CN104392212A - Method for detecting road information and identifying forward vehicles based on vision - Google Patents

Abstract

The invention belongs to the field of intelligent automobile road detection and relates to a method for detecting road information and identifying forward vehicles based on vision. The method comprises image preprocessing, lane line characteristic parameter extraction, region-of-interest segmentation and vehicle outline identification. According to the method for detecting the road information and identifying the forward vehicles based on vision, the complexity of computation is simplified by extracting the regions of interest, filtering out background regions and reducing the processing scope; the processing result of each frame of image is obtained through fixed computation times by use of a line-by-line retrieval method, and differing from the characteristic, namely linear fitting on each bright spot, of Hough conversion, the method has outstanding advantages in algorithm instantaneity; the instantaneity of the Robinson direction model operator is improved, and an intermediate variable is set to reduce the computation times of each pixel point. The regions of interest are screened and judged by use of an information entropy and a vehicle tail symmetry characteristic in the target region, and therefore, the omission factor and the false drop rate of the algorithm are reduced.

Description

A kind of road information of view-based access control model detects and front vehicles recognition methods

Technical field

The invention belongs to intelligent automobile field of road detection, the road information being specifically related to a kind of view-based access control model detects and front vehicles recognition methods.

Background technology:

Intelligent vehicle is the system ensemble that integrates environment sensing, programmed decision-making, many grades assist the functions such as driving, it concentrates technology such as having used computing machine, modern sensing, information fusion, communication, artificial intelligence and automatic control, is typical new and high technology synthesis.At present the security, the comfortableness that improve automobile are mainly devoted to the research of intelligent vehicle, and excellent people's car interactive interface is provided.In recent years, the new power that oneself focus and auto industry through becoming world's Vehicle Engineering research of intelligent vehicle increases, a lot of developed country has been included in the intelligent transportation system given priority to separately all.Road information detection technique is the core link of intelligent vehicle control loop always, is the important technology of intelligent transportation system.And the detection and indentification of lane line and front vehicles is the matter of utmost importance realizing this technology.

In this field, there has been proposed many technology, the automatic driving vehicle ARGO system of VisLab development uses vision as main sensor, by setting up two degrees of freedom kinetic model and the preview follower model of vehicle, introduces feedback supervisory signals.Because after image reconstruction road environment, need just to obtain suitable bearing circle by the fit procedure of complexity and export, computation complexity in method very high, the consumption of hardware resource is very large.Tzomakas and Seelen achieves a kind of method obtaining road surface gray threshold, but cannot solve the problem of road surface grey scale change.The research of Marola belongs to Knowledge based engineering method, and the weak point of the method is that the false drop rate under complex environment can obviously increase.The people such as Wang propose the method for detecting lane lines based on B-spline.Have benefited from the arbitrariness that splines is expressed profile, the method can accurately identify straight way and bend, and road pavement shade has certain robustness.The profile allocation control points of B-spline is positioned at curved exterior, and therefore convergence needs successive ignition, and adds system complexity.At home, Chen Zhi proposes a kind of vehicle identification method based on wavelet transformation, but cannot meet the extensive adaptability of system matches.Main flow algorithm is the straight line by meeting most track feature in Hough transform recognition image now, thus demarcates.The advantage of this algorithm is that real-time is high, and weak point is that result is based on straight-line segment, is difficult to carry out providing effective parameter in the process of turning at vehicle, and calculated amount is large, and real-time is difficult to ensure.

Summary of the invention

Cannot meet the problem of early warning mechanism requirement for the robustness existed in prior art or real-time, the road information that the present invention proposes a kind of view-based access control model detects and front vehicles recognition methods, and first the method carries out self-adaption binaryzation segmentation to image; Then the region of interest ROI (Region Of Interes) in image is extracted; Adopt the method retrieved line by line to carry out the screening of unique point inside lane line, thus the left-right marker line parameter obtaining actual track is to carry out road model reconstruction.By burn into plavini filtering interfering point; Carry out the extraction of hatched merging and ROI region; Utilize the information entropy in target area, tailstock symmetric characteristics screen ROI region and differentiate, reduce the undetected of algorithm and false drop rate; Use the Robinson angle detecting operator extraction vehicle border of improving, achieve good effect.

The technical solution used in the present invention is for achieving the above object:

A kind of road information of view-based access control model detects and front vehicles recognition methods, the system realizing described method comprises: camera, installs the Measurement &control computer of video frequency collection card, the LAN (Local Area Network) that router is built, programmed decision-making host computer, intelligent vehicle BJUT-SHEV experiment porch.Camera is arranged on the ceiling front dead center position of intelligent vehicle BJUT-SHEV experiment porch, Real-time Collection road image; Camera is connected video frequency collection card with Measurement &control computer by USB, realizes the function of video data acquiring; The LAN (Local Area Network) that the controling parameters that Measurement &control computer process obtains is built by router passes to programmed decision-making host computer (Measurement &control computer); After programmed decision-making host computer resolves above-mentioned information, BJUT-SHEV experiment porch is controlled.It is characterized in that, described method performs following steps in Measurement &control computer:

Step 1, Image semantic classification.

Comprise: carry out gray processing to coloured image, adopt single maximum between-cluster variance OTSU method to carry out binarization segmentation, Sobel operator edge detection, image thinning process, determines road area.

Step 2, lane detection and depart from early warning.

Take the method retrieved line by line to obtain lane boundary point, utilize least square method to carry out matching to frontier point, obtain the quafric curve describing track.Judge the direction of vehicle front Road turnings, to vehicle, whether run-off-road line carries out early warning.

