Lane line extraction and parking space detection method based on laser radar intensity value

Abstract

The invention discloses a lane line extraction and parking space detection method based on a laser radar intensity value. The method comprises the steps of 1, preprocessing radar data; 2, identifying and removing a target of a dynamic object in the point cloud with the intensity value information obtained in the step 1 by using a dynamic detection algorithm, and retaining data of a static object for subsequent processing; 3, identifying the boundary and characteristics of the lane line of the static object, determining a lane line candidate region, accurately extracting the lane line through geometric constraint, obtaining a linear equation of the lane line, and obtaining a parking space region; and 4, detecting and numbering the parking spaces according to the extracted lane lines and parking space areas, and judging the occupation state of each parking space by analyzing the point cloud density and intensity value distribution in the parking space areas. According to the invention, the radar data can stably and accurately identify the lane line boundary and the parking space state under different environmental conditions, so that the environmental perception capability in automatic driving is improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving perception technology, and in particular to a lane line extraction and parking space detection method based on laser radar intensity values. Background Art

[0002] With the rapid development of autonomous driving technology, the demand for high-precision environmental perception and target detection is growing. In urban parking lots and complex traffic scenarios, accurate lane line and parking space detection is of great significance for realizing automatic parking and parking lot navigation. Although traditional vision-based detection methods can provide high accuracy under certain conditions, they have great limitations when dealing with complex environments (such as low light, occlusion, rain and fog, etc.). This is because visual methods are highly dependent on ambient lighting, and recognition failures are prone to occur in the case of lighting changes or occlusions, thus affecting the stability of the autonomous driving system.

[0003] In contrast, radar data, especially radar intensity value information, has stronger environmental adaptability. Radar can work stably in different weather and lighting conditions and provide reflection intensity information for objects in the environment, which makes it more robust in detecting ground features and static objects.

[0004] However, in the prior art, radar parking may be disturbed by dynamic objects, affecting the stability of detection.

[0005] For example, a laser radar parking space recognition method and device with publication number CN118447471A applied to complex scenes will be disturbed by dynamic objects (such as pedestrians and vehicles), affecting the stability of detection.

[0006] In addition, the accuracy of extracting lane line boundaries is also a key problem, which has a direct impact on the detection effect. When LiDAR detects stop lines, it lacks texture and color information, resulting in insufficient recognition ability. Summary of the invention

[0007] In order to overcome the defects of the above existing technologies, the present invention provides a lane line extraction and parking space detection method based on laser radar intensity value, which can improve the accuracy and robustness of lane line and parking space detection in the presence of dynamic interference objects (such as pedestrians and vehicles) and in complex lighting environments. The present invention uses technical means such as dynamic detection and removal, candidate frame extraction, etc. to enable radar data to stably and accurately identify lane line boundaries and parking space status under different environmental conditions, thereby improving the environmental perception ability in autonomous driving.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A lane line extraction and parking space detection method based on laser radar intensity value, comprising the following steps;

[0010] Step 1: Get radar data and preprocess it:

[0011] In a parking environment, by installing a laser radar on a vehicle or robot, the original point cloud data in the environment is collected in real time during its movement or task execution; the original point cloud data is preprocessed to extract intensity value information; it is used to distinguish different types of object surfaces (such as the ground, lane lines, vehicles, etc.), and is divided into dynamic objects and static objects according to the object surface;

[0012] Step 2: Dynamic detection and removal;

[0013] Using dynamic detection algorithm to identify, the detected objects are divided into dynamic objects and static objects according to the intensity value information, the point cloud data of dynamic objects are removed, and the point cloud data of static objects are retained. The point cloud data of static objects contains the point cloud intensity value information of parking lines;

[0014] Step 3: Extract candidate boxes based on intensity values:

[0015] The lane line boundary and features are identified through the change of radar intensity value, and the candidate lane line area is determined by using the local maximum value and gradient information of the intensity value. The lane line is accurately extracted through geometric constraints and the straight line equation of the lane line is obtained, thereby describing the spatial position of the parking line and obtaining the parking space area.

[0016] Step 4: Parking space detection and status judgment

[0017] According to the extracted lane lines and parking space areas, the parking spaces are detected and numbered, and the occupancy status (free or occupied) of each parking space is determined by analyzing the point cloud density and intensity value distribution within the parking space area.

[0018] The step 1 is specifically as follows:

[0019] Get the original point cloud data P in the environment = {(x i ,y i ,z i ,I i )|i=1,2,…,N}, where P is a point cloud collection containing N points, and each point (x i ,y i ,z i ,I i ) represents the spatial coordinates of the point (x i ,y i ,z i ) and intensity value I i ;

[0020] Intensity value Ii Represents the reflection intensity of radar waves, which is used to assist in distinguishing ground and non-ground points. The reflection intensity of radar waves will be used for subsequent separation of ground and non-ground points.

[0021] The CSF bag filter is used to simulate the sagging process of the bag under the action of gravity. The point cloud data is regarded as a surface. In each iteration, the height of the bag The update formula is as follows:

[0022]

[0023] is the height of the i-th point at the t-th iteration, is the height of the i-th point at the t+1th iteration, Δt is the time step, which is an adjustment parameter used to control the update speed of each iteration, g is the acceleration of gravity, which represents the force of the simulated bag under the action of gravity, and k is the elastic coefficient or recovery coefficient, which represents the recovery ability of the bag when it is deformed.

