CN102314599A - Identification and deviation-detection method for lane - Google Patents

Abstract

The present invention relates to a lane recognition deviation detection method, comprising the following steps: (1) acquiring a lane image, and performing preprocessing on the lane image; (2) performing Canny operator edge detection on the preprocessed lane image, Obtain the lane edge image; (3) According to the obtained lane edge image based on the lane tracking method of the Kalman predictor, determine the position of the exit lane line, select the Kalman prediction area, use the distance discrimination method to filter out the effective point set, and finally at the point (4) According to the obtained lane parameters, use the Hough transform with straight line fitting to extract the lane line; (5) use the starting point position determined in step (3) and the dynamic prediction of the lane, in Count the number of background points and lane line points in the Kalman prediction area, and calculate the ratio between background points and lane line points. The invention can quickly and stably realize the monitoring of the condition of the lane.

Description

A lane recognition deviation detection method

technical field

The invention relates to the technical field of lane recognition, in particular to a lane recognition deviation detection method.

Background technique

Fatigue driving is one of the important hidden dangers of today's traffic safety. When the driver is fatigued, his ability to perceive the surrounding environment, the ability to judge the situation and the ability to control the vehicle all decline to varying degrees, so traffic accidents are prone to occur. In the anti-fatigue safe driving intelligent system, the extraction and processing of lane lines, as an important indicator for judging whether a person is fatigued, is a key link in the entire system. Therefore, the lane line is separated from the lane picture and processed in real time to calculate the parameters and determine the current state of the vehicle in the lane, so as to monitor the vehicle in real time and effectively, and then make an effective judgment on the driver's state , Avoid traffic accidents by reminding.

After searching the existing technology, it was found that the application number of the Chinese invention patent "A Method, Device and System for Determining Lane Departure" is 201010033839.7, and the publication number is CN 101804814A. This patent discloses a lane departure detection method. Firstly, edge detection is performed on the lane image to obtain the gradient magnitude and gradient direction of each point in the lane image, and then determine the gradient direction of the lane boundary. Using the gradient size and gradient direction of each pixel point and the gradient direction of the lane boundary to fit the boundary with a straight line to obtain a lane boundary line. In this recognition method, after the lane line is transformed, there may be too many straight lines that cannot be fitted well, and when the lane has a certain width and curvature, the above method has a large error and cannot better identify the current lane status. The application number of the Chinese invention patent "A Simulated Lane Recognition Method Based on Statistical Threshold Segmentation" is 200710168943.5, and the publication number is CN 101187976A. This patent uses the method of making difference between the pixel values of any data point in the image, uses the black and white boundary threshold to segment the image, and calculates the threshold by counting the percentage of lanes in the image. In the process of dealing with the selection of black-white boundary threshold, this method may have fuzzy deviation, and the error is relatively large in actual application, and it cannot achieve real-time and self-adaptive effects. Chinese invention patent "A method for fast multi-threshold segmentation of grayscale images", the application number is 200810064059.1, and the publication number is CN 101236607A. This invention proposes a histogram-based grayscale image threshold segmentation method. Due to the existence of multiple targets, the grayscale histogram used in this method has multiple peaks, so the grayscale corresponding to the midpoint of two adjacent peaks is used as the threshold The segmentation threshold. Due to the existence of edge fluctuations, this method has poor resistance to image interference noise or uneven illumination, and it is prone to large errors in threshold selection, which greatly limits its application. The Japanese patent "TRAFFIC LANE BOUNDARY DECISIONDEVICE", the application number is JP2005258846A, this patent proposes a lane identification method, which limits the lane to be in an ideal state, and has strict requirements on the environment, and does not address changes in light, weather, etc. come up with effective solutions. US patent "Vehicle and Lane Mark Detection Device", the application number is US2009167864A1. This patent proposes a lane image processing method based on the CCD imaging principle. This method is prone to dynamic blur errors when the image resolution is low and the light changes, causing large recognition errors. When splitting, the real-time effect is not achieved.

Contents of the invention

The technical problem to be solved by the present invention is to provide a lane recognition deviation detection method, so that it can quickly and stably realize the monitoring of lane conditions.

The technical solution adopted by the present invention to solve the technical problem is: provide a lane recognition deviation detection method, comprising the following steps:

(1) Acquire a lane image, and preprocess the lane image;

(2) Carry out Canny operator edge detection on the preprocessed lane image to obtain the lane edge image;

(3) According to the obtained lane edge image based on the lane tracking method of the Kalman predictor, determine the position of the exit lane line, select the Kalman prediction area, and use the distance discrimination method to screen out the effective point set, and finally based on the optimized point set Extract the lane parameters;

(4) According to the obtained lane parameter, utilize the Hough transformation with straight line fitting to extract the lane line;

(5) Using the starting point position determined in step (3) and the dynamic prediction of the lane, count the number of background points and lane line points in the Kalman prediction area, and calculate the ratio between the background point and lane line points.