Step 3, ROI region is extracted.

Adopt the method dividing vehicle bottom shadow that road area gray scale combines with two OTSU, corrosion expansion process is carried out to segmentation image, fills gap regions, obtain vehicle ROI region based on shade at the bottom of car.

Step 4, vehicle ' s contour identification.

Carry out screening with the multiple features that information entropy and symmetry are Primary Reference foundation to region.Use the Robinson operator improved to process the part that remains after screening, ask for grey scale change Grad, and with Hough transform method identification vehicle outer contour.

Compared with prior art, the present invention has the following advantages:

(1) the present invention is by carrying out self-adaption binaryzation segmentation to image, reaches the effect of Adaptive matching image.

(2) real-time of the present invention by taking following measures to improve system: the area-of-interest in image is extracted, filter background region, reduce the process range of subsequent algorithm, simplify the complexity of calculating; Adopt the method retrieved line by line just can obtain the result of every two field picture through fixing calculation times, be different from Hough transform will carry out linear fit feature to each bright spot point, algorithm real-time has outstanding advantage; The present invention has carried out real-time improvement to Robinson direction template operator, arranges intermediate variable thus decreases the calculation times to each pixel.

(3) Hough transform from conventional in prior art is different, the present invention adopts the method retrieved line by line to carry out the screening of unique point inside lane line, the lane line that testing result is more fitted in real road can be made, there is not the limitation of line characteristics, thus cross in curved way as system provides more effective information at vehicle.

(4) cannot solve the problem of road surface grey scale change in the method for Tzomakas and Seelen proposition, the present invention, on the basis that self-adaption binaryzation is split, carries out second time OTSU Threshold segmentation, has extracted shade at the bottom of car exactly.By burn into plavini filtering interfering point, simplify and improve the efficiency that hacures merge and ROI region is extracted.

(5) the obvious problem increased of false drop rate meeting be under complex environment is caused for Marola Knowledge based engineering method, the present invention utilizes the information entropy in target area, tailstock symmetric characteristics screen ROI region and differentiate, reduce the undetected of algorithm and false drop rate, improve the feasibility of system under complex environment.

Accompanying drawing explanation

Fig. 1 is embodiment of the present invention hardware system composition frame chart;

Fig. 2 is the method for the invention main flow chart;

Fig. 3 is lane line Image semantic classification process flow diagram;

Fig. 4 is that vehicle departs from early warning principle schematic;

Fig. 5 is run-off-road line model figure: (a), for depart from left, (b) is for depart to the right;

Fig. 6 is that lane line extracts and method for early warning process flow diagram;

Fig. 7 is the position relationship of hacures length and image coordinate;

Fig. 8 is that ROI extracts process flow diagram;

Fig. 9 is vehicle ' s contour identification process figure.

Embodiment

Below in conjunction with drawings and Examples, the present invention will be further described.

The hardware system composition frame chart that embodiment adopts as shown in Figure 1, comprising:

Camera: after camera employing wind, mirror king series is a, is connected with Measurement &control computer with USB line.Camera is arranged on intelligent vehicle BJUT-SHEV experiment porch ceiling front dead center position, along with the advance of intelligent vehicle, just can collect the real-time information of road ahead.

The Measurement &control computer of video frequency collection card is installed: video frequency collection card adopts Si Tejiatu ST-769 capture card, and the simulated roadway information that Measurement &control computer receives by it is converted to digital image information.In addition, Measurement &control computer installs VS2010 and OPENCV2.4.5 and configuration software running environment, and realize the software program of the method for the invention.

The LAN (Local Area Network) that router is built: the WNR2000 that router adopts Netgear company to produce.The data message that Measurement &control computer is packed by the LAN (Local Area Network) that router is built, by this LAN (Local Area Network), uploads to programmed decision-making host computer for it.

Programmed decision-making host computer: for resolving aforementioned data information and obtaining control command, thus control BJUT-SHEV experiment porch, implements Vehicular turn, lifts throttle or step on the actions such as brake.

A kind of road information detection of view-based access control model and front vehicles recognition methods process flow diagram as shown in Figure 2, are realized by the software program be arranged in Measurement &control computer, comprise the following steps:

Step 1, Image semantic classification, idiographic flow as shown in Figure 3.

Step 1.1, coloured image gray processing.

If pixel color is RGB (R, G, B) in original color image, the pixel gray-scale value after process is Gray, and coloured image gray processing can be expressed as follows:

Gray＝R×0.299+G×0.587+B×0.144

Step 1.2, single OTSU method binary image.

OTSU method purposes in the process of pattern-recognition is relatively extensive, can selected threshold adaptively, distinguishes background and target area.First the characteristic parameter of gray level image is calculated:

μ＝ω ₀μ ₀+ω ₁μ ₁

σ ²(K)＝ω ₀(μ ₀-μ) ²+ω ₁(μ ₁-μ) ²

Wherein, ω ₀, ω ₁be respectively the probability of background and the appearance of target area pixel gray-scale value, μ ₀, μ ₁be respectively the average gray value of background and target area pixel, μ is the average statistical of general image gray scale, σ ²(K) be background area and target area between-group variance, K=1,2,3 ..., ask K when making variance obtain maximal value, obtain optimal threshold K.

Step 1.3, adopts Sobel operator to carry out rim detection.