[0024] During the bag deformation process, the height of each point i is also affected by its neighboring points. To ensure the smoothness of the bag deformation, the height of each point is adjusted by the following formula:

[0025]

[0026] is the average height of the neighboring points of the i-th point at the t-th iteration. This value is used to calculate the smooth height of point i during the bag deformation process and is an important basis for updating the height. is the neighborhood of point i, which means the set of points adjacent to point i. The neighborhood is determined by defining a certain distance threshold, that is, points within a certain threshold are regarded as neighborhood points. is the number of neighborhood points, that is, the total number of points contained in the neighborhood of point i. It represents the number of points in the neighborhood of the i-th point that participate in calculating the average height. is the height of the jth neighborhood point at the jth iteration. The smoothness of the bag deformation is ensured by averaging the heights of the neighborhood points. t is the number of iterations, which is used to describe the current state of the bag deformation process.

[0027] After multiple iterations, the bag reaches a stable state and forms a surface that fits the ground. i And bag height The difference δz i To classify ground and non-ground points:

[0028]

[0029] Set a threshold δ thresholdIf δz i ≤δ threshold , then the point is considered a ground point, otherwise it is a non-ground point:

[0030] Ground point ifδz i ≤δ threshold

[0031] Non-ground point ifδz i >δ threshold

[0032] After obtaining the parking lot ground point cloud, RANSAC is used for plane fitting to obtain a more accurate plane point cloud.

[0033] Furthermore, the RANSAC algorithm is used to perform plane fitting on the parking lot ground point cloud to obtain accurate ground point cloud;

[0034] RANSAC fits an initial plane by randomly selecting a small number of points and checks the distances of other points to the plane. If they are within the set range, the points are considered to conform to the plane.

[0035] This process is repeated multiple times to find the best plane model containing the most conforming points (inliers), and these inliers are used to refit a more accurate ground plane.

[0036] After fitting the plane, the parking line point cloud is extracted through the change of point cloud intensity value, and the non-ground point cloud provides vehicle detection information for subsequent parking space occupancy detection;

[0037] Assume that the intensity value of each point in the point cloud data on the same scan line is I i , where i represents the order of the points.

[0038] In the scan line, the rules for determining the stop line are as follows:

[0039] Judgment of strength value mutation:

[0040] For point i, if the intensity values ​​of the adjacent points meet the following conditions:

[0041] I i+1 -I i >T

[0042] It is considered that the intensity value at point i rises sharply, T is the threshold of intensity mutation, and point i can be marked as the starting point cloud of the stop line;

[0043] On the scan line after point i, if point j appears, it satisfies the following conditions:

[0044] I j -I j+1 >T

[0045] It is considered that the intensity value at point j drops sharply, and the point cloud from point i to point j can be marked as the stop line point cloud.

[0046] The threshold T is determined using the statistical characteristics of the ground point cloud data;

[0047] Assume that the intensity value of all ground point clouds in the point cloud data is I 1 ,I 2 ,...,I n , then calculate its mean μ and variance σ 2 as follows:

[0048] Average value μ:

[0049] Variance σ 2 :

[0050] Threshold setting T:

[0051] To eliminate the effect of illumination, the threshold T is set to the mean plus twice the standard deviation:

[0052] T=μ+2σ

[0053] The parking lot ground parking line point cloud information is obtained through this method.

[0054] The step 2 is specifically as follows:

[0055] The M-detetor dynamic detection module is used to eliminate the interference of dynamic targets (such as moving vehicles and pedestrians) on lane line and parking space detection.

[0056] The dynamic detection module processes the input point cloud data, whether it is a single point or a frame of point cloud data. First, the point cloud data is serialized and these serial point clouds are input into the event detection module. At the same time, the point cloud depth image accumulated to the current moment is also input into the module.

[0057] The event detection module uses three methods to identify dynamic points:

[0058] Occlusion background detection: Identify moving targets perpendicular to the laser emission direction by detecting whether the input point at the current moment occludes the background point at the past moment;

[0059] Distance occlusion detection: By detecting whether the point at the past moment is blocked by the current input point, the target movement far away from the radar and parallel to the laser emission direction is identified;

[0060] Proximity occlusion detection: Identify target movement close to the radar and parallel to the laser emission direction by detecting whether the current input point occludes the point in the past.

[0061] The step 3 is specifically as follows:

[0062] Assume that the length and width of the preset parking space are L and W. By screening out candidate boxes that meet these conditions, the length, width and corresponding straight line equation of the candidate boxes are extracted. Each candidate box contains two sets of straight line equations: the equation group L in the length direction i : and the width direction equation system W j :

[0063] The linear translation and marking of the length equation L for each candidate box i and width equation W j Conduct independent processing;

[0064] Select the equation with the smaller intercept b for translation operation.

[0065] During the translation process, the number of points near the translated line is calculated. If the number of points N near the line is greater than the threshold T p , then mark the translated straight line as a length straight line or a width straight line;

[0066] Calculation of the distance between adjacent straight lines For the extracted length straight lines and width straight lines, calculate the distance d between adjacent length straight lines L The distance d between the adjacent width lines W :

[0067]

[0068] According to the preset length L and width W of the parking space, the straight lines that meet the distance requirements are screened out, and the redundant straight lines that are too close are eliminated. The retained straight lines are called effective length straight lines and effective width straight lines.

[0069] Determination of actual length and width of the straight line:

[0070] Count the number of effective length lines and width lines. If the number is larger, they are straight lines in the width direction, and they are recorded as actual width lines; if the number is smaller, they are straight lines in the length direction, and they are recorded as actual length lines.

