysis, in conjunction with repeatedly traffic sign example laboratory, arrange correlation parameter: $T_{d\max }=\frac{1}{4}\min (\mathrm{rows},,\mathrm{columns}),T_{d\min }=\frac{1}{48}\min (\mathrm{rows},\mathrm{columns});$ Relative elliptical degree detect that the sampling number that a circle allows is F=3K, having at most K=3773 in 6 round situations.It is N that marginal point concentrates permission left point minimum _min=100.If point set size is less than N _min, then end is detected.

As the description that Fig. 6 is to this paper circular traffic sign localization method.Fig. 7 is the simple process figure of circle detection.

Circular traffic sign positioning experiment interpretation of result: test circular traffic sign this paper algorithm of several actual photographed, picture size is all 1280 × 960.Adopt MATLAB simulation software as experimental tool, host CPU frequency is 1.66GHz, and below completing, the positioning time of every width image is at 4s to 5s, comprises color space conversion, IRCD algorithm and circle mark.Under complicated nature background, circular traffic sign yardstick is larger or less, or can orient traffic sign region exactly under traffic sign geometric deformation.Meanwhile, for different light conditions, also achieve good effect, show that algorithm also has good robustness to illumination.

Claims (3)

1., based on a circular traffic sign localization method for colouring information and random loop truss, it is characterized in that: described localization method comprises the following steps:

1) coloured image of acquisition camera acquisition, described coloured image is that rgb format stores, and the image of RGB color space is transformed into HSV space, and sets the H of traffic sign redness and traffic sign blueness, S components range, the color characteristic according to circular traffic sign carries out pre-segmentation;

2) in edge image, carry out loop truss, comprise following process:

2.1) edge extracting: medium filtering is carried out to pre-segmentation result, and adopt Sobel to carry out rim detection, obtain the bianry image comprising object edge information, i.e. edge point set;

2.2) constraint condition of stochastic sampling is set: point-to-point transmission is maximum, minor increment T _dmin, T _dmax; Sampled probability estimates maximum sampling number F; Edge point set allows minimum points N _min;

2.3) Stochastic choice 4 v are concentrated at described marginal point _i, v _j, v _k, c _l;

If d _{i → j}, d _{j → k}, d _{k → l}, d _{l → i}> T _dmin; d _{i → j}, d _{j → k}, d _{k → i}< T _dmax, judge v _i, v _j, v _k3 on possible candidate circle; Otherwise enter and sample, sample counter adds 1 next time;

2.4) v is calculated _ldistance d between point to possibility candidate circle _{l → ijk}if, d _{l → ijk}be less than T _d, T is set _d∈ (1,3) then confirms as candidate's circle C _ijk; Otherwise enter and sample, sample counter adds 1 next time;

2.5) ovality is set $O_r=\frac{\mathrm{\&Delta;r}}{r}\&times;100\%;$

Wherein, r is candidate's radius of a circle, and Δ r is radial misalignment value;

And set up annulus

Wherein, (a _ijk, b _ijk) be the center of circle that candidate justifies; r _ijkit is candidate's radius of a circle;

2.6) accumulative annulus the number count of inward flange point; If count>=2 π r _ijkt _r, T _rfor the integrity degree parameter threshold of circle, then confirm C _ijkbe a proper circle, and point on circle is concentrated removal from marginal point, sample counter zero setting;

2.7) accumulative stochastic sampling number of times f, if f is less than the maximum sampling number F of permission, and the residue edge amount of counting is greater than N _min, then step 2.3 is turned back to), otherwise location is terminated.

2. as claimed in claim 1 based on the circular traffic sign localization method of colouring information and random loop truss, it is characterized in that: if round A, B, C that image has three distributions of differing in size different, count as N on supposing to justify ₁, N ₂, N ₃, noise is counted as N ₄, then the probability while of stochastic sampling 4 on circle A is

$P_A=\frac{N_1(N_1-1)(N_1-2)(N_1-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_1}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4---\left(1\right)$

In like manner, have

$P_B=\frac{N_2(N_2-1)(N_2-2)(N_2-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_2}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4---\left(2\right)$

$P_C=\frac{N_3(N_3-1)(N_3-2)(N_3-3)}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-1)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-2)({\mathrm{\&Sigma;}}_{i=1}^4{N}_i-3)}\&ap;{\left(\frac{N_3}{{\mathrm{\&Sigma;}}_{i=1}^4{N}_i}\right)}^4---\left(3\right)$

The Effective Probability that sampling should be carried out is

$P=P_A+P_B+P_C\&ap;\frac{{\mathrm{\&Sigma;}}_{i=1}^4{N_i}^4}{{\left({\mathrm{\&Sigma;}}_{i=1}^4{N}_i\right)}^4}---\left(4\right)$

If counting of three circles is all equal to noise spot N, draw if there be S circle in figure, so the probability of an efficiently sampling is estimate the minimum sampling number 1/P determining candidate's circle, suppose to have on same circle and imitate sampling n time, just have 4n to put concyclic; When n>=2, this candidate circle is that true round possibility is just very large; Extrapolating the minimum sampling number confirming as true circle when n=2 is $K=\frac{2S(2S-1)}{S\left(\begin{array}{c}2\\ 1\end{array}\right)P}={(S+1)}^3(2S-1);$ Arrange according to picture noise degree and detect that the permission sampling number of a circle is F=2K ~ 8K.

3. as claimed in claim 1 or 2 based on the circular traffic sign localization method of colouring information and random loop truss, it is characterized in that: described step 2.2) and 2.3) in, three non-colinear marginal points can determine a circle, be too near in order to avoid 3 middle any two points and can't detect true circle, by image size setting threshold value T _dmax, image size is rows × columns, and the size proportion accounting for original image according to the circle of reality detection is arranged $T_{d\max }=\{1,\frac{1}{2},\frac{1}{4},\frac{1}{8}\}\min (\mathrm{rows},\mathrm{columns}),$ Image size is rows × columns.