Image border presents the transition of gray level usually, and this transition can describe with the differential of image.So be the one relatively commonly used based on the method for detecting image edge of differentiating operator class.Most of algorithm in these class methods uses filter template, even the center superposition of handled pixel and template, coefficients is with after corresponding pixel value weighting, and its result is as the Grad of this pixel.Mobile filter device template in view picture digital image matrix, just can obtain a width gradient map.The result of the method reflects the gradient of pixel grey scale change in digital picture, detects the edge of image according to the situation of change of gradient in gradient map.The present invention adopts Sobel operator to detect, and its principle template is as shown in table 1.

Table 1Sobel operator principle template

If image is after binaryzation, pixel coordinate is (i, j), carries out template computing to entire image, along the Grad Gx in x, y direction _{(i, j)}and Gy _{(i, j)}, think that when meeting following formula this point is marginal point:

|Gx|+|Gy|＞nThreshold

Wherein, nThreshold is threshold value, and the present embodiment gets nThreshold=138.

Step 1.4, image thinning process.

The edge that Sobel rim detection draws is comparatively thick, affects next step process.Image after edge detection below carries out refinement.Thinking is, thick edge means that edge pixel has certain width, only retains the pixel in the middle of this width, and the pixel " corrosion " of surrounding is fallen, just can reduce the width of edge pixel, refinement edge.

Each the white pixel point detected is judged, if be only less than the white pixel (k get in the detection be 7) of k in its eight neighborhood, then illustrate, this point is a brighter point, belong to the edge pixel in above-mentioned border width, so, such pixel is set to 0, the thinning processing to image can be completed.

Step 1.5, determines road area.

Determine the road area upper bound: downward retrieval from first pixel of each row of image, find first black pixel point of these row, mark its line number y _r, at y _rmaximal value on to increase the line number that m pixel obtain be the upper bound of processing region.The value of m is determined by experiment.The present embodiment m=15.

Determine the right boundary of road area: best straight line border should be two, the left and right straight line comprising whole road area, and a point in straight line, should on the inner boundary of Road.And according to characteristics of image, the Road on both sides, image lower end, must respectively on the RC both sides of image.So, from the center of image left, upwards search line by line from row bottom, using first white point finding as the point of first on road inner boundary, then using straight slope k as parameter, build straight-line equation.The span of left margin slope k is [0.2,6], and the increment with 0.1 increases, and calculates the number of the white point on this straight line according to straight-line equation, will obtain k value that maximum white the counts slope as this edge fitting straight line.After straight slope is determined, y value raises straight line the fitting a straight line that an increment b obtains left margin.Get the span of right margin slope k for [-0.5 ,-6], determine right margin fitting a straight line in the same way.Article two, the region in the middle of straight line is exactly road area.Experimental result shows, this region is very effective.

Step 2, the detection of lane line and early warning, idiographic flow as shown in Figure 6.

Step 2.1, determines road edge point.

Upwards search for until upper confinement boundary from last column of entire image, its longitudinal span is picture altitude height.Retrieve the line segment of white pixel line by line, write down the length l of n-th line segment _n.The line segment that end column coordinate is no more than 3/4 row of entire image is classified as left side Road, and the line segment that initial row coordinate is no less than entire image 1/4 row is classified as the right Road.Setting pixel distance threshold value d, when when the latter half of whole image, d=100; When the first half at image, d=30.Search in the Road sequence of left and right respectively between line segment adjacent rows effective row-coordinate i _jand i _j-1if the difference between them is greater than d, illustrates that this line segment belongs to noise, then reject from this sequence.Finally, from the Road sequence of left and right, find out the wherein the most obvious line segment of feature respectively, mark effective coordinate of this line segment, left side sequential stroke is (i _l, j _l), right side is (i _r, j _r).

Step 2.2, the matching of Road inner boundary.

Least square method is taked to carry out matching to lane line.If (x ₁, y ₁), (x ₂, y ₂) ... (x _n, y _n) be the one group of data provided under rectangular coordinate system, if there is x ₁< x ₂< ... < x _n, then this can be organized the discrete point set that data regard a function as.If treat that the straight-line equation of matching is: y=f (x)+ε.

F (x) represents ideally function when not having noise, and ε is noise; It is minimum that least square method is exactly the error sum of squares Q that noise is produced, that is:

$Q=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&epsiv;}}_i^2=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{[y_i-f\left(x_i\right)]}^2$

Use least square method carry out matching, its advantage be speed quickly, as long as traversal once just can calculate matched curve.

Step 2.3, Road turnings walking direction and depart from early warning.

Step 2.3.1, Road turnings walking direction.

The intersection point of note right boundary straight line is (x ₀, y ₀), the intersection point of left and right road inner boundary matched curve is (x ₁, y ₁).If there is x ₀< x ₁-δ ₀, then illustrate that road is being turned right; If there is x ₀> x ₁+ δ ₀, then illustrate that road is being turned left; If x ₁-δ ₀≤ x ₀≤ x ₁+ δ ₀, then road is described linearly.δ ₀for by testing determine one very little numerical value.

Step 2.3.2, departs from early warning.

The object of lane line matching is the position obtaining vehicle place.Therefore, a mathematical model can be set up to describe the positional information of vehicle.As shown in Figure 4, obtained the center identification line of left and right lane line by lane line fitting algorithm, i.e. the angular bisector of two the lane line angles in left and right.Track center line can represent with a linear equation in two unknowns, if:

y＝ax+b

Obtained by triangle angular bisector theorem and trigonometric function relation:

$a=\frac{k_1{k}_2+\sqrt{1+k_1^2{k}_2^2+k_1^2+k_2^2}-1}{k_1+k_2}$

$b=\frac{(b_1-b_2)(\sqrt{1+k_1^2{k}_2^2+k_1^2+k_2^2}-1)+a_1^2{b}_2-a_2^2{b}_1}{k_1^2-k_2^2}$

Wherein, k ₁, k ₂be respectively the slope of left and right lane line.