[0071] Line completion judgment and operation:

[0072] If the number of actual length straight lines is equal to 2, no completion is required; if the number of actual length straight lines is equal to 1, completion is required.

[0073] The completion steps are as follows:

[0074] Assume that the equation of the actual length straight line is L actual :y=m actual x+b actual ;

[0075] L actual : represents the actual length of a straight line (or true straight line), used to describe the position and direction of a straight line;

[0076] m actual : The slope of a line indicates the degree of inclination of the line. The slope is defined as the ratio of the difference between the ordinate and the abscissa between any two points on the line, reflecting the angle between the line and the horizontal axis.

[0077] b actual : The intercept of the straight line, which indicates the intersection of the straight line and the y-axis, that is, the starting position of the straight line on the vertical axis;

[0078] Calculate the distance between the actual length line and the two edges in the candidate box, and select the candidate box line L that is closest to it. near And the candidate box line L with the longest distance far ;

[0079] Calculate the distance between the nearest and farthest candidate box lines: d nf =|b near -b far |;

[0080] If d nf <0, then in b actual Add a fixed value C to generate the completed line: b new =b actual +C;

[0081] If d nf >0, then in b actual Subtract a fixed value C from the top to generate the completed line: b new =b actual -C;

[0082] Calculation and labeling of the corner point coordinates of the parking space. According to the equations of the effective length line and the width line obtained, the coordinates of the four corner points of each parking space rectangle are calculated. For each corner point P of the parking space k (x k ,y k ), obtained by the intersection of the straight line equations, these corner point information is stored in the dictionary.

[0083] The step 4 is specifically as follows:

[0084] By calculating the number of non-ground point clouds above the parking area, it is determined whether the parking space is occupied;

[0085] The specific steps are as follows:

[0086] Parking area definition:

[0087] Set the angle P of each parking space 1 (x1 ,y 1 ),P 2 (x 2 ,y 2 ),P 3 (x 3 ,y 3 ),P 4 (x 4 ,y 4 ) These corner points define the rectangular area R of the parking space;

[0088] Non-ground point cloud extraction:

[0089] Extract and filter the non-ground point cloud data above the parking space, and set the non-ground point cloud set as S = {(x i ,y i ,z i )|i=1,2,…,n}, where z i is the height of the point cloud;

[0090] Determine the number of point clouds above the rectangular area:

[0091] For a point Q(x i ,y i ), calculate the cross product value with each side of the rectangle to determine their positional relationship:

[0092] cross 12 =(x 2 -x 1 )(y i -y 1 )-(y 2 -y 1 )(x i -x 1 )

[0093] cross 23 =(x 3 -x 2 )(y i -y 2 )-(y 3 -y 2 )(x i -x 2 )

[0094] cross 34 =(x 4 -x 3 )(y i -y 3 )-(y 4 -y 3 )(x i -x 3 )

[0095] cross 41 =(x 1 -x 4 )(y i -y 4 )-(y 1 -y 4 )(x i -x 4 )

[0096] If all cross products 12 cross 23 cross 34 cross 41 have the same sign (i.e. both are positive or both are negative), then point Q(x i ,y i ) is within the rectangular region R; otherwise, point Q(x i ,y i )) is located outside the region;

[0097] Count the points that meet the conditions, and count the points that meet the conditions in the rectangular area, recorded as:

[0098] Among them, indicator function 1 in_rectangle (x i ,y i ) if and only if the point Q(x i ,y i ) is within the rectangular area, the value is 1, otherwise the value is 0;

[0099] Occupancy judgment: set a point cloud quantity threshold T occupy , if N>T occupy , the parking space is considered to be occupied; otherwise, it is considered to be free:

[0100] Occupancy status condition: Occupancy status: if N ≥ T occupy

[0101] Idle state condition: Idle state: ifN <T occupy .

[0102] The method is applied to parking space occupancy detection in complex environments.

[0103] Beneficial effects of the present invention:

[0104] 1. Enhanced environmental adaptability: Lane line and parking space detection is performed through the intensity value information of the LiDAR, which effectively improves the robustness of the system under various lighting conditions. The intensity value is not affected by light changes and can work stably in low light, strong light, or even no light environments, overcoming the dependence of traditional visual methods on light.

[0105] 2. Effective filtering of dynamic targets: Using dynamic detection and removal technology, dynamic targets in the parking lot (such as pedestrians and moving vehicles) can be effectively identified and removed, thereby retaining interference-free static point cloud data for subsequent lane line and parking space extraction, greatly improving the accuracy of detection.

[0106] 3. Accurate lane and stop line extraction: Through the candidate area extraction method based on intensity values, combined with the local maximum of the intensity value and gradient information, the present invention can accurately identify the boundaries of lane lines and stop lines, and further optimize them through geometric constraints, thereby achieving high-precision stop line extraction.

[0107] 4. Parking space status judgment: By using the point cloud distribution of intensity values, the present invention can accurately judge the occupancy status of each parking space and timely detect the vacant or occupied status of the parking space. This still has strong applicability and robustness in parking lot environments with many dynamic targets.

[0108] 5. Wide applicability and high robustness: This method is applicable to a variety of complex parking environments, including open-air parking lots, underground parking lots, etc. It is not affected by light and weather conditions and has a wide range of adaptability. The system can still provide stable detection results in a variety of complex environments, reflecting strong practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 It is the original point cloud map provided by the present invention.

[0110] Figure 2 It is the CSF ground filtering and plane fitting diagram provided by the present invention.

