CN105966314A - Lane departure pre-warning method based on double low-cost cameras - Google Patents

Abstract

The invention belongs to the field of aided driving, and relates to a lane departure pre-warning method based on double low-cost cameras. A lane departure pre-warning algorithm is achieved by using the double low-cost cameras. The method includes the steps that real-time images are obtained through the cameras below a left vehicle lug and a right vehicle lug of an intelligent vehicle; pretreatment, IPM and feature point extraction are carried out on the two images; feature points of candidate lane lines are clustered to obtain feature points of lane lines; Hoff straight line conversion detection is carried out on the images; the rightmost lane line of the left camera and the leftmost lane line of the right camera are extracted to calculate departure distance. The departure distance is calculated based on a model set up on the IPM images obtained after image inverse perspective conversion, and the problem that detection of the camera at one side fails is effectively solved. The double low-cost cameras are used for replacing a traditional single high-definition industrial camera, cost is greatly reduced, and the algorithm is low in operation cost and high in accuracy and real-time performance.

Description

Lane departure warning methods based on double low cost photographic head

Technical field

The present invention is lane departure warning methods based on double low cost photographic head, belongs to intelligence auxiliary driving technology neck Territory.

Background technology

Lane departure warning is a critical aspects of intelligence DAS (Driver Assistant System) research, along with safe driving, intelligence are handed over Logical increasingly by people's concern and attention, lane departure warning method research has become as the hot issue of research.Deviation The design of method for early warning is according to road ahead environment and this truck position relation, it is judged that automotive run-off-road distance is also carried out in time Remind, prevent owing to driver neglects the generation of the Lane Departure caused.Show according to research, lane departure warning method Realization and application can avoid the generation of deviation vehicle accident of 30-70%, to improving traffic safety, reduce and hand over Interpreter thus cause be significant.Its design head-up position under monocular-camera of traditional lane departure warning method enters OK, including image capture module, lane detection module and warning module.Typically first pass through imageing sensor and obtain vehicle row Road image real-time during sailing, then the run-off-road line that lane detection is obtained vehicle on hardware processing platform Whether distance parameters etc., then have the probability that deviation occurs to be estimated by method for early warning to current running state, Finally show early warning information.

Summary of the invention

The present invention utilizes the deviation that the photographic head of two low costs immediately below vehicle car ear (rearview mirror) is carried out Method for early warning, it is possible to lane departure warning information is provided correctly, in real time.Its advantage be the cost requirement to photographic head very Low, and ensure that computing overhead minimum and real-time.The design of algorithm is photographic head based on double low costs, it is possible to adapt to car Produce family to cost, real-time and requirement accurately, market prospect will be the best.

To achieve these goals, this invention takes following technical scheme:

Step 1: sensor is installed；

Video camera is separately mounted to the underface of intelligent vehicle sided mirror unit parallel with the longitudinal coordinate axle of car body, When video camera is installed, should ensure that two video cameras must be positioned in same plane, and vehicle both sides lane line image can be collected；

Step 2: camera calibration；

Demarcating two photographic head respectively, the field range of demarcation is: wide 600cm, remote 1000cm.Adjust photographic head Left and right, upper lower angle, after enabling demarcation, the position of left and right photographic head (to require that the picture obtained is same in the same plane Process in plane)；

Step 3: the pretreatment of image；

According to lane line feature, first image being carried out gray proces, the formula of gray proces is

G_ray=R*0.5+G*0.5, wherein R, G represent red blue channel component value, G respectively_rayRepresent the pixel after conversion Gray value.Then image is carried out medium filtering, this experiment uses the square field of 3 × 3 be filtered image processing.

Step 4:IPM；

Gray level image carrying out inverse perspective process and obtains birds-eye view picture, perspective matrix is calibrated H-matrix.

Step 5: feature point extraction；

The gray value of lane line is higher than the value on its both sides, forms a crest；Present and from left to right fall after rising Trend；When in lane line region, average is higher, average differs bigger with its summit value.We utilize these characteristics, by calculating The change of adjacent image pixels judges the edge of lane line.