Can be obtained fom the above equation, the angle of vehicle heading and track centerline direction as shown in Figure 5.

Adopt based on the transversal displacement d of vehicle in current lane and the lane departure warning model of fleet angle θ.The advantage of this model is not rely on the change of lane width and the travel speed of vehicle, has higher real-time and accuracy, and as shown in Figure 5, treatment scheme as shown in Figure 6 for its model.According to the lateral attitude of vehicle in current lane and fleet angle, set up left and right deviation criterion.First extract the single-frame images information in automobile video frequency, this image is processed with reference to above-mentioned steps and calculates characteristic parameter d and θ of lane line.As d > d ₀and θ > θ ₀time, be judged as that left side is departed from, send early warning signal; As d > d ₀and θ <-θ ₀time, be judged as that right side is departed from, send early warning signal; If do not meet above-mentioned condition, do not send early warning signal.Lower two field picture is entered after processing.

Step 3, ROI region is extracted, as shown in Figure 8.

Step 3.1, self-adaption binaryzation.

Adopt two OTSU method to Image Segmentation Using, first use the threshold value T of OTSU method computed image entirety ₁; All pixels in traversing graph, with threshold value T ₁classify, be greater than T ₁classify as background.Filtering out background region, is less than T to grey scale pixel value in former figure ₁all pixels reuse OTSU method, obtain new threshold value T ₂.With T ₂for segmentation threshold carries out binaryzation again to image, be greater than T ₂classify as background, pixel value is set to 255; Be less than T ₂be set to object pixel, pixel value is set to 0.Gray level image chooses m ₁individual length and width are n ₁the region, road surface of individual pixel, m ₁, n ₁determine according to vide image resolution.The present embodiment vide image resolution is 640*480, m ₁get 5, n ₁get 25.Statistics m ₁the average gray value μ in region, individual road surface _iand standard deviation sigma _i, remove μ _ibe greater than μ ₀and σ _ibe greater than σ ₀region, μ ₀, σ ₀determined by experiment, the present embodiment μ ₀, σ ₀get 180 and 90 respectively, just can get rid of road surface regional window and be divided in situation above zebra stripes or index line, and set remaining road surface areal as N.Can in the hope of the average gray value in this N number of region through calculating and average variance can obtain optimal threshold is:

$T=\stackrel{\&OverBar;}{\mathrm{\&mu;}}-4\stackrel{\&OverBar;}{\mathrm{\&sigma;}}$

If because N is too small or T < 0 cannot carry out local gray-value computing, by as shown in the formula carrying out choosing of algorithm, this kind of mode can make system take into account robustness and real-time simultaneously.

Step 3.2, burn into expansion process: through process to multiple image, find some shadow regions obtained occur sometimes breach or with the situation such as surrounding environment adhesion, for above-mentioned situation, can adopt expansion, the morphological method such as corrosion process.Burn into expansion process is carried out to the binary image after segmentation before, the impurity in the former figure of treated image filtering and noise, and there is good shape facility.

Step 3.3, shadow region extraction and merging.

First from top to bottom, search for hatched reference position and final position line by line from left to right, thus determine its length and location.Then think when meeting following formula and have found hatched starting point x respectively _startand terminal x _end.

$\left\{\begin{array}{c}f(x-1,y)-f(x,y)=255\\ f(x,y)-f(x+1,y)=0\end{array}\right.$

$\left\{\begin{array}{c}f(x,y)-f(x-1,y)=0\\ f(x+1,y)-f(x,y)=255\end{array}\right.$

Vehicle bottom hacures length in the picture should in certain scope, according to the demarcation to camera position and parameter, the difference that this scope can be expert at along with hacures and changing, according to this feature, system is that often row chooses a threshold value, if the hacures length length=x detected _end-x _startthreshold value difference too much therewith, then filter out this hatched interference.Corresponding relational expression is as follows:

$w=\frac{w_p}{H}(y-\frac{\mathrm{height}}{2})$

Wherein, w is hacures length scalars (pixel) in the picture; w _pfor vehicle real wide (rice); H is the height on camera light wheelbase ground, value 1.6 meters; Y is the place line number (pixel) of target on image y direction; Height is the height (pixel) of image.When meeting following formula, then can think that these hacures are vehicle bottom shade:

0.75*w＜length＝x _start-x _end＜1.25*w

Step 3.4, ROI region is extracted.

Definition rectangular degree SQ is the ratio of region inner area and its boundary rectangle area, the more rectangular shape in larger then this region of SQ.Making quadrilateral measure QM is measuring of shadow region the ratio of width to height, as QM=1, can be similar to and think that quadrilateral is equilateral.Utilize SQ and QM to screen the above-mentioned shadow region detected, method is as follows:

In the process extracting ROI region, need the uncertainty of the height, width, position etc. considering shadow region, the ROI region that therefore first selection range is larger.Need entire vehicle to be included in region in the process of preliminary extraction, and consider that the change of intensity of illumination and angle can make shade at the bottom of car and car body itself present different proportionate relationships, concrete grammar is:

$\left\{\begin{array}{c}{R}_{v\_}x=R_{s\_}x-\frac{\mathrm{\&delta;}}{2}\\ R_{v\_}y=R_{s\_}y-R_{v\_}\mathrm{hright}\\ R_{v\_}\mathrm{width}=R_{s\_}\mathrm{width}+\mathrm{\&delta;}\\ R_{v\_}\mathrm{height}=\mathrm{\&lambda;}*R_{s\_}\mathrm{width}\end{array}\right.$

Wherein, (R _v_ x, R _v_ y) be ROI region lower left corner coordinate points, (R _s_ x, R _s_ y) be the lower left corner, shadow region coordinate points, R _v_ width, R _v_ height is respectively width and the height of ROI region, R _s_ width is the width of shadow region, parameter lambda=1.2, δ=50.