[0111] Figure 3 The present invention provides a method for filtering and extracting a lane line point cloud image based on intensity values.

[0112] Figure 4 It is a flow chart of extracting lane lines based on geometric constraints provided by the present invention.

[0113] Figure 5 This is the result graph of lane line extraction based on geometric constraints provided by the present invention.

[0114] Figure 6 This is a result diagram of the parking space occupancy detection provided by the present invention. DETAILED DESCRIPTION

[0115] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0116] A lane line extraction and parking space detection method based on laser radar intensity value, comprising the following steps;

[0117] Step 1: Radar data preprocessing: Obtain point cloud data of the parking environment, where each point contains location, height and intensity value information; Use the CSF bag filter method to deform the point cloud data under the action of gravity, remove noise points, and separate ground and non-ground points; Use the RANSAC algorithm to perform plane fitting on the filtered ground point cloud data and extract the ground plane model; Extract the parking line point cloud information based on the sudden change of the point cloud intensity value.

[0118] Furthermore, the point cloud intensity value information is used to judge the parking line, and the parking line point cloud is identified by the surge and sharp drop in the intensity values ​​of adjacent points in the scanning line; to reduce the impact of illumination changes, the threshold T = μ + 2σ is set, where μ is the average value of the ground point cloud intensity value and σ is its standard deviation.

[0119] The step 1 is specifically as follows:

[0120] Get the point cloud data P in the environment = {(x i ,y i ,z i ,I i )|i=1,2,…,N}, where each point (x i ,y i ,z i ,I i ) contains position, altitude and intensity value information;

[0121] Intensity value I i Represents the reflection intensity of radar waves, which is used to assist in distinguishing ground and non-ground points; these data will be used for subsequent separation of ground and non-ground points;

[0122] The CSF bag filter is used to simulate the sagging process of the bag under the action of gravity. The point cloud data is regarded as a surface. In each iteration, the height of the bag The update formula is as follows:

[0123]

[0124] is the height of the i-th point at the t-th iteration, is the height of the i-th point at the t+1th iteration, Δt is the time step, which is an adjustment parameter used to control the update speed of each iteration, g is the acceleration of gravity, which represents the force of the simulated bag under the action of gravity, and k is the elastic coefficient or recovery coefficient, which represents the recovery ability of the bag when it is deformed.

[0125] During the bag deformation process, the height of each point i is also affected by its neighboring points. To ensure the smoothness of the bag deformation, the height of each point is adjusted by the following formula:

[0126]

[0127] The average height of the neighboring points of the i-th point at the t-th iteration. This value is used to calculate the smooth height of point i during the bag deformation process and is an important basis for updating the height. is the neighborhood of point i, which means the set of points adjacent to point i. The neighborhood is determined by defining a certain distance threshold, that is, points within a certain threshold are regarded as neighborhood points. is the number of neighborhood points, that is, the total number of points contained in the neighborhood of point i. It represents the number of points in the neighborhood of the i-th point that participate in calculating the average height. is the height of the jth neighborhood point at the jth iteration. The smoothness of the bag deformation is ensured by averaging the heights of the neighborhood points. t is the number of iterations, which is used to describe the current state of the bag deformation process.

[0128] After multiple iterations, the bag reaches a stable state and forms a surface that fits the ground. i And bag height The difference δz i To classify ground and non-ground points:

[0129]

[0130] Set a threshold δ threshold If δz i ≤δ threshold , then the point is considered a ground point, otherwise it is a non-ground point:

[0131] Ground point ifδz i ≤δ threshold

[0132] Non-ground point ifδz i >δ threshold

[0133] After obtaining the parking lot ground point cloud, RANSAC is used for plane fitting to obtain a more accurate plane point cloud.

[0134] Furthermore, in order to obtain more accurate ground point cloud, the RANSAC algorithm is used to perform plane fitting on the parking lot ground point cloud;

[0135] RANSAC fits an initial plane by randomly selecting a small number of points and checks the distances of other points to the plane. If they are within the set range, the points are considered to conform to the plane.

[0136] This process is repeated multiple times to find the best plane model containing the most conforming points (inliers), and these inliers are used to refit a more accurate ground plane. In this way, noise points can be effectively removed and an accurate ground plane point cloud can be obtained, thereby improving the accuracy of subsequent parking space detection.

[0137] After fitting the plane, the parking line point cloud is extracted through the change of point cloud intensity value, and the non-ground point cloud provides vehicle detection information for subsequent parking space occupancy detection;

[0138] Assume that the intensity value of each point in the point cloud data on the same scan line is I i , where i represents the order of the points.

[0139] In the scan line, the rules for determining the stop line are as follows:

[0140] Judgment of strength value mutation:

[0141] For point i, if the intensity values ​​of the adjacent points meet the following conditions:

[0142] I i+1 -I i >T

[0143] It is considered that the intensity value at point i rises sharply, T is the threshold of intensity mutation, and point i can be marked as the starting point cloud of the stop line;

[0144] On the scan line after point i, if point j appears, it satisfies the following conditions:

[0145] I j -I j+1 >T

[0146] It is considered that the intensity value at point j drops sharply, and the point cloud from point i to point j can be marked as the stop line point cloud.

[0147] Since different lighting conditions will cause the radar intensity value I of the same object to change, in order to reduce the impact of lighting changes, the statistical characteristics of the ground point cloud data are used to determine the threshold T. This allows the threshold to dynamically adapt to the lighting conditions of the environment and is more robust.

