CN109145746A - A kind of signal lamp detection method based on image procossing - Google Patents

Abstract

Translated fromChinese

本发明公开了一种基于图像处理的信号灯检测方法，包括如下步骤：步骤1：从前端相机读取图像文件F₀；步骤2：读取系统配置文件中的感兴趣区域，并从F₀中的拷贝感兴趣区域图像F₁；步骤3：根据系统当前时间对图像F₁进行图像增强；步骤4：对图像F₂进行颜色区域分割，步骤5：对图像F₄进行形态学的闭运算处理；步骤6：提取图像F₄中的连通区域，步骤7：提取C_L中区域的外接矩形，并根据外接矩形位置从图像F₂中拷贝候选信号灯图像，得信号灯图像集合L；步骤8：提取L中所有图像的Hog特征，本发明的有益效果是，能实现复杂场景的信号灯检测，并同时满足系统的稳定性和实时性的需求。

The invention discloses a signal light detection method based on image processing, comprising the following steps: step 1: reading an image file F₀ from a front-end camera_; Copy the area of interest image F1; Step₃ : Perform image enhancement_on the image F1 according to the current time of the system; Step₄ : Perform color area segmentation on the image F2, Step₅ : Perform morphological closing operation processing on the image F4 ; Step 6: extract the connected area in the image F₄ , step 7: extract the circumscribed rectangle of the area in_CL , and copy the candidate signal light image from the image F₂ according to the position of the circumscribed rectangle to obtain the signal light image set L; Step 8: extract Hog feature of all images in L, the beneficial effect of the present invention is that it can realize signal light detection of complex scene, and simultaneously meet the requirements of stability and real-time performance of the system.

Description

A kind of signal lamp detection method based on image procossing

Technical field

The present invention relates to technical field of image processing, and in particular to a kind of signal lamp detection method based on image procossing.

Background technique

With the rapid development of internet and hardware art, the artificial intelligence epoch have arrived.Researcher gradually startsThe problems in daily life is solved using various related advanced technologies, so that it is growing to quality of life to meet peopleIt pursues.Intelligent transportation system is exactly one of this product of the time, it together, mentions the technological incorporation of multiple fields for peopleFor safety and convenience.And key components of the detection of traffic lights as intelligent transportation system, there is huge researchValue and potentiality, but since signal lamp detection has higher requirement to method robustness, energy stable operation is in complicated road dayIn gas field scape, and requirement of real-time is higher, and most detection methods can not meet above-mentioned requirements simultaneously.Therefore it is based on imageThe signal lamp detection method of processing is a kind of preferable solution, has not only been able to satisfy the requirement of real-time of signal lamp detection, but also energyStable operation is in the weather scene of Various Complex.

For the accurate detection property and the problem of real-time of signal lamp, domestic and international academia, industry propose many schemes.The technical solution being wherein closer to the present invention includes: that (Wu Ying, Zhang little Ning, He Bin are believed Wu Ying based on the traffic of image procossingSignal lamp recognition methods [J] traffic information and safety, 2011,29 (3): 51-54.) using the good HSI of complex but stabilityColor space carries out the segmentation of image, by calculating circularity and traffic lights bottom plate length-width ratio, filters candidate region, finallyTraffic signals lamp type is confirmed by template matching.But algorithm accuracy in complex scene is not high, and needs to know in advance, there are many restrictions in practical applications in the placement direction of road traffic lights.Gu Mingqin (Gu Mingqin, Cai Zixing, Huang Zhenwei,Real-time recognizer [J] Central South University's journal (natural science edition) of arrowhead-shaped traffic lights in equal urban environment, 2013,44 (4): 1403-1408.) image under RGB color is transformed into YCbCr color space, recycle Gabor wavelet transformationThe feature of traffic lights candidate region is obtained with two-dimentional independent component analysis, finally with nearest neighbor classifier to traffic lightsType identified.Since the conversion of RGB color to YCbCr color space is linear, so the algorithm real-timeRelatively good, under the relatively good environment of light condition on daytime, recognition accuracy is relatively high, but is not suitable for night environment still.(Zhou Xuanru, Yuan Jiazheng, Liu Hongzhe wait based on real-time recognizer research [J] of the traffic lights of HOG feature to Zhou XuanruCalculation machine science, 2014,41 (7): 313-317.) algorithm that proposes carries out image segmentation also in YCbCr color space, thenObtain traffic lights candidate region by area, shape, density triple filter, then after extraction process image HOG feature,It is identified using SVM classifier.The method that algorithm uses machine learning, can carry out traffic in more complicated environmentThe detection of signal lamp, while round traffic lights can be detected but also detect arrow-shaped traffic lights, but for havingThe traffic lights recognition effect of number is bad.