Step 4, vehicle ' s contour extracts, as shown in Figure 9.

Step 4.1, information entropy is screened.

The entropy of image obviously can increase in containing the candidate region of vehicle.For region, road surface, because its gray-scale value is more single, so the Pixel Information value comprised is lower.Therefore the less region of some entropy can be filtered out by this character.The height being located at said extracted is in the ROI region of h, and the entropy of often going is H (a), then its average is:

$\stackrel{\&OverBar;}{H}\left(a\right)=\frac{1}{h}\underset{j=1}{\overset{h}{\mathrm{\&Sigma;}}}H\left(a_j\right)$

Have chosen 80 different images in experiment to process, final selected threshold is T=3.1, then when time, can think that ROI region comprises information of vehicles, otherwise filtering can be carried out to this ROI region.

Step 4.2, symmetry is filtered.

Applied mathematics principle, if R (x) is the one-time continuous function in ROI region, then can be split as odd function R _o(x) and even function R _ex (), therefore the symmetry of a function just can be determined by the proportion shared by its isolated parity function.For the ROI region R that said extracted arrives, make area size be w × h, axis of symmetry is then for the y line section of image, the expression formula of odd function and even function is respectively:

$R_o(u,y)=\frac{1}{2}[R(x_s+u,y)-R(x_s-u,y)]$

$R_e(u,y)=\frac{1}{2}[R(x_s+u,y)-R(x_s-u,y)]$

Algorithm needs dual function correction, makes its revised average the same with odd function, levels off to zero, so that the relation both contrasting with energy function.After correction:

$R_e^{\&prime;}(u,y)=R_e(u,y)-\frac{1}{w}\underset{v=0}{\overset{w}{\mathrm{\&Sigma;}}}{R}_e(v,y)$

The energy function that can obtain odd function and even function is thus respectively:

$E\left[R_o(u,y)\right]=\underset{u=-w/2}{\overset{w/2}{\mathrm{\&Sigma;}}}{R}_o^2(u,y)$

$E\left[R_e^{\&prime;}(u,y)\right]=\underset{u=-w/2}{\overset{w/2}{\mathrm{\&Sigma;}}}{[R_e(u,y)-\frac{1}{w}\underset{v=0}{\overset{w}{\mathrm{\&Sigma;}}}{R}_e(v,y)]}^2$

Then can carry out metric calculation to the symmetry of the capable pixel of y thus:

$S\left[R(u,y)\right]=\frac{E\left[R_e^{\&prime;}(u,y)\right]-E\left[R_o(u,y)\right]}{E\left[R_e^{\&prime;}(u,y)\right]+E\left[R_o(u,y)\right]}$

Then have:

Final selected threshold is 0.15, as S > 0.15, then can be judged to be vehicle ROI region, otherwise delete this region.

Step 4.3, adopts the Robinson detective operators improved to carry out rim detection.

After above-mentioned steps process completes, obtain ROI vehicle region more accurately, in this region, adopt Robinson direction template operator to carry out the Grad that rim detection asks for pixel grey scale.Detect the point in region by the template of eight in table 1 respectively, wherein maximum output valve and direction thereof are just as gray-scale value and the direction thereof of gained after this some process, M in table ₀(↑), m ₂(→), m ₄(↓), m ₆(←), represent respectively just going up, upper right, positive right side, bottom right, just under, lower-left, a positive left side, upper left masterplate operator.

Table 2Robinson detective operators

When adopting Robinson operator to detect the pixel in ROI region, arbitrary pixel and M ₁the operational method of template is as follows:

e ₀＝p ₀+2×p ₁+p ₂-p ₅-2×p ₆-p ₇

Wherein, e ₀to e ₇be respectively target pixel points and template M ₀(↑) extremely carry out the result of rim detection computing; p ₀to p ₇be respectively be positioned at target pixel points upper left, just go up, upper right, positive left, the positive right side, lower-left, just under, the gray-scale value of bottom right pixel point.

To arbitrary pixel, need to be similar to above-mentioned computing 8 times, could accurate result have been obtained.As can be seen here, need to carry out 16 multiplyings and 40 plus and minus calculations to the detection of a point, this drags the travelling speed of slow system greatly.So propose to improve one's methods, introduce variable with the following method:

$\left\{\begin{array}{c}{x}_0=p_0-p_4\\ x_1=p_1-p_5\\ x_2=p_2-p_6\\ x_3=p_3-p_7\end{array}\right.,\left\{\begin{array}{c}{y}_0=x_0+x_1\\ y_1=x_1+x_2\\ y_2=x_2+x_3\\ y_3=x_3-x_0\end{array}\right.$

Carry out arithmetic speed optimization process, can obtain result is:

$\left\{\begin{array}{c}{e}_0=y_0+y_1\\ e_1=y_1+y_2\\ e_2=y_2+y_3\\ e_0=y_3-y_0\end{array}\right.,\left\{\begin{array}{c}{e}_4=-e_0\\ e_5={-e}_1\\ e_6={-e}_2\\ e_7={-e}_3\end{array}\right.$

After improvement, only need carry out 12 plus and minus calculations and just can complete detection to a pixel.Finally respectively horizontal and longitudinal Detection Information is extracted.