Step 5-1: calculate the average of adjacent image pixels；

If certain point be (x, y), meet y ∈ [0, h) and x ∈ [2, w-2).X, y are the columns and rows of pixel respectively, and w is figure The width of picture, h is the height of image.The average then having an adjacent image pixels is:

${\mathrm{avg}}_{(x,y)}=\frac{1}{t}\underset{i=-t/2}{\overset{i=t/2}{\Σ}}f(i+x,y)$

Wherein t ∈ [1,3,5,7 ...], t=5 can obtain good effect.

Step 5-2: calculating edge extracting threshold value T, its computing formula is expressed as follows:

$T=\left\{\begin{array}{cc}\frac{{\mathrm{avg}}_{(x,y)}}{12}& {\mathrm{avg}}_{(x,y)}>200\\ \frac{{\mathrm{avg}}_{(x,y)}}{5}& 50<{\mathrm{avg}}_{(x,y)}\&le;200\\ \frac{{\mathrm{avg}}_{(x,y)}}{10}& other\end{array}\right.$

Wherein avg_(x,y)For point (x, y) average near horizontal line.

Step 5-3: calculate liter height e at edge_pWith fall height e_v。

e_p∈{f(x+2,y)-f(x,y)>T}

e_v∈{f(x+2,y)-f(x,y)<-T}

(x is y) that ((x+2 y) is the gray value of spaced points of this horizontal direction to f to current point for x, gray value y) to f.

Step 5-4: the liter height of lane line and fall height are to occur in pairs in the picture, and between meet certain Distance.Relatively rise height and the width of fall height, reject ungratified point.

Δ w=e_p(x)-e_v(x)

If Δ w > W, then it is assumed that be it is unlikely that lane line, then to give up.Wherein, e_p(x) and e_vX () represents respectively Rising height and the row pixel coordinate of fall height, W is the maximum number of pixels that lane line occupies in the picture.

Step 6: clustering method based on lane line.

The lane line candidate domain obtained after feature point extraction, clusters lane line characteristic point.Assuming that starting point is (x₀, y₀), set the square window of a m*n, i.e. m row n row, the most from bottom to top, scanning characteristic point from left to right.If The characteristic point that in window, often row runs into for the first time is validity feature point, and other characteristic points of this row will be not scanned；If this Not having characteristic point in individual window, window will translate n-1 row, not have characteristic point, cluster to terminate if accumulative more than 2n row；If y > h Or x > w, cluster terminates, and wherein w is the width of image, and h is the height of image.

Step 7: the characteristic point generated after cluster is carried out Hough transform respectively, and extends straight line, take y=nearby respectively 180pix, at a distance y=120pix.

Step 8: calculate abscissa distance D of these 2 isolated edges at distance respectively^rAnd D^l, then have W=D^r+D^l, so There is Δ D=D^r-D^l, as Δ D > 0, inclined left lane line；When Δ D < when 0, inclined right lane line；As Δ D=0, do not occur Skew.

Step 9: the result of detection shown or is sent to policymaker, in order to being adjusted in time；

When wherein calculating at distance, may have a relative error, this error utilizes the seriality of intra-frame trunk, Can reduce.I.e. as continuous 5 frame images above error delta D ' > T_D, T_DTake 30pix, then reinitialize, and by up-to-date detection Result is issued or is shown to policymaker；Pix represents pixel.

Further illustrate:

1. two video cameras installed in step 1 must be positioned in same plane, and can collect vehicle both sides lane line Image；Do not require that there is automatic exposure, the function such as area-of-interest can be arranged, AWB；

2. the field range demarcated in step 2 to be determined according to the ultimate resolution of photographic head, and in experiment, we use Sieve skill C170, ultimate resolution is 640*480, therefore, is designed as wide 600cm, and the visual field of remote 1000cm is the most suitable.

