A millimeter wave radar lane detection method based on optical flow method

Abstract

The present application discloses a millimeter-wave radar lane detection method based on the optical flow method, which relates to the field of traffic technology. The method uses the point cloud data provided by the millimeter-wave radar as the data basis. Considering the data sparsity and the fact that the millimeter-wave radar is greatly affected by noise, the traditional optical flow method is modified to construct an error function based on the change of cluster point positions to achieve motion estimation. Then, based on the extracted optical flow vector, the point cloud data is converted into a binary bitmap that shows the traffic flow. Finally, a sliding window is used for local fitting and splicing to obtain the lane detection result. Since the high-precision point cloud data of the millimeter-wave radar is more flexible and can provide accurate information about the lane, this method can show excellent flexibility in complex weather conditions and changing light environments. And the sliding window fitting method is also the reason why this method performs well in processing nonlinear lanes, overcoming the limitations of traditional methods.

Description

Technical Field

[0001] The present application relates to the field of traffic technology, and in particular to a millimeter wave radar lane detection method based on optical flow method. Background Art

[0002] The rapid development of intelligent transportation technology has created unprecedented prospects for traffic management and smart city planning. Since better traffic and urban planning outcomes depend on accurate understanding of road conditions, road lane detection remains an important task in this field. Traditional lane detection methods usually rely on camera images or lidar data, but these data are greatly affected by factors such as lighting and weather conditions, resulting in low lane detection accuracy. Summary of the invention

[0003] In view of the above problems and technical requirements, this application proposes a millimeter wave radar lane detection method based on optical flow method. The technical solution of this application is as follows:

[0004] A millimeter wave radar lane detection method based on an optical flow method, the millimeter wave radar lane detection method comprising:

[0005] The t-th frame point cloud data of the traffic scene is obtained by using the road-side millimeter-wave radar, and each cluster point in the first frame point cloud data is initialized as a key cluster point, and the initial value of the integer parameter t is 1;

[0006] Starting from any t≥2, the optical flow method is used to determine the cluster points in the t-th frame point cloud data that have a matching relationship with the key cluster points in the t-1th frame point cloud data based on the change in cluster point position, and the angle between the optical flow vector formed by each pair of matching key cluster points and the horizontal line is calculated;

[0007] The t-th frame point cloud data is rasterized, and the grids whose number of target key cluster points reaches the number threshold are set as valid grids, and the grids whose number of target key cluster points does not reach the number threshold are set as invalid grids, so as to obtain a binary bitmap of the t-th frame point cloud data; wherein the target key cluster point is a key cluster point whose angle between the optical flow vector and the horizontal line formed in the t-th frame point cloud data is within a predetermined angle range;

[0008] The sliding window is used to locally fit and splice the valid grids in the binary bitmap of the t-th frame point cloud data to obtain the lane detection result.

[0009] A further technical solution is to use the optical flow method to determine the cluster points in the t-th frame point cloud data that have a matching relationship with the key cluster points in the t-1-th frame point cloud data based on the change in cluster point positions, including:

[0010] The error function E is constructed based on the change of cluster point position:

[0011]

[0012] Among them, (x t-1 ,y t-1 ) is the position of a key cluster point in the t-1th frame point cloud data, (dx, dy) is the key cluster point (x t-1 ,y t-1 ) displacement; Ω t-1 is the set of all key cluster points in the t-1th frame point cloud data, C t-1 is the cluster point set composed of all cluster points in the t-1th frame point cloud data, C t is the cluster point set composed of all cluster points in the t-th frame point cloud data; G(C t-1 ,(x t-1 ,y t-1 )) is based on the cluster point set C t-1 The calculated position (x t-1 ,y t-1 ), G(C t ,(x t-1 +dx,y t-1 +dy)) is based on the cluster point set C t The calculated position (x t-1 +dx,y t-1 +dy) at the fitting pixel value;

[0013] Based on the constructed error function E, the optical flow method is used to perform optical flow calculation to determine the optical flow vector composed of key cluster points with matching relationship.

[0014] A further technical solution is to calculate the fitting pixel value G(C, (α, β)) at the position (α, β) according to the distribution of cluster point positions in any cluster point set C, including:

[0015] According to the distribution of cluster point positions in the cluster point set C, the candidate pixel values ​​at the position (α, β) and its multiple neighboring positions are calculated respectively; for any position, when there is a cluster point in the cluster point set C and the Euclidean distance between the position is less than the distance threshold, the candidate pixel value at the position is determined to be 255, otherwise the candidate pixel value at the position is determined to be 0;

[0016] The candidate pixel values ​​at the position (α, β) and its multiple neighboring positions are weighted to obtain the fitting pixel value G(C, (α, β)) at the position (α, β).