[0148] Assume that the intensity value of all ground point clouds in the point cloud data is I 1 ,I 2 ,...,I n , then calculate its mean μ and variance σ 2 as follows:

[0149] Average value μ:

[0150] Variance σ 2 :

[0151] Threshold setting T:

[0152] To eliminate the effect of illumination, the threshold T is set to the mean plus twice the standard deviation:

[0153] T=μ+2σ

[0154] The parking lot ground parking line point cloud information is obtained through this method.

[0155] Step 2: Dynamic detection and removal;

[0156] By detecting whether the current point occludes the background point at the past moment, whether the point at the past moment occludes the current point, and whether the current point occludes the point at the past moment, the dynamic point cloud perpendicular and parallel to the laser emission direction is identified and removed.

[0157] Through dynamic point detection, the interference of moving objects such as vehicles and pedestrians is eliminated to obtain a static scene point cloud without interference from dynamic objects.

[0158] This step eliminates the interference of dynamic targets (such as moving vehicles and pedestrians) on lane line and parking space detection by using the M-detetor dynamic detection method.

[0159] Specifically, a dynamic target recognition method based on event detection is used to remove dynamic object interference in the point cloud to retain the static parking environment data for subsequent processing.

[0160] The dynamic detection module processes the input point cloud data, whether it is a single point or a frame of point cloud data. First, the point cloud data is serialized and these serial point clouds are input into the event detection module. At the same time, the point cloud depth image accumulated to the current moment is also input into the module.

[0161] The event detection module uses three methods to identify dynamic points:

[0162] Occlusion background detection: Identify moving targets perpendicular to the laser emission direction by detecting whether the input point at the current moment occludes the background point at the past moment;

[0163] Distance occlusion detection: By detecting whether the point at the past moment is blocked by the current input point, the target movement far away from the radar and parallel to the laser emission direction is identified;

[0164] Proximity occlusion detection: Identify target movement close to the radar and parallel to the laser emission direction by detecting whether the current input point occludes the point in the past.

[0165] This dynamic detection mechanism based on occlusion relationships achieves low-latency and fast dynamic object recognition, which can effectively eliminate dynamic interference in parking lot scenes and obtain static scene data without dynamic object interference. This is especially important for complex parking lot environments with many vehicles and pedestrians, making subsequent lane line and parking space detection more accurate and reliable.

[0166] Step 3: Extract candidate frames based on intensity values: Filter candidate frames according to the length and width information of the preset parking space; extract the length and width equations of the candidate frames; perform independent operations on each candidate frame, select the equation with a smaller intercept to translate and calculate the number of nearby points. If the number of points is greater than the threshold, mark it as a length or width line; calculate the distance between adjacent length and width lines, remove lines that are too close, and retain valid length and width lines; determine the number of actual length lines and width lines. If the actual length lines are less than 2, fill them in according to the candidate frame boundary.

[0167] Assume that the length and width of the preset parking space are L and W. By screening out candidate boxes that meet these conditions, the length, width and corresponding straight line equation of the candidate boxes are extracted. Each candidate box contains two sets of straight line equations: the equation group L in the length direction i : and the width direction equation system W j :

[0168] The linear translation and marking of the length equation L for each candidate box i and width equation W j Conduct independent processing;

[0169] Select the equation with the smaller intercept b for translation operation.

[0170] During the translation process, the number of points near the translated line is calculated. If the number of points N near the line is greater than the threshold T p , then mark the translated straight line as a length straight line or a width straight line;

[0171] Calculation of the distance between adjacent straight lines For the extracted length straight lines and width straight lines, calculate the distance d between adjacent length straight lines L The distance d between the adjacent width lines W :

[0172]

[0173] According to the preset length L and width W of the parking space, the straight lines that meet the distance requirements are screened out, and the redundant straight lines that are too close are eliminated. The retained straight lines are called effective length straight lines and effective width straight lines.

[0174] Determination of actual length and width of the straight line:

[0175] Count the number of effective length lines and width lines. If the number is larger, they are straight lines in the width direction, and they are recorded as actual width lines; if the number is smaller, they are straight lines in the length direction, and they are recorded as actual length lines.

[0176] Line completion judgment and operation:

[0177] If the number of actual length straight lines is equal to 2, no completion is required; if the number of actual length straight lines is equal to 1, completion is required.

[0178] The completion steps are as follows:

[0179] Assume that the equation of the actual length straight line is L actual :y=m actual x+b actual

[0180] L actual : represents the actual length of a straight line (or true straight line), used to describe the position and direction of a straight line;

[0181] m actual : The slope of a line indicates the degree of inclination of the line. The slope is defined as the ratio of the difference between the ordinate and the abscissa between any two points on the line, reflecting the angle between the line and the horizontal axis.

[0182] b actual : The intercept of the straight line, which indicates the intersection of the straight line and the y-axis, that is, the starting position of the straight line on the vertical axis;

[0183] Calculate the distance between the actual length line and the two edges in the candidate box, and select the candidate box line L that is closest to it. near And the candidate box line L with the longest distance far ;

[0184] Calculate the distance between the nearest and farthest candidate box lines: d nf =|b near -b far |

[0185] If d nf <0, then in b actual Add a fixed value C to generate the completed line: b new =b actual +C

[0186] If d nf >0, then in b actual Subtract a fixed value C from the top to generate the completed line: b new =b actual -C

[0187] Calculation and labeling of the corner point coordinates of the parking space. According to the equations of the effective length line and the width line obtained, the coordinates of the four corner points of each parking space rectangle are calculated. For each corner point P of the parking space k (x k ,y k ), obtained by the intersection of the straight line equations, these corner point information is stored in the dictionary.