3. step 3, step 4 are carried out on a hardware platform, it is desirable to the platform of work at least meets more than internal memory 2G, place Reason more than device 2.1Hz.

4. the feature point extraction of step 5 is the important module of lane detection, when calculating edge extracting threshold value T, and basis Real road is adjusted, and under super expressway, the threshold value that we set will be most suitable.Additionally, it is unconspicuous at lane line In the case of, may be greatly reduced even empty, for this situation, Wo Men by lane line information after feature point extraction Algorithm devises the increase reliable point module of lane line.This requires when being IPM, and a image of reservation is as backup, in feature When point extracts less, we contrast two parts of images, increase lane line information in region.Formula is as follows:

{(x₁,y₁),…,(x_n,y_n)}∈{(x₁,y₁),…,(x_m,y_m),(x_m+1,y_m+1)…,(x_m+k,y_m+k)}

5. the clustering method in step 6 is can to determine for the result in step 5, for left and right photographic head, it is known that The lane line information of the left-hand lane line that left photographic head obtains, what right photographic head obtained is the lane line information on right side.At algorithm In design, we are it is of concern that current lane line information, and therefore, the IPM image of left photographic head is most concerned is proximate to right side The lane line characteristic point at edge, the most concerned lane line characteristic point being proximate to left side edge of the IPM image of right photographic head.In choosing Take starting point (x₀,y₀) time it is noted that belong to the image of which photographic head.

6. in step 7, the lane line of all bends and straight way is contrasted by we, finds under our model, institute Having bend the most all can be approximately straight line, 120-180 on hand is the most suitable for selection range.

7. the computational methods of the ratio k of the actual range in step 8 and pixel distance: intelligent vehicle is stopped in track And parallel with lane line, from birds-eye view picture, calculate horizontal pixel distance P (unit pixel) in two adjacent lane lines, so After measure the width W (unit cm), then k=W/p in a track；.

The invention has the beneficial effects as follows:

The present invention, by feasible technical scheme, not only can meet real-time during actual application, it can also be ensured that identifies Accuracy is more than 95%, additionally significantly reduces cost.

Accompanying drawing explanation

The schematic flow sheet of Fig. 1 present invention

Fig. 2 camera calibration figure

Fig. 3-(a) bend turnout gray-scale map

Fig. 3-(b) straight way gray-scale map

Fig. 3-(c) bend gray-scale map

Fig. 4-(a) bend IMP figure-1

Fig. 4-(b) bend IMP figure-2

Fig. 4-(c) straight way IMP figure-1

Fig. 4-(d) straight way IMP figure-2

Fig. 5-(a) bend feature point extraction-1

Fig. 5-(b) bend feature point extraction-2

Fig. 5-(c) straight way I feature point extraction-1

Fig. 5-(d) straight way feature point extraction-2

Fig. 6 lane line dendrogram picture

Fig. 7 Hough transform image (extending)

Fig. 8 deviation distance result figure

Detailed description of the invention

The method using the present invention, provides the example of an indefiniteness, concrete real to the present invention further in conjunction with Fig. 1 The process of executing illustrates.The present invention realizes at intelligent vehicle platform, intelligent vehicle test site, in order to ensure driving intelligent vapour Car and personal security, platform used and place are intelligent driving technology specialty experiment porch and test site.Used Some current techiques such as image acquisition, image conversion etc. are not in narration in detail.

Embodiments of the present invention are as follows:

1. requiring to install video camera according to step 1, required device installed by platform used by this example, it is only necessary to slightly adjust Just can test.

2. according to step 2, photographic head is demarcated, photographic head is finely adjusted whole.

3. realize according to the detailed step of 3,4,5,6,7,8,9.