[0017] A further technical solution is to weight the candidate pixel values ​​at the position (α, β) and its multiple neighboring positions to obtain the fitting pixel value G(C, (α, β)) at the position (α, β):

[0018]

[0019] Among them, spv(C, (α, β)) is the candidate pixel value at the position (α, β) calculated according to the distribution of cluster point positions in the cluster point set C. It is the position calculated based on the distribution of cluster points in the cluster point set C. The candidate pixel value at It is the position calculated based on the distribution of cluster points in the cluster point set C. The candidate pixel value at Based on the cluster point set C The location of the cluster point distribution in the calculation Candidate pixel value at ; means rounding α upwards. represents rounding up of β; ω1, ω2, ω3 and ω4 are weights respectively.

[0020] Its further technical solution is:

[0021] Among them, {α} represents taking the decimal part of α, and {β} represents taking the decimal part of β.

[0022] A further technical solution is to use a sliding window to locally fit and splice the valid grids in the binary bitmap of the t-th frame point cloud data to obtain a lane detection result including:

[0023] Determine the number n of sliding windows and the window size of each sliding window;

[0024] Divide the binary bitmap of the point cloud data of the tth frame into n regions along the row direction, and take the grid column with the largest number of valid grids in each region as the initial detection column of the current region;

[0025] Initialize a sliding window at the initial detection column of each region according to the window size, the window center of the sliding window is located at the initial detection column, and the lower edge of the sliding window coincides with the lower edge of the binary bitmap of the t-th frame point cloud data;

[0026] For each sliding window, extract the valid grids in the sliding window. When the number of valid grids contained in the sliding window reaches the minimum number of grids, update the position of the sliding window according to the valid grids contained in the sliding window, and execute the step of extracting the valid grids in the sliding window again. All valid grids extracted from the same sliding window during the position update process are spliced ​​and merged to obtain the detection result of a lane.

[0027] A further technical solution is that updating the position of the sliding window according to the valid grid contained in the sliding window includes:

[0028] The average value of the horizontal coordinates of all valid grids contained in the sliding window in the grid coordinate system is calculated as the horizontal coordinate of the window center of the updated sliding window, and the vertical coordinate of the window center of the current sliding window in the grid coordinate system is increased by the height of the sliding window as the vertical coordinate of the window center of the updated sliding window; wherein, the grid coordinate system takes the lower left corner of the binary bitmap as the origin, the rightward direction along the row direction is the positive direction of the horizontal coordinate, and the upward direction along the column direction is the positive direction of the vertical coordinate.

[0029] The beneficial technical effects of this application are:

[0030] The present application discloses a lane detection method for millimeter-wave radar based on optical flow method. The method uses point cloud data provided by millimeter-wave radar for lane detection. Considering the data sparsity and the fact that millimeter-wave radar is greatly affected by noise, the traditional optical flow method is modified to construct an error function based on the change of cluster point position to achieve motion estimation. The improved optical flow method is more suitable for processing the data characteristics unique to millimeter-wave radar, and can improve the accuracy of cluster point position estimation, and provide more stable motion estimation in millimeter-wave radar scenes with noise, thereby improving the adaptability to low-resolution and long-distance targets. Finally, a sliding window is used for sliding fitting to achieve lane detection. Since the high-precision point cloud data of millimeter-wave radar is more flexible and can provide accurate information about the lane, this method can show excellent flexibility in complex weather conditions and changing light environments. The sliding window fitting method is also the method that performs well in processing nonlinear lanes, overcoming the limitations of traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a method flow chart of a millimeter wave radar lane detection method according to an embodiment of the present application.

[0032] Figure 2 is a schematic diagram of two sliding windows initialized in an example.

[0033] Figure 3 yes Figure 2 In the example, when the positions of two initialized sliding windows are updated once respectively, a schematic diagram of the sliding windows before and after the position update is shown.

[0034] Figure 4 yes Figure 2 In the example, a schematic diagram is shown of continuously updating the positions of two sliding windows for multiple times. DETAILED DESCRIPTION

[0035] The specific implementation of the present application is further described below in conjunction with the accompanying drawings.

[0036] This application discloses a millimeter wave radar lane detection method based on optical flow method, please refer to Figure 1 The method comprises the following steps:

[0037] Step 1: Use the road-side millimeter-wave radar to obtain the t-th frame point cloud data of the traffic scene, and the initial value of the integer parameter t is 1.

[0038] Step 2: When t = 1, initialize each cluster point in the first frame of point cloud data as a key cluster point. Starting from any t ≥ 2, use the optical flow method to determine the cluster points in the t-th frame of point cloud data that have a matching relationship with the key cluster points in the t-1th frame of point cloud data based on the change in cluster point position as the key cluster points in the t-th frame of point cloud data.