Step 4.4, adopts Hough transform to extract vehicle boundary line.

Hough transform extracts the comparatively general of straight line and the method for maturation, and weak point is that operand is larger.Applying Hough transform is herein carry out in ROI region, instead of extracts view picture figure, therefore reduces calculated amount largely.The present invention limits in straight line angle, thus accelerates arithmetic speed further.Extract in the process of transverse edge and make angle-5 ° of < θ < 5 °, extract in the process of longitudinal edge and make 60 ° of < θ < 120 °, order longitudinally left (right side) boundary line is as the criterion with the most left (right side) point, slope value is drawn close to infinite, then can obtain rectangle vehicle outer boundary.

Claims (9)

1. the road information of a view-based access control model detects and front vehicles recognition methods, the system realizing described method comprises: camera, the inner Measurement &control computer installing video frequency collection card, the LAN (Local Area Network) of being built by router, programmed decision-making host computer, intelligent vehicle experiment porch; Camera is arranged on the ceiling front dead center position of intelligent vehicle experiment porch, connects the video frequency collection card in described Measurement &control computer, Real-time Collection road image by USB; Measurement &control computer processes the LAN (Local Area Network) that the controling parameters that obtains built by router and passes to programmed decision-making host computer to image; Programmed decision-making host computer controls experiment porch after resolving above-mentioned information; It is characterized in that, described method performs following steps in Measurement &control computer:

Step 1, Image semantic classification;

Step 1.1, coloured image gray processing;

If pixel color is RGB (R, G, B) in original color image, the pixel gray-scale value after process is Gray, and coloured image gray processing can be expressed as follows:

Gray＝R×0.299+G×0.587+B×0.144

Step 1.2, single OTSU method binary image;

Calculate the characteristic parameter of gray level image:

μ＝ω ₀μ ₀+ω ₁μ ₁

σ ²(K)＝ω ₀(μ ₀-μ) ²+ω ₁(μ ₁-μ) ²

Wherein, ω ₀, ω ₁be respectively the probability of background and the appearance of target area pixel gray-scale value, μ ₀, μ ₁be respectively the average gray value of background and target area pixel, μ is the average statistical of general image gray scale, σ ²(K) be background area and target area between-group variance, K=1,2,3 ..., ask K when making variance obtain maximal value, obtain optimal threshold K;

Step 1.3, adopts Sobel operator to carry out rim detection;

If image pixel coordinate after binaryzation is (i, j), template computing is carried out to entire image, thus obtain the Grad Gx of each pixel along x, y direction _{(i, j)}and Gy _{(i, j)}, think that when meeting following formula this point is marginal point:

|Gx|+|Gy|＞nThreshold

Wherein, nThreshold is threshold value;

Step 1.4, image thinning process;

Thick edge pixel has certain width, only retain the pixel in the middle of this width, the pixel " corrosion " of surrounding is fallen: each white pixel point that step 1.3 detects is judged, if be only less than the white pixel of k in its eight neighborhood, illustrate that this point is a brighter point, belong to the edge pixel in border width, such pixel is set to 0; K is by testing the positive integer determined;

Step 1.5, determines road area;

Determine the road area upper bound: downward retrieval from first pixel of each row of image, find first black pixel point of these row, mark its line number y _r, at y _rmaximal value on to increase the line number that m pixel obtain be the upper bound of processing region; The value of m is determined by experiment;

Determine the right boundary of road area: from the center of image left, upwards search line by line from row bottom, using first white point finding as the point of first on road inner boundary, then using straight slope k as parameter, build straight-line equation; Calculate the number of the white point on this straight line according to straight-line equation, will k value that maximum white the counts slope as this edge fitting straight line be obtained; After straight slope is determined, y value raises straight line the fitting a straight line that an increment b obtains left margin; Determine right margin fitting a straight line in the same way; Article two, the region in the middle of straight line is exactly road area;

Step 2, lane detection and depart from early warning;

Take the method determination lane boundary point retrieved line by line, adopt least square method to carry out matching to frontier point, obtain the quafric curve describing track; Judge the direction of vehicle front Road turnings, to vehicle, whether run-off-road line carries out early warning;

Step 3, ROI region is split;

Adopt the method dividing vehicle bottom shadow that road area gray scale combines with two OTSU, corrosion expansion process is carried out to segmentation image, fills gap regions, obtain vehicle ROI region based on shade at the bottom of car;

Step 4, vehicle ' s contour identification;

Carry out screening with the multiple features that information entropy and symmetry are Primary Reference foundation to region; Adopt the Robinson operator improved to process the part that remains after screening, ask for grey scale change Grad, and with Hough transform method identification vehicle outer contour.

2. the road information of a kind of view-based access control model according to claim 1 detects and front vehicles recognition methods, it is characterized in that, determines that the method for lane boundary point is as follows described in step 2:

Upwards search for until upper confinement boundary from last column of entire image, its longitudinal span is picture altitude height; Retrieve the line segment of white pixel line by line, write down the length l of n-th line segment _n; The line segment that end column coordinate is no more than 3/4 row of entire image is classified as left side Road, and the line segment that initial row coordinate is no less than entire image 1/4 row is classified as the right Road; The effective row-coordinate i between line segment adjacent rows is searched respectively in the Road sequence of left and right _jand i _j-1if the difference between them is greater than pixel distance threshold value d, illustrates that this line segment belongs to noise, then reject from this sequence; Finally, from the Road sequence of left and right, find out the wherein the most obvious line segment of feature respectively, mark effective coordinate of this line segment, left side sequential stroke is (i _l, j _l), right side is (i _r, j _r).