In conclusion existing in current signal light detection scheme following insufficient: 1) algorithm accuracy is not high, may not apply toComplicated weather scene；2) traffic lights various shapes, algorithm cannot be suitable for a variety of traffic signals of different shapes simultaneouslyLamp inspection is surveyed；3) most of traffic lights detection algorithms are detected for the traffic scene on daytime, to the signal lamp inspection at nightIt is poor to survey effect.

Traffic lights identification is the base application of wisdom traffic system, determines the accurate of break in traffic rules and regulations and traffic schedulingProperty and real-time, but from the point of view of existing achievement, intelligent transportation system is also and not perfect, and the detection of traffic lights and identificationAlso there are many areas for improvement for key components as system.From the point of view of current many achievements, in traffic signalsLamp context of detection, for color and shape as most important two features of signal lamp, many algorithms are unfolded around them,But since the detection method is higher to real-time and stability requirement, most algorithm, which can not meet simultaneously the two, to be wantedIt asks, and by image enhancement, color segmentation, Morphological scale-space, geometrical characteristic filtering and to signal lamp area in image in the present inventionDomain extracts and passes through machine learning method marker lamp, is able to achieve the signal lamp detection of complex scene, and meets simultaneouslyThe stability of system and the demand of real-time.

Summary of the invention

In order to solve the problems in the existing technology, the present invention provides accurate, real-time a kind of based on image procossingSignal lamp detection method.

Technical scheme is as follows:

A kind of signal lamp detection method based on image procossing, which comprises the steps of:

Step 1: reading image file F from front end camera₀；

Step 2: reading the area-of-interest in system configuration file, and from F₀In copy region of interest area image F₁；

Step 3: according to the current time in system to image F₁Carry out image enhancement；6:00 between 18:00 be daytime, it is rightImage F₁Histogram equalization is carried out, remaining time is night, to F₁Carry out gamma correction, image F after must enhancing₂；

Step 4: to image F₂Carry out color region segmentation, the specific steps are as follows:

Step 4.1 is by image F₂It is transformed into hsv color space；

Step 4.2 is according to system time to F₂Color threshold segmentation is carried out, Pixel Information in set red and green is retained,And the lightness of other pixels is set as 0, obtain Threshold segmentation image F₃；Retain the pixel value for meeting the pixel of formula (1) daytime；Night retains the pixel value for meeting the pixel of formula (2)；

In formula, p (x, y) indicates location of pixels as (x, y) pixel, and h, s, v respectively indicate the point in Color Channel HSVThe channel value in each channel, red indicate red pixel point set, and green indicates green pixel point set；

Step 4.3 is to image F₃Gray processing is carried out, and binaryzation is carried out by OTSU, obtains binary image F₄；

Step 5: to image F₄Carry out morphologic closed operation processing；

Step 6: extracting image F₄In connected region, obtain connected region set C={ A_i| i=1,2,3 ..., n }, A_iTableShow i-th of connected region, n indicates total connected region quantity；According to formula (3) (4) (5), to object in C, it carries out geometrical characteristic mistakeFilter, ungratified object is rejected from C, obtains candidate signal lamp regional ensemble C_L；

S_min< A_i.S < S_max (3)

r_whmin< A_i.r_wh< r_whmax (4)

K_min< A_i.K < K_max (5)