[0188] The step 3 can effectively obtain the rectangular structure of each parking space and its corner point coordinates, providing support for subsequent parking space management.

[0189] Step 4: Parking space detection and status judgment Define the rectangular parking space area according to the parking space corner point information; extract and filter the non-ground point cloud above the area; use the cross product method to determine whether the point cloud is within the parking space area; count the number of non-ground point clouds N in the area, if N>T occupy , then the parking space is considered to be occupied, otherwise it is vacant.

[0190] The specific steps are as follows:

[0191] Parking area definition:

[0192] Set the angle P of each parking space 1 (x 1 ,y 1 ),P 2 (x 2 ,y 2 ),P 3 (x 3 ,y 3 ),P 4 (x 4 ,y 4 ) These corner points define the rectangular area R of the parking space;

[0193] Non-ground point cloud extraction:

[0194] Extract and filter the non-ground point cloud data above the parking space, and set the non-ground point cloud set as S = {(x i ,y i ,z i )|i=1,2,…,n}, where z i is the height of the point cloud;

[0195] Determine the number of point clouds above the rectangular area:

[0196] For a point Q(x i ,y i ), calculate the cross product value with each side of the rectangle to determine their positional relationship:

[0197] cross 12 =(x 2 -x1 )(y i -y 1 )-(y 2 -y 1 )(x i -x 1 )

[0198] cross 23 =(x 3 -x 2 )(y i -y 2 )-(y 3 -y 2 )(x i -x 2 )

[0199] cross 34 =(x 4 -x 3 )(y i -y 3 )-(y 4 -y 3 )(x i -x 3 )

[0200] cross 41 =(x 1 -x 4 )(y i -y 4 )-(y 1 -y 4 )(x i -x 4 )

[0201] If all cross products 12 cross 23 cross 34 cross 41 have the same sign (i.e. both are positive or both are negative), then point Q(x i ,y i ) is within the rectangular region R; otherwise, point Q(x i ,y i )) is located outside the region;

[0202] Count the points that meet the conditions, and count the points that meet the conditions in the rectangular area, recorded as:

[0203] Among them, indicator function 1 in_rectangle (x i ,y i ) if and only if the point Q(xi ,y i ) is within the rectangular area, the value is 1, otherwise the value is 0;

[0204] Occupancy judgment: set a point cloud quantity threshold T occupy , if N>T occupy , the parking space is considered to be occupied; otherwise, it is considered to be free:

[0205] Occupancy status condition: Occupancy status: if N ≥ T occupy

[0206] Idle state condition: Idle state: ifN <T occupy

[0207] The step 4 can effectively determine whether the parking space is occupied according to the distribution of the non-ground point cloud in the parking space area. The method has high robustness and accuracy and is suitable for parking space occupancy detection in complex environments.

[0208] Figure 1 The original point cloud image of the parking environment collected by radar is shown. The point cloud data contains the three-dimensional spatial information of the parking lot, and provides the location information, height information and radar reflection intensity value information of each point. By processing these data, the present invention can realize functions such as ground separation, lane line extraction, dynamic object detection and parking space occupancy judgment.

[0209] Figure 2 The results of ground separation using the CSF ground filtering algorithm are shown. The CSF algorithm filters the point cloud data to remove non-ground points by simulating the deformation of the bag under gravity. After CSF processing, the remaining point cloud is further plane-fitted using the RANSAC algorithm to extract the plane terrain of the parking lot. This plane fitting step improves the smoothness of the ground data, laying the foundation for the subsequent accurate detection of parking spaces and lane lines.

[0210] exist Figure 3 The results of lane line point cloud extraction based on intensity value are shown in . The present invention uses the intensity value information of radar point cloud to detect the mutation area of ​​intensity value to identify stop line and lane line. When the intensity value of adjacent points in the point cloud changes significantly (such as surge or drop), it is considered that a possible lane line or stop line is detected. This method has strong adaptability under different lighting conditions, and through the adjustment of intensity threshold, it can achieve robust extraction in different environments.

[0211] Figure 4The flowchart of lane line extraction based on geometric constraints is shown. First, candidate boxes are selected based on the preset lane line length and width, and their edge line equations are extracted. For each candidate box, its length and width equations are translated respectively, and the number of points near the translated straight line is counted to determine the exact position of the length and width straight line. The distance calculation of adjacent length and width straight lines is used to select valid straight lines with appropriate distances. Through geometric constraints, accurate lane line recognition is completed.

[0212] Figure 5 The actual effect of lane line extraction based on geometric constraints is demonstrated. The lane line straight lines obtained based on geometric constraints have high accuracy and consistency, and have good adaptability to lane lines in parking spaces. The lane lines extracted by this method effectively eliminate the interference of noise and dynamic objects, thereby ensuring the stability of subsequent parking space detection.

[0213] Figure 6 The results of the parking space occupancy detection method provided by the present invention are shown. After the rectangular area of ​​the parking space is identified, the occupancy of each parking space is determined by the number of non-ground point clouds above the detection area. By calculating the point cloud density and the preset occupancy threshold T occupy , determine whether the parking space is occupied. This step enables the present invention to accurately determine the idle or occupied state of the parking space, meeting the requirements of automatic parking and intelligent navigation.