3. the road information of a kind of view-based access control model according to claim 1 detects and front vehicles recognition methods, it is characterized in that, judges that the method in the direction of vehicle front Road turnings is as follows described in step 2:

The intersection point of note right boundary straight line is (x ₀, y ₀), the intersection point of left and right road inner boundary matched curve is (x ₁, y ₁); If there is x ₀< x ₁-δ ₀, then illustrate that road is being turned right; If there is x ₀> x ₁+ δ ₀, then illustrate that road is being turned left; If x ₁-δ ₀≤ _x0≤x ₁+ δ ₀, then road is described linearly; δ ₀for by testing determine one very little numerical value.

4. the road information of a kind of view-based access control model according to claim 1,2 or 3 detects and front vehicles recognition methods, it is characterized in that, described in step 2 to vehicle whether run-off-road line to carry out the method for early warning as follows:

The center identification line of left and right lane line is obtained, i.e. the angular bisector of two the lane line angles in left and right by lane line fitting algorithm; Track center line is represented with a linear equation in two unknowns:

y＝ax+b

$a=\frac{k_1{k}_2+\sqrt{1+{k_1}^2{k}_2^2+{k_1}^2+k_2^2}-1}{k_1+k_2}$

$b=\frac{(b_1-b_2)(\sqrt{1+{k_1}^2{k}_2^2+{k_1}^2{k}_2^2}-1)+{a_1}^2{b}_1-a_2^2{b}_1}{{k_1}^2-k_2^2}$

Wherein, k ₁, k ₂be respectively the slope of left and right lane line;

The angle of vehicle heading and track centerline direction

If the transversal displacement of vehicle in current lane is d; As d > d ₀and θ > θ ₀time, be judged as that left side is departed from, send early warning signal; As d > d ₀and θ <-θ ₀time, be judged as that right side is departed from, send early warning signal; If do not meet above-mentioned condition, do not send early warning signal.

5. the road information of a kind of view-based access control model according to claim 1 detects and front vehicles recognition methods, and it is characterized in that, described in step 3, the method for dividing vehicle bottom shadow is as follows:

First the threshold value T of OTSU method computed image entirety is used ₁; All pixels in traversing graph, with threshold value T ₁classify, be greater than T ₁classify as background; Filtering out background region, is less than T to grey scale pixel value in former figure ₁all pixels reuse OTSU method, obtain new threshold value T ₂; With T ₂for segmentation threshold carries out binaryzation again to image, be greater than T ₂classify as background, pixel value is set to 255; Be less than T ₂be set to object pixel, pixel value is set to 0; Gray level image chooses m ₁individual length and width are n ₁the region, road surface of individual pixel, m ₁, n ₁determine according to vide image resolution; Statistics m ₁the average gray value μ in region, individual road surface _iand standard deviation sigma _i, remove μ _ibe greater than μ ₀and σ _ibe greater than σ ₀region, μ ₀, σ ₀determined by experiment, just can get rid of road surface regional window and be divided in situation above zebra stripes or index line, and set remaining road surface areal as N; Can in the hope of the average gray value in this N number of region through calculating and average variance can obtain optimal threshold is:

$T=\stackrel{\&OverBar;}{\mathrm{\&mu;}}-4\stackrel{\&OverBar;}{\mathrm{\&sigma;}}$

If because of N is too small or T < 0 time cannot carry out local gray-value computing, by as shown in the formula carrying out choosing of algorithm:

6. a kind of road information of view-based access control model detects and front vehicles recognition methods according to claim 1 or 5, it is characterized in that, as follows based on the method for shade acquisition vehicle ROI region at the bottom of car described in step 3:

From top to bottom, search for hatched reference position and final position line by line from left to right, thus determine its length and location; Then think when meeting following formula and have found hatched starting point x respectively _startand terminal x _end:

$\left\{\begin{array}{c}f(x-1,y)-f(x,y)=255\\ f(x,y)-f(x+1,y)=0\end{array}\right.$

$\left\{\begin{array}{c}f(x,y)-f(x-1,y)=0\\ f(x+1,y)-f(x,y)=255\end{array}\right.$

Often row chooses a threshold value, and when meeting following formula, these hacures are vehicle bottom shade:

0.75*w＜length＝x _start-x _end＜1.25*w

$w=\frac{w_p}{H}(y-\frac{\mathrm{height}}{2})$

Wherein, w is hacures length scalars in the picture, and unit is pixel; w _pfor vehicle is real wide; H is the height on camera light wheelbase ground; Y is the place line number of target on image y direction, and unit is pixel; Height is the height unit of image is pixel;

Definition rectangular degree SQ is the ratio of region inner area and its boundary rectangle area, the more rectangular shape in larger then this region of SQ; Making quadrilateral measure QM is measuring of shadow region the ratio of width to height, as QM=1, can be similar to and think that quadrilateral is equilateral; Utilize SQ and QM to screen the above-mentioned shadow region detected, method is as follows:

In the process extracting ROI region, consider the uncertainty of the height of shadow region, width, position etc., the ROI region that first selection range is larger; In the process of preliminary extraction, entire vehicle is included in region, and considers that the change of intensity of illumination and angle can make shade at the bottom of car and car body itself present different proportionate relationships, concrete grammar is:

$\left\{\begin{array}{c}{R}_{v\_}x=R_{s\_}x-\frac{\mathrm{\&delta;}}{2}\\ R_{v\_}y=R_{s\_}y-R_{v\_}\mathrm{height}\\ R_{v\_}\mathrm{width}=R_{s\_}\mathrm{width}+\mathrm{\&delta;}\\ R_{v\_}\mathrm{height}=\mathrm{\&lambda;}*R_{s\_}\mathrm{width}\end{array}\right.$

Wherein, (R _{v_}x,R _{v_}y) be ROI region lower left corner coordinate points, (R _{s_}x,R _{s_}y) be the lower left corner, shadow region coordinate points, R _{v_}width, R _{v_}height is respectively width and the height of ROI region, R _{s_}width is the width of shadow region, parameter lambda=1.2, δ=50.