[0039] The Lucas-Kanade (LK) optical flow algorithm is a classic sparse optical flow algorithm. The traditional optical flow method estimates the movement direction and speed of feature points in the image based on the assumption that the local pixel brightness is constant. This application uses the traditional optical flow method used for image pixel processing to process the point cloud data acquired by the millimeter-wave radar. However, the millimeter-wave radar does not rely on high-resolution pixels like the image sensor. The point cloud data it collects is often sparse and greatly affected by noise. Therefore, it is difficult to rely on direct pixel intensity changes for motion estimation like the traditional optical flow method. Therefore, in order to better adapt to the application scenarios and data characteristics of road-side millimeter-wave radars, this application performs motion estimation based on changes in cluster point positions.

[0040] Specifically, the present application improves the construction method of the error function in the optical flow method. The traditional optical flow method directly uses the square sum of the pixel intensity differences in the image to construct the error function, while the present application constructs the error function based on the change in cluster point position. In one embodiment, the error function E constructed based on the change in cluster point position is:

[0041]

[0042] Among them, (x t-1 ,y t-1 ) is the position of a key cluster point in the t-1th frame point cloud data, (dx, dy) is the key cluster point (x t-1 ,y t-1 ) displacement. Ω t-1 It is the set of all key cluster points in the t-1th frame point cloud data. t-1 is the cluster point set composed of all cluster points in the t-1th frame point cloud data, C t It is the cluster point set composed of all cluster points in the t-th frame point cloud data.

[0043] G(C t-1 ,(x t-1 ,yt-1 )) is based on the cluster point set C t-1 The calculated position (x t-1 ,y t-1 ), G(C t ,(x t-1 +dx,y t-1 +dy)) is based on the cluster point set C t The calculated position (x t-1 +dx,y t-1 +dy).

[0044] After constructing the error function E, the optical flow method can be used to calculate the optical flow based on the error function E. This part is the same as the traditional optical flow method, which also uses matrix operations to find the displacement of the key cluster points that minimize the error function, thereby realizing motion estimation, and then being able to determine the optical flow vector composed of the key cluster points that have a matching relationship in two adjacent frames of point cloud data. This application will not elaborate on the specific solution method based on the error function E.

[0045] As can be seen from the above introduction, since it is impossible to directly use the pixel intensity in the image like the traditional optical flow method, the function G() is introduced in this embodiment to obtain the fitting pixel value according to the distribution of cluster point positions, which is also the focus of constructing the error function E in this embodiment. t-1 ,(x t-1 ,y t-1 )) and G(C t ,(x t-1 +dx,y t-1 +dy)) both use the function G( ), and the calculation principle is the same. Next, the method of calculating the fitting pixel value G(C, (α, β)) at the position (α, β) according to the distribution of cluster point positions in an arbitrary cluster point set C is introduced, including the following steps:

[0046] First, the candidate pixel values ​​at the position (α, β) and its multiple neighboring positions are calculated according to the distribution of the cluster point positions in the cluster point set C. For any position, when there is a cluster point in the cluster point set C and the Euclidean distance of the position is less than the distance threshold, the candidate pixel value at the position is determined to be 255, otherwise the candidate pixel value at the position is determined to be 0. The distance threshold here is a custom value, such as 1.0.

[0047] The above method can be used to determine that the candidate pixel values ​​at the position (α, β) and its multiple neighboring positions are 255 or 0. Then, the idea of ​​bilinear interpolation is used to weight the candidate pixel values ​​at the position (α, β) and its multiple neighboring positions, and finally the fitting pixel value G(C, (α, β)) at the position (α, β) is obtained.

[0048] In this process, the multiple neighboring positions of the position (α, β) used include: Location and location in, means rounding α upwards. Indicates that β is rounded up. Then the above weighted process of obtaining the fitted pixel value G(C, (α, β)) can be expressed as the following formula:

[0049]

[0050] Wherein, spv(C, (α, β)) is the candidate pixel value at the position (α, β) calculated according to the distribution of cluster point positions in the cluster point set C. It is the position calculated based on the distribution of cluster points in the cluster point set C. The candidate pixel value at . It is the position calculated based on the distribution of cluster points in the cluster point set C. The candidate pixel value at . It is the position calculated based on the distribution of cluster points in the cluster point set C. The candidate pixel values ​​at these four positions are 255 or 0 determined by the above method.

[0051] ω1, ω2, ω3 and w4 are weights respectively. These four weights can be set by user, or calculated based on the distance between positions based on the idea of ​​bilinear interpolation. In another embodiment, they are calculated and determined based on the position (α, β), including:

[0052]

[0053] Among them, {α} represents taking the decimal part of α, and {β} represents taking the decimal part of β.