Claims

1. A lane line extraction and parking space detection method based on laser radar intensity value, characterized in that: The steps include: Step 1: Get radar data and preprocess it: In a parking lot environment, by installing a laser radar on a vehicle or robot, the original point cloud data in the environment is collected in real time during the movement or execution of the task; the original point cloud data is preprocessed to extract the intensity value information; Used to distinguish different types of object surfaces, divided into dynamic objects and static objects according to the object surface; Step 2: Dynamic detection and removal; Using dynamic detection algorithm to identify, the detected objects are divided into dynamic objects and static objects according to the intensity value information, the point cloud data of dynamic objects are removed, and the point cloud data of static objects are retained. The point cloud data of static objects contains the point cloud intensity value information of parking lines; Step 3: Extract candidate boxes based on intensity values: The lane line boundary and features are identified through the change of radar intensity value, and the candidate lane line area is determined by using the local maximum value and gradient information of the intensity value. The lane line is accurately extracted through geometric constraints and the straight line equation of the lane line is obtained, thereby describing the spatial position of the parking line and obtaining the parking space area. Step 4: Parking space detection and status judgment According to the extracted lane lines and parking space areas, the parking spaces are detected and numbered, and the occupancy status of each parking space is determined by analyzing the point cloud density and intensity value distribution within the parking space area.

2. The lane line extraction and parking space detection method based on laser radar intensity value according to claim 1 is characterized in that: The step 1 is specifically as follows: Get the original point cloud data P in the environment = {(x i ,y i ,z i ,I i )|i=1,2,…,N}, where each point (x i ,y i ,z i ,I i ) contains position, altitude and intensity value information; Intensity value I i Represents the reflection intensity of radar waves, which is used to assist in distinguishing ground and non-ground points; The reflection intensity of the radar wave will be used for the subsequent separation of ground and non-ground points; The CSF bag filter is used to simulate the sagging process of the bag under the action of gravity. The point cloud data is regarded as a surface. In each iteration, the height of the bag The update formula is as follows: is the height of the i-th point at the t-th iteration, is the height of the i-th point at the t+1th iteration, Δt is the time step, which is a regulation parameter used to control the update speed of each iteration, g is the acceleration of gravity, which represents the force of the simulated bag under the action of gravity, and k is the elastic coefficient or recovery coefficient, which represents the recovery ability of the bag when it is deformed; During the bag deformation process, the height of each point is adjusted by the following formula: is the average height of the neighboring points of the i-th point at the t-th iteration, which is used to calculate the smooth height of point i during the bag deformation process. is the neighborhood of point i, which means the set of points adjacent to point i. The neighborhood is determined by defining a certain distance threshold, that is, points within a certain threshold are regarded as neighborhood points. is the number of neighborhood points, that is, the total number of points contained in the neighborhood of point i. It represents the number of points in the neighborhood of the i-th point that participate in calculating the average height. is the height of the jth neighboring point at the jth iteration. The smoothness of the bag deformation is ensured by averaging the heights of the neighboring points. t is the number of iterations, which is used to describe the current state of the bag deformation process. After multiple iterations, the bag reaches a stable state and forms a surface that fits the ground. i And bag height The difference δz i To classify ground and non-ground points: Set a threshold δ threshold ; If δz i ≤δ threshold , then the point is considered a ground point, otherwise it is a non-ground point: Ground point ifδz i ≤δ threshold Non-ground point ifδz i >δ threshold After obtaining the parking lot ground point cloud, RANSAC is used for plane fitting.

3. The lane line extraction and parking space detection method based on laser radar intensity value according to claim 2 is characterized in that: Use the RANSAC algorithm to perform plane fitting on the parking lot ground point cloud to obtain accurate ground point cloud; RANSAC fits an initial plane by randomly selecting a small number of points and checks the distances of other points to the plane. If they are within the set range, the points are considered to conform to the plane. This process is repeated multiple times to find the best plane model containing the most conforming points (inliers), and these inliers are used to refit a more accurate ground plane; After fitting the plane, the parking line point cloud is extracted through the change of point cloud intensity value, and the non-ground point cloud provides vehicle detection information for subsequent parking space occupancy detection; Assume that the intensity value of each point in the point cloud data on the same scan line is I i , where i represents the order of the points.

4. The lane line extraction and parking space detection method based on laser radar intensity value according to claim 3 is characterized in that: In the scan line, the rules for determining the stop line are as follows: Judgment of strength value mutation: For point i, if the intensity values ​​of the adjacent points meet the following conditions: I i+1 -I i >T It is considered that the intensity value at point i rises sharply, T is the threshold of intensity mutation, and point i can be marked as the starting point cloud of the stop line; On the scan line after point i, if point j appears, it satisfies the following conditions: I j -I j+1 >T It is considered that the intensity value at point j drops sharply, and the point cloud from point i to point j can be marked as the stop line point cloud; The threshold T is determined using the statistical characteristics of the ground point cloud data; Assume that the intensity values ​​of all ground point clouds in the point cloud data are I1, I2, ..., I n , then calculate its mean μ and variance σ 2 as follows: Average value μ: Variance σ 2 : Threshold setting T: The threshold T is set to the mean plus two times the standard deviation: T=μ+2σ The parking lot ground parking line point cloud information is obtained through this method.