The preprocessing in the step (1) also includes the following substeps:

(11) ROI processing is carried out to the acquired lane image;

(12) Ashing processing is carried out to the lane image after ROI processing;

(13) Carry out median filter processing to the lane image after graying processing;

(14) Contrast enhancement processing is carried out to the lane image after median filtering;

(15) The lane image after the contrast enhancement is divided into m levels, and the probability of occurrence of pixels of each level is reflected and analyzed in the form of a histogram, wherein, m>1;

(16) Utilize the difference in grayscale between the target object and the background in the lane image, obtain the first binarization segmentation threshold based on prior knowledge, and automatically obtain the next segmentation threshold in an adaptive way, so as to determine the every point of .

The Hough transformation in the described step (4) also includes the following steps:

(41) Determine the polar coordinate system, and correspond the lane parameters to the polar coordinate system;

(42) Carry out Hough transform to each pixel point on the roadway image, traverse the polar angles of all points, calculate the polar diameters of all points, add 1 in the parameter array corresponding to the point of the same polar diameter and polar angle;

(43) set the straight line length threshold, obtain the polar coordinate parameter of straight line transformation;

(44) Mark the straight line on the lane image according to the parameters. If there are multiple lane lines in the obtained polar coordinate parameters that cause the lane line to be too wide or have multiple boundaries, then adopt the lane line fitting method to fit the wide straight line or multiple boundaries. Perform lane fitting and extract lane lines.

In the step (5), the Kalman threshold of the next image is determined by the ratio of the previous image, and the ratio of the next image is calculated according to the Kalman threshold.

The Kalman prediction area in the step (3) is an area centered on the lane line and having a width of five lane lines.

Beneficial effect

Due to the adoption of the above-mentioned technical solution, the present invention has the following advantages and positive effects compared with the prior art: the present invention adopts an adaptive method to obtain the threshold value of Hough transform, and adds a straight line fitting algorithm to accurately locate the straight line position, which can better identify lane lines with a certain width and curvature. The present invention also adopts a brand-new dynamic threshold binarization method based on prior knowledge, which can adapt to changes in weather conditions and adaptively acquire thresholds for image threshold segmentation. The present invention uses the result of the last positioning to limit the image area of the lane line search of the current lane, so that the range of the sensitive area can be narrowed according to the dynamic information of the movement, the lane line can be reliably located, and the accuracy and real-time performance can be improved.

Description of drawings

Fig. 1 is a flow chart of the present invention;

Fig. 2 is the flowchart of image preprocessing among the present invention;

Fig. 3 is a flow chart of Hough transform in the present invention.

Detailed ways

Below in conjunction with specific embodiment, further illustrate the present invention. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the teachings of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of the present application.

The embodiment of the present invention relates to a lane recognition deviation detection method, as shown in Figure 1, is divided into the following steps, lane image preprocessing, Canny operator edge detection, Kalman filter prediction, Hough transform with straight line fitting , select the threshold of the Hough transform in an adaptive way.

The preprocessing step of the lane image in the present invention, as shown in Figure 2, includes ROI processing, graying processing, smoothing filter processing, contrast enhancement processing, histogram analysis area and dynamic threshold binarization based on prior knowledge to the image deal with.

ROI processing is based on CCD and CMOS imaging principles to divide the picture into regions of interest and regions of non-interest for modeling. First, the camera obtains an image based on CCD or CMOS. After obtaining the image, the image is segmented and processed. According to empirical data, the sky component generally accounts for 5/12 of the image. Therefore, we take the lower 7/12 part of the image as the image processing Base. The ROI processing of the image aims to approximately take a cutoff point Y _m as the cutoff point between the near vision area and the far vision area. Wherein, the near vision area is the region of interest, and the far vision area is the non-interest area. In the near vision area, the lane is approximately seen as a straight line.

Then, ash processing is carried out. The ash processing uses the LAB mode as an intermediary to convert the RGB image acquired by the camera into the LAB mode, and then generates the corresponding equivalent RGB gray scale. That is to say, convert the RGB image to LAB mode, then decolorize in LAB mode, then return to the RGB image and generate an equivalent RGB grayscale, and finally convert to the grayscale space according to the grayscale, and generate the corresponding The grayscale K. The value of the gray scale K is between 0 and 255.

There are many noise points in the image obtained after graying, and the median filter method is used for processing. Replace the value of a point in a digital image or digital sequence with the median value of each point in a neighborhood (3*3) of the point, so that the surrounding pixel values are close to the real value, thereby eliminating isolated noise points. Using the sliding template from left to right and from top to bottom, the pixels in the display screen are sorted according to the size of the pixel value, and a monotonically rising (or falling) data sequence is generated. The smoothing filtering process adopts the median filtering method, which can nonlinearly and smoothly remove the noise points after graying, and at the same time protect the target boundary so that it is not blurred.