7. the road information of a kind of view-based access control model according to claim 1 detects and front vehicles recognition methods, it is characterized in that, the method for carrying out screening with the multiple features that information entropy and symmetry are Primary Reference foundation to region described in step 4 is as follows:

(1) information entropy screening

The entropy of image obviously increases in containing the candidate region of vehicle; For region, road surface, because its gray-scale value is more single, so the Pixel Information value comprised is lower; The less region of some entropy is filtered out by this character; The height being located at said extracted is in the ROI region of h, and the entropy of often going is H (a), then its average is:

$\stackrel{\&OverBar;}{H}\left(a\right)=\frac{1}{h}\underset{j=1}{\overset{h}{\mathrm{\&Sigma;}}}H\left(a_j\right)$

When time, think that ROI region comprises information of vehicles, otherwise filtering is carried out to this ROI region; T is by testing the threshold value determined;

(2) symmetry is filtered

If R (x) is the one-time continuous function in ROI region, be split as odd function R _o(x) and even function R _e(x); For the ROI region R that said extracted arrives, make area size be w × h, axis of symmetry is then for the y line section of image, the expression formula of odd function and even function is respectively:

$R_o(u,y)=\frac{1}{2}[R(x_s+u,y)-R(x_s-u,y)]$

$R_e(u,y)=\frac{1}{2}[R(x_s+u,y)-R(x_s-u,y)]$

Dual function is revised, and makes its revised average the same with odd function, levels off to zero, so that the relation both contrasting with energy function; After correction:

$R_e^{\&prime;}(u,y)=R_e(u,y)-\frac{1}{w}\underset{v=0}{\overset{w}{\mathrm{\&Sigma;}}}{R}_e(v,y)$

The energy function obtaining odd function and even function is thus respectively:

$e\left[R_o(u,y)\right]=\underset{u=-w/2}{\overset{w/2}{\mathrm{\&Sigma;}}}{R}_o^2(u,y)$

$E\left[R_e^{\&prime;}(u,y)\right]=\underset{u=-w/2}{\overset{w/2}{\mathrm{\&Sigma;}}}{[R_e(u,y)-\frac{1}{w}\underset{v=0}{\overset{w}{\mathrm{\&Sigma;}}}{R}_e(v,y)]}^2$

Metric calculation is carried out to the symmetry of the capable pixel of y:

$S\left[R(u,y)\right]=\frac{E[R_e^{\&prime;}(u,y)-E[R_o(u,y)]}{E\left[R_e^{\&prime;}(u,y)\right]+E\left[R_o(u,y)\right]}$

Then have:

As S > S ₀time, then can be judged to be vehicle ROI region, otherwise delete this region; S ₀for the threshold value determined by experiment.

8. the road information of a kind of view-based access control model according to claim 1 detects and front vehicles recognition methods, and it is characterized in that, the method improving Robinson operator described in step 4 is as follows:

When adopting Robinson operator to detect the pixel in ROI region, arbitrary pixel and M ₁the operational method of template is as follows:

e ₀＝p ₀+2×p ₁+p ₂-p ₅-2×p ₆-p ₇

Wherein, e ₀~ e ₇be respectively target pixel points and template M ₀(↑) ~ M ₇ carry out the result of rim detection computing; p ₀~ p ₇be respectively be positioned at target pixel points upper left, just go up, upper right, positive left, the positive right side, lower-left, just under, the gray-scale value of bottom right pixel point;

The method improved is to introduce variable:

$\begin{array}{cc}\left\{\begin{array}{c}{x}_0=p_0-p_4\\ x_1=p_1-p_5\\ x_2=p_2-p_6\\ x_3=p_3-p_7\end{array}\right.,& \left\{\begin{array}{c}{y}_0=x_0+x_1\\ y_1=x_1+x_2\\ y_2=x_2+x_3\\ y_3=x_3-x_0\end{array}\right.\end{array}$

Carry out arithmetic speed optimization process, can obtain result is:

$\begin{array}{cc}\left\{\begin{array}{c}{e}_0=y_0+y_1\\ e_1=y_1+y_2\\ e_2=y_2+y_3\\ e_0=y_3-y_0\end{array}\right.,& \left\{\begin{array}{c}{e}_4={-e}_0\\ e_5={-e}_1\\ e_6={-e}_2\\ e_7={-e}_3\end{array}\right.\end{array}$

Finally respectively horizontal and longitudinal Detection Information is extracted.

9. the road information of a kind of view-based access control model according to claim 1 or 8 detects and front vehicles recognition methods, it is characterized in that, adopts the method for Hough transform method identification vehicle outer contour to carry out in ROI region, to reduce calculated amount described in step 4; Straight line angle limits, thus accelerates arithmetic speed further; Extract in the process of transverse edge and make angle-5 ° of < θ < 5 °, extract in the process of longitudinal edge and make 60 ° of < θ < 120 °, order is longitudinally as the criterion with the most left and right point left and right boundary line respectively, slope value is drawn close to infinite, can obtain rectangle vehicle outer boundary.