Wherein, wherein S_minAnd S_maxRespectively indicate minimum and maximum area threshold given in advance, A_i.S A is indicated_iRegionArea；r_whminAnd r_whmaxRespectively indicate minimum and maximum the ratio of width to height given in advance, A_i.r_whIndicate A_iRegion the ratio of width to height；K_minAnd K_maxRespectively indicate minimum and maximum areal concentration given in advance, A_i.K A is indicated_iAreal concentration, i.e. A_iIn connected regionPixel quantity and the region minimum circumscribed rectangle area ratio；

Step 7: extracting C_LThe boundary rectangle in middle region, and according to boundary rectangle position from image F₂Middle copy candidate signalLamp image obtains signal lamp image collection L；

Step 8: the Hog feature of all images in L is extracted, and Feature Dimension Reduction is carried out to it by PCA method, finally byTrained SVM classifier identifies signal lamp according to characteristics of image after dimensionality reduction.

The beneficial effects of the present invention are: guaranteeing that traffic lights have good detection effect under complex environment；According to letterThe priori features such as signal lamp shape color carry out rapidly extracting to signal lamp region, and are dropped by PCA method to Hog featureDimension, improves the speed of algorithm, is able to satisfy the real-time demand of system；The image of day and night is handled respectively, it can be sameWhen meet day and night difference illumination scene signal lamp detection, have higher robustness；It being capable of real-time detection various shapesWith the traffic lights of state, and accuracy with higher.

Detailed description of the invention

Fig. 1 is that the present invention is based on the signal lamp detection method flow charts of image procossing；

Fig. 2 is histogram equalization on daytime front and back comparison diagram in step 3 in the present invention；

Fig. 3 is comparison diagram before and after night gamma correction in step 3 in the present invention；

Fig. 4 is closing operation of mathematical morphology comparison diagram before and after the processing in step 4 in the present invention；

Fig. 5 is the signal lamp result schematic diagram obtained after step 8 in the present invention.

Specific embodiment

Elaborate that the present invention is based on the specific implementations of the signal lamp detection method of image procossing below with reference to embodimentMode.

Specific step is as follows for a kind of signal lamp detection method based on image procossing:

Step 1: reading image file F from front end camera₀；

Step 2: reading the area-of-interest in system configuration file, and from F₀In copy region of interest area image F₁；

Step 3: according to the current time in system to image F₁Carry out image enhancement；6:00 between 18:00 be daytime, it is rightImage F₁Histogram equalization is carried out, remaining time is night, to F₁Carry out gamma correction, image F after must enhancing₂；

Step 4: to image F₂Carry out color region segmentation, the specific steps are as follows:

Step 4.1 is by image F₂It is transformed into hsv color space；

Step 4.2 is according to system time to F₂Color threshold segmentation is carried out, Pixel Information in set red and green is retained,And the lightness of other pixels is set as 0, obtain Threshold segmentation image F₃；Retain the pixel value for meeting the pixel of formula (1) daytime；Night retains the pixel value for meeting the pixel of formula (2)；

In formula, p (x, y) indicates location of pixels as (x, y) pixel, and h, s, v respectively indicate the point in Color Channel HSVThe channel value in each channel, red indicate red pixel point set, and green indicates green pixel point set；

Step 4.3 is to image F₃Gray processing is carried out, and binaryzation is carried out by OTSU, obtains binary image F₄；

Step 5: to image F₄Carry out morphologic closed operation processing；

Step 6: extracting image F₄In connected region, obtain connected region set C={ A_i| i=1,2,3 ..., n }, A_iTableShow i-th of connected region, n indicates total connected region quantity；According to formula (3) (4) (5), to object in C, it carries out geometrical characteristic mistakeFilter, ungratified object is rejected from C, obtains candidate signal lamp regional ensemble C_L；

S_min< A_i.S < S_max (3)

r_whmin< A_i.r_wh< r_whmax (4)

K_min< A_i.K < K_max (5)