Then, the contrast enhancement process is performed. Since the fullness of the grayscale image obtained after the median filter is low, the contrast enhancement method is used to improve the fullness of the image, which helps the system to clearly distinguish the position of the lane line.

Afterwards, use the histogram to analyze the area. The histogram analysis divides the image into m levels, and there are n pixels with gray value i. The probability of occurrence of pixels at each level is reflected and analyzed in the form of a histogram.

In the dynamic threshold binarization process based on prior knowledge, the image is regarded as a combination of two types of regions with different gray levels by using the difference in gray level between the target object in the image and its background, and an appropriate threshold is selected to Identify each point in the image. First, the method based on prior knowledge is used to perform the first binarization segmentation, and then the first segmentation threshold is determined. When the vehicle enters the lane, the area of the unit area is manually selected on the lane line and the lane respectively, and the mean values W and B of the gray levels in the two areas are calculated. Combining each color gray value feature of the reference template (that is, the standard road image with prior knowledge) into a 1×m vector A=[q ₁ , q _W ,...,q _m-1 ,q _m ], such that the vector A Each component in represents the proportion of this color feature component in the reference template. According to the prior knowledge, the mean values W and B of the gray levels in the two areas selected for the first time are binarized and segmented for the first time using the "valley method". The trough between W and B, that is, the lowest point, indicates that within the gray value, the probability of the image point is the smallest, which can be used as the basis for segmentation, and the gray value of this point is recorded as K. It has been verified by experiments that under the conditions of light and weather changes, the grayscale of the grayscale image does not change more than 18 grayscale values, which is used as the benchmark for the next frame of image binarization. Map the first binarized segmented image to the original image in the way of Y=X mapping, take the gray value of each point in the original image corresponding to the white point in the binarized image, and obtain the gray average by means of arithmetic mean Value K _W1 . When the grayscale arithmetic is averaged, the grayscale histogram is reflected as the kth level, then centering on this value, the corresponding items of the k±1, k±2, ..., k±9 groups in the histogram are paired For comparison, obtain the larger of the two comparison results, determine the group number k _m of the group with the larger gray value, obtain a set of data k _m1 , k _m2 , ..., k _m9 , and then determine the gray value of this group of data Degree range, denoted as K _W2 . Taking K _W2 as the benchmark for the second binarization, the points within the range of K _W2 are used as white points for the second binarization, and the points outside K _W2 are used as black points. Repeatedly, the adaptive binarization threshold determination is realized.

It is not difficult to find that a dynamic threshold binarization method based on prior knowledge can adapt to changes in weather conditions and adaptively obtain thresholds for image threshold segmentation.

Canny operator edge detection essentially uses a quasi-Gaussian function for smoothing operations, and then uses a directional first-order differential operator to locate the maximum value of the derivative, and then judges whether the point is an edge point according to the three judgment principles of the Canny operator. Specifically, according to the definition of Canny, the central edge point is the maximum value in the area of the convolution between the operator and the image in the edge gradient direction, so that it can be judged in the gradient direction of each point whether the intensity of this point is its The maximum value of the neighborhood is used to determine whether the point is an edge point. When a pixel satisfies the following three conditions, it is considered to be an edge point of the image: 1) The edge intensity of the point is greater than two pixels along the gradient direction of the point. The edge strength of adjacent pixels; 2) The direction difference between two adjacent points in the gradient direction of this point is less than 45 degrees; 3) The edge strength maximum value in the 3×3 neighborhood centered on this point is less than a certain a threshold. The lane edge image is obtained by Canny operator edge detection.

During the image acquisition process, the camera shake caused by light intensity, obstacle occlusion, roadside trees and uneven road surface will all affect the lane line information in the image. In such a case, the extraction of lane line parameters will produce a relatively large error, and it may also be impossible to track the lane due to the lane turning state. The present invention introduces the lane tracking method based on the Kalman predictor to determine the position of the lane line. After the lane line position is determined at the initial stage of vehicle driving, take the lane line as the center and take the width as five lane lines (about 50 in the image. Pixel area) is used as the Kalman prediction area to track and predict the possible position of the lane, and then use the distance discriminant method to filter out the effective point set, and finally extract the lane parameters on the basis of the optimized point set. Through such a prediction method, the change of the lane can be better obtained, which is more accurate than other lane detection methods.