5. The lane line extraction and parking space detection method based on laser radar intensity value according to claim 1 is characterized in that: The step 2 is specifically as follows: Use the M-detetor dynamic detection module to eliminate the interference of dynamic targets on lane line and parking space detection; The dynamic detection module processes the input point cloud data, whether it is a single point or a frame of point cloud data. First, the point cloud data is serialized and these sequenced point clouds are input to the event detection module. At the same time, the point cloud depth image accumulated to the current moment is also input to the module; The event detection module uses three methods to identify dynamic points: Occlusion background detection: Identify moving targets perpendicular to the laser emission direction by detecting whether the input point at the current moment occludes the background point at the past moment; Distance occlusion detection: By detecting whether the point at the past moment is blocked by the current input point, the target movement far away from the radar and parallel to the laser emission direction is identified; Proximity occlusion detection: Identify target movement close to the radar and parallel to the laser emission direction by detecting whether the current input point occludes the point in the past.

6. The lane line extraction and parking space detection method based on laser radar intensity value according to claim 1 is characterized in that: The step 3 is specifically as follows: Assume that the length and width of the preset parking space are L and W. By screening out candidate boxes that meet these conditions, the length, width and corresponding straight line equations of the candidate boxes are extracted. Each candidate box contains two sets of straight line equations: the equation set in the length direction and the width direction equations The linear translation and marking of the length equation L for each candidate box i and width equation W j Conduct independent processing; Select the equation with the smaller intercept b for translation operation; During the translation process, the number of points near the translated line is calculated. If the number of points N near the line is greater than the threshold T p , then mark the translated straight line as a length straight line or a width straight line; Calculation of the distance between adjacent straight lines For the extracted length straight lines and width straight lines, calculate the distance d between adjacent length straight lines L The distance d between the adjacent width lines W : According to the preset length L and width W of the parking space, the straight lines that meet the distance requirements are screened out, and the redundant straight lines that are too close are eliminated. The remaining straight lines are called effective length straight lines and effective width straight lines; Determination of actual length and width of the straight line: Count the number of effective length lines and width lines. If the number is larger, they are straight lines in the width direction, and they are recorded as actual width lines; if the number is smaller, they are straight lines in the length direction, and they are recorded as actual length lines. Line completion judgment and operation: If the number of actual length straight lines is equal to 2, no completion is required; if the number of actual length straight lines is equal to 1, completion is required.

7. The lane line extraction and parking space detection method based on laser radar intensity value according to claim 6 is characterized in that: The completion steps are as follows: Assume that the equation of the actual length straight line is L actual :y=m actual x+b actual ; L actual : Indicates the actual length of a straight line, used to describe the position and direction of the straight line; m actual : The slope of a line indicates the degree of inclination of the line. The slope is defined as the ratio of the difference between the ordinate and the abscissa between any two points on the line, reflecting the angle between the line and the horizontal axis. b actual : The intercept of the straight line, which indicates the intersection of the straight line and the y-axis, that is, the starting position of the straight line on the vertical axis; Calculate the distance between the actual length line and the two edges in the candidate box, and select the candidate box line L that is closest to it. near And the candidate box line L with the longest distance far ; Calculate the distance between the nearest and farthest candidate box lines: d nf =|b near -b far |; If d nf <0, then in b actual Add a fixed value C to generate the completed line: b new =b actual +C; If d nf >0, then in b actual Subtract a fixed value C from the top to generate the completed line: b new =b actual -C; Calculation and labeling of the corner point coordinates of the parking space. According to the equations of the effective length line and the width line obtained, the coordinates of the four corner points of each parking space rectangle are calculated. For each corner point P of the parking space k (x k ,y k ), obtained by the intersection of the straight line equations.

8. The method for lane line extraction and parking space detection based on laser radar intensity value according to claim 7 is characterized in that: The step 4 is specifically as follows: By calculating the number of non-ground point clouds above the parking area, it is determined whether the parking space is occupied; The specific steps are as follows: Parking area definition: Set the corners of each parking space P1(x1,y1), P2(x2,y2), P3(x3,y3), P4(x4,y4). These corner points define the rectangular area R of the parking space; Non-ground point cloud extraction: Extract and filter the non-ground point cloud data above the parking space, and set the non-ground point cloud set as S = {(x i ,y i ,z i )|i=1,2,…,n}, where z i is the height of the point cloud; Determine the number of point clouds above the rectangular area: For a point Q(x i ,y i ), calculate the cross product value with each side of the rectangle to determine their positional relationship: cross 12 =(x2-x1)(y i -y1)-(y2-y1)(x i -x1) <h2 style=";text-align:left;direction:ltr">cross<h2 style=";text-align:left;direction:ltr"> 23 <h2 style=";text-align:left;direction:ltr"> (x3-x2)(y)<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> -y2)-(y3-y2)(x<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> -x2) cross 34 =(x4-x3)(y i -y3)-(y4-y3)(x i -x3) cross 41 =(x1-x4)(y) i -y4)-(y1-y4)(x i -x4) If all cross products 12 cross 23 cross 34 cross 41 have the same sign, then the point Q(x i ,y i ) is within the rectangular region R; otherwise, point Q(x i ,y i )) is located outside the region; Count the points that meet the conditions, and count the points that meet the conditions in the rectangular area, recorded as: Among them, indicator function 1 in_rectangle (x i ,y i ) if and only if the point Q(x i ,y i ) is within the rectangular area, the value is 1, otherwise the value is 0; Occupancy judgment: set a point cloud quantity threshold T occupy , if N>T occupy , the parking space is considered to be occupied; otherwise, it is considered to be free: Occupancy status condition: Occupancy status: if N ≥ T occupy ; Idle state condition: Idle state: ifN <T occupy .

9. Application of the lane line extraction and parking space detection method based on laser radar intensity value according to any one of claims 1 to 8, characterized in that: The method is applied to parking space occupancy detection in complex environments.

Images