Wherein, wherein S_minAnd S_maxRespectively indicate minimum and maximum area threshold given in advance, A_i.S A is indicated_iRegionArea；r_whminAnd r_whmaxRespectively indicate minimum and maximum the ratio of width to height given in advance, A_i.r_whIndicate A_iRegion the ratio of width to height；K_minAnd K_maxRespectively indicate minimum and maximum areal concentration given in advance, A_i.K A is indicated_iAreal concentration, i.e. A_iIn connected regionPixel quantity and the region minimum circumscribed rectangle area ratio；In this example, S_min=30, S_max=120；r_whmin=0.5, r_whmax=2.0；K_min=0.5, K_max=0.9；

Step 7: extracting C_LThe boundary rectangle in middle region, and according to boundary rectangle position from image F₂Middle copy candidate signalLamp image obtains signal lamp image collection L；

Step 8: the Hog feature of all images in L is extracted, and Feature Dimension Reduction is carried out to it by PCA method, finally byTrained SVM classifier identifies signal lamp according to characteristics of image after dimensionality reduction.

Content described in this specification embodiment is only enumerating to the way of realization of inventive concept, protection of the inventionRange should not be construed as being limited to the specific forms stated in the embodiments, and protection scope of the present invention is also and in this field skillArt personnel conceive according to the present invention it is conceivable that equivalent technologies mean.

Claims (1)

1. a kind of signal lamp detection method based on image procossing, which comprises the steps of:

Step 1: reading image file F from front end camera₀；

Step 2: reading the area-of-interest in system configuration file, and from F₀In copy region of interest area image F₁；

Step 3: according to the current time in system to image F₁Carry out image enhancement；In 6:00 to being daytime between 18:00, to imageF₁Histogram equalization is carried out, remaining time is night, to F₁Carry out gamma correction, image F after must enhancing₂；

Step 4: to image F₂Carry out color region segmentation, the specific steps are as follows:

Step 4.1 is by image F₂It is transformed into hsv color space；

Step 4.2 is according to system time to F₂Color threshold segmentation is carried out, retains Pixel Information in set red and green, and willThe lightness of other pixels is set as 0, obtains Threshold segmentation image F₃；Retain the pixel value for meeting the pixel of formula (1) daytime；Night,Retain the pixel value for meeting the pixel of formula (2)；

In formula, p (x, y) indicates that location of pixels is (x, y) pixel, and h, s, it is each logical in Color Channel HSV that v respectively indicates the pointThe channel value in road, red indicate red pixel point set, and green indicates green pixel point set；

Step 4.3 is to image F₃Gray processing is carried out, and binaryzation is carried out by OTSU, obtains binary image F₄；

Step 5: to image F₄Carry out morphologic closed operation processing；

Step 6: extracting image F₄In connected region, obtain connected region set C={ A_i| i=1,2,3 ..., n }, A_iIndicate i-thA connected region, n indicate total connected region quantity；According to formula (3) (4) (5), to object in C, it carries out geometrical characteristic filtering, willUngratified object is rejected from C, obtains candidate signal lamp regional ensemble C_L；

S_min< A_i.S < S_max (3)

r_whmin< A_i.r_wh< r_whmax (4)

K_min< A_i.K < K_max (5)

Wherein, wherein S_minAnd S_maxRespectively indicate minimum and maximum area threshold given in advance, A_i.S A is indicated_iArea surfaceProduct；r_whminAnd r_whmaxRespectively indicate minimum and maximum the ratio of width to height given in advance, A_i.r_whIndicate A_iRegion the ratio of width to height；K_minWithK_maxRespectively indicate minimum and maximum areal concentration given in advance, A_i.K A is indicated_iAreal concentration, i.e. A_iIn connected regionThe ratio of pixel quantity and the minimum circumscribed rectangle area in the region；

Step 7: extracting C_LThe boundary rectangle in middle region, and according to boundary rectangle position from image F₂Middle copy candidate signal lamp figurePicture obtains signal lamp image collection L；

Step 8: extracting the Hog feature of all images in L, and Feature Dimension Reduction is carried out to it by PCA method, finally by trainingGood SVM classifier identifies signal lamp according to characteristics of image after dimensionality reduction.