In the lane line extraction algorithm, Hough transform is one of the most commonly used methods, and its advantages are good anti-noise performance and stable algorithm. Pixels in image space can be represented by straight lines in parameter space through projection. As shown in Figure 3, first determine a polar coordinate system, that is, initialize a two-dimensional array buffer to store the values of the parameter planes ρ, θ, and first set all the data in the array to 0. Then perform Hough transform on each pixel of the road image, traverse the θ angle of all points (the traversed area can be selected according to needs), that is, the polar angle, and calculate all the values of ρ, that is, the polar radius. Add 1 to the parameter array for points corresponding to the same ρ, θ. Finally, find the position of the larger point of the parameter array on the parameter plane. The selection of the larger point can be selected by setting different thresholds according to the transformation needs. This position is the parameter corresponding to the straight line on the road image. In the obtained parameters, there may be multiple lane lines that cause the lane line to be too wide or polygonal. Here, the lane fitting method is used to perform lane fitting on wide straight lines and multiple boundaries. After the Hough transform accumulates all the points, the threshold of the number of accumulated points is not easy to determine. If the threshold is too large, when the lane line has a dashed line, the discontinuous part of the lane line will affect the cumulative value of the straight line, and the lane line may be missed at this time. If the threshold is too small, other lane lines and road boundaries other than the current lane line will also be considered as lane lines, and certain strategies need to be adopted to extract lane lines.

Finally, the threshold of the Hough transform is selected in an adaptive way, that is to say, through the above Kalman filter prediction steps, after the starting point is determined, the dynamic prediction of the lane can be obtained. The selected lane is the center and the width is five In the image sensitive area of the lane width, count the number of pixels identified within the range of the background point and the lane line obtained through the Hough transform, and calculate the ratio T. According to the ratio T, it can be known whether the lane deviates. It is determined by the ratio T of the previous image, and the Kalman threshold ΔH is obtained through the relationship with the Kalman prediction point of the next image, and then the ratio T' of the next image is calculated according to the Kalman threshold ΔH.

It can be seen that the present invention adopts an adaptive method to obtain the threshold value of the Hough transform, and adds a straight line fitting algorithm to accurately locate the position of the straight line, and can better identify lane lines with a certain width and curvature. At the same time, the result of the last positioning is used to limit the image area of the lane line search of the current lane, so that the range of sensitive areas can be narrowed according to the dynamic information of the movement, which can not only reliably locate the lane line, but also improve accuracy and real-time performance.

Claims (5)

1. A lane recognition deviation detection method, is characterized in that, comprises the following steps: (1) Acquire a lane image, and preprocess the lane image; (2) Carry out Canny operator edge detection on the preprocessed lane image to obtain the lane edge image; (3) According to the obtained lane edge image based on the lane tracking method of the Kalman predictor, determine the position of the exit lane line, select the Kalman prediction area, and use the distance discrimination method to screen out the effective point set, and finally based on the optimized point set Extract the lane parameters; (4) According to the obtained lane parameter, utilize the Hough transformation with straight line fitting to extract the lane line; (5) Using the starting point position determined in step (3) and the dynamic prediction of the lane, count the number of background points and lane line points in the Kalman prediction area, and calculate the ratio between the background point and lane line points. 2. lane recognition deviation detection method according to claim 1, is characterized in that, the preprocessing in described step (1) also comprises the following substeps: (11) ROI processing is carried out to the acquired lane image; (12) Ashing processing is carried out to the lane image after ROI processing; (13) Carry out median filter processing to the lane image after graying processing; (14) Contrast enhancement processing is carried out to the lane image after median filtering; (15) The lane image after the contrast enhancement is divided into m levels, and the probability of occurrence of pixels of each level is reflected and analyzed in the form of a histogram, wherein, m>1; (16) Utilize the difference in grayscale between the target object and the background in the lane image, obtain the first binarization segmentation threshold based on prior knowledge, and automatically obtain the next segmentation threshold in an adaptive way, so as to determine the every point of . 3. lane identification deviation detection method according to claim 1, is characterized in that, the Hough transformation in described step (4) also comprises the following steps: (41) Determine the polar coordinate system, and correspond the lane parameters to the polar coordinate system; (42) Carry out Hough transform to each pixel point on the roadway image, traverse the polar angles of all points, calculate the polar diameters of all points, add 1 in the parameter array corresponding to the point of the same polar diameter and polar angle; (43) set the straight line length threshold, obtain the polar coordinate parameter of straight line transformation; (44) Mark the straight line on the lane image according to the parameters. If there are multiple lane lines in the obtained polar coordinate parameters that cause the lane line to be too wide or have multiple boundaries, then adopt the lane line fitting method to fit the wide straight line or multiple boundaries. Perform lane fitting and extract lane lines. 4. lane recognition departure detection method according to claim 1, is characterized in that, in described step (5), by determining the Kalman threshold value of next image by the ratio of last image, and according to this Kalman threshold Threshold finds the ratio of the next image. 5. lane identification deviation detection method according to claim 1, is characterized in that, the Kalman prediction area in the described step (3) is centered on the lane line, and width is the area of five described lane lines .

Images