Claims (1)

1. lane departure warning methods based on double low cost photographic head, it is characterised in that comprise the following steps:

Step 1: sensor is installed；

Video camera is separately mounted to the underface of intelligent vehicle sided mirror unit parallel with the longitudinal coordinate axle of car body, installs During video camera, should ensure that two video cameras must be positioned in same plane, and the lane line image of vehicle both sides can be collected；

Step 2: camera calibration；

Demarcating two photographic head respectively, the field range of demarcation is: wide 600cm, remote 1000cm；Adjust a left side for photographic head Lower angle right, upper, makes after demarcation that the position of left and right photographic head will be in the same plane；

Step 3: the pretreatment of image；

First image being carried out gray proces, the formula of gray proces is G_ray=R*0.5+G*0.5, wherein R, G represent red indigo plant respectively Channel components value, G_rayRepresent the gray value of the pixel after conversion；Then image is carried out medium filtering, use 3 × 3 square Image is filtered processing by field；

Step 4:IPM；

Gray level image carrying out inverse perspective process and obtains birds-eye view picture, perspective matrix is calibrated H-matrix；

Step 5-1: calculate the average of adjacent image pixels；

If certain point be (x, y), meet y ∈ [0, h) and x ∈ [2, w-2)；X, y are the columns and rows of pixel respectively, and w is image Width, h is the height of image；The average then having an adjacent image pixels is:

${\mathrm{avg}}_{(x,y)}=\frac{1}{t}\underset{i=-t/2}{\overset{i=t/2}{\&Sigma;}}f(i+x,y)$

Wherein t=5；

Step 5-2: calculating edge extracting threshold value T, its computing formula is expressed as follows:

$T=\left\{\begin{array}{cc}\frac{{\mathrm{avg}}_{(x,y)}}{12}& {\mathrm{avg}}_{(x,y)}>200\\ \frac{{\mathrm{avg}}_{(x,y)}}{5}& 50<{\mathrm{avg}}_{(x,y)}\&le;200\\ \frac{{\mathrm{avg}}_{(x,y)}}{10}& other\end{array}\right.$

Wherein avg_(x,y)For point (x, y) average near horizontal line；

Step 5-3: calculate liter height e at edge_pWith fall height e_v；

e_p∈ { f (x+2, y)-f (x, y) ＞ T}

e_v∈ { f (x+2, y)-f (x, y) ＜-T}

(x is y) that ((x+2 y) is the gray value of spaced points of this horizontal direction to f to current point for x, gray value y) to f；

Step 5-4: compare liter height and the width of fall height, reject ungratified point；

Δ w=e_p(x)-e_v(x)

If Δ w ＞ W, then it is assumed that be it is unlikely that lane line, then to give up；Wherein, e_p(x) and e_vX () represents that liter becomes respectively Point and the row pixel coordinate of fall height, W is the maximum number of pixels that lane line occupies in the picture；

Step 6: clustering method based on lane line；

The lane line candidate domain obtained after feature point extraction, clusters lane line characteristic point；Assuming that starting point is (x₀,y₀), if The square window of a fixed m*n, i.e. m row n row, the most from bottom to top, scanning characteristic point from left to right；If in window The characteristic point that often row runs into for the first time is validity feature point, and other characteristic points of this row will be not scanned；If this window Inside not having characteristic point, window will translate n-1 row, not have characteristic point, cluster to terminate if accumulative more than 2n row；If y ＞ h or x ＞ W, cluster terminates, and wherein w is the width of image, and h is the height of image；

Step 7: the characteristic point generated after cluster is carried out Hough transform respectively, and extends straight line, take y=nearby respectively 180pix, at a distance y=120pix；

Step 8: calculate abscissa distance D of these 2 isolated edges at distance respectively^rAnd D^l, W=D^r+D^l, then have Δ D=D^r- D^l, as Δ D ＞ 0, inclined left lane line；As Δ D ＜ 0, inclined right lane line；As Δ D=0, do not offset；

Step 9: the result of detection shown or is sent to policymaker, in order to being adjusted in time；When continuous 5 frame images above are missed Difference Δ D ' ＞ T_D, T_DTake 30pix, then reinitialize, and up-to-date testing result is issued or is shown to policymaker.