[0054] For example, in one example, assuming that the position (α, β) is (3.1, 14.7), its neighboring positions include (4, 14.7), (3.1, 15), and (4, 15), where:

[0055] Assume that the candidate pixel value at the position (3.1, 14.7) is 255, and the corresponding calculation weight w1=(1-0.1)×(1-0.7)=0.27.

[0056] Assume that the candidate pixel value of the neighborhood position (4, 14.7) is 255, and the corresponding calculation weight is w2=0.1×(1-0.7)=0.03.

[0057] Assume that the candidate pixel value of the neighborhood position (3.1, 15) is 255, and the corresponding calculation weight is w3 = (1-0.1) × 0.7 = 0.63.

[0058] Assume that the candidate pixel value of the neighborhood position (4, 15) is 0, and the corresponding calculation weight is w4=0.1×0.7=0.07.

[0059] The weighted fitting pixel value at the position (α, β) is G(C, (α, β))=0.27×255+0.03×255+0.63×255+0.07×0=237.15.

[0060] Step 3, calculate the angle between the optical flow vector formed by each pair of key cluster points with a matching relationship and the horizontal line. The calculated optical flow vector is a vector in the radar coordinate system of the road-side millimeter-wave radar. The horizontal line is the horizontal direction of the radar coordinate system of the road-side millimeter-wave radar. Therefore, the angle between the optical flow vector and the horizontal line can be directly calculated by trigonometric functions.

[0061] Step 4, rasterize the point cloud data of the tth frame, and set the grids whose number of target key cluster points reaches the number threshold as valid grids, and set the grids whose number of target key cluster points does not reach the number threshold as invalid grids, to obtain the binary bitmap of the point cloud data of the tth frame. The grid size during rasterization is generally selected to be smaller to improve the accuracy of subsequent fitting detection, for example, it is generally set to 2.

[0062] Among them, the target key cluster point is the key cluster point whose angle between the optical flow vector formed in the t-th frame point cloud data and the horizontal line is within a predetermined angle range. The predetermined angle range is a predetermined range of the angle between the vehicle flow motion vector and the horizontal line, which can be obtained through statistics and generally includes an angle range of 60° to 120° and 240° to 300°.

[0063] The optical flow vectors formed by all the key cluster points with matching relationships screened out in the above steps are all motion vectors. The vehicle flow motion vectors can be further screened out using the angle with the horizontal line, so that subsequent lane detection can be performed based on the vehicle flow motion vectors.

[0064] Step 5: Use a sliding window to locally fit and splice the valid grids in the binary bitmap of the t-th frame point cloud data to obtain the lane detection result.

[0065] Fit the effective grids in the local area of ​​each sliding window, then slide the sliding window to different positions to perform local fitting on different local areas, and then splice these local fitting results to form the overall lane detection result. This method has good flexibility and only processes local data in each sliding window, which means it can better cope with drastic changes in road curves without over-smoothing or ignoring local features due to global fitting. Compared with the global fitting approach, it can dynamically respond to changes in road curvature, reduce the impact of noise data on the entire lane fitting, and improve the accuracy and stability of fitting, which is particularly suitable for nonlinear lanes.

[0066] In one embodiment, the method of fitting using a sliding window includes:

[0067] (1) Determine the number of sliding windows n and the window size of each sliding window. The number of sliding windows is consistent with the number of lanes included in the traffic scene range covered by the road-side millimeter-wave radar. The sliding window adopts a rectangular window, and the window size of each sliding window includes a window width w and a window height h.

[0068] (2) Divide the binary bitmap of the point cloud data of the tth frame into n regions along the row direction, and take the grid column with the largest number of valid grids in each region as the initial detection column of the current region. Figure 2 For example, take n as 2. Figure 2 The white grid in the middle represents a valid grid. In this example, the binary bitmap is divided into left and right areas, and then a grid column containing the largest number of valid grids is searched in the left and right areas respectively.

[0069] (3) Initialize a sliding window at the initial detection column in each area according to the window size. The window center of the initialized sliding window is located at the initial detection column, and the lower edge of the sliding window coincides with the lower edge of the binary bitmap of the t-th frame point cloud data.

[0070] For the convenience of description, a grid coordinate system is established with the lower left corner of the binary bitmap as the origin, the rightward direction along the row direction as the positive direction of the horizontal coordinate, and the upward direction along the column direction as the positive direction of the vertical coordinate. Then, each grid in the binary bitmap can be represented by its horizontal and vertical coordinates in the grid coordinate system. During initialization, the horizontal coordinate of the initial detection column is used as the horizontal coordinate of the window center, and the vertical coordinate of the window center is set to h / 2, and then a sliding window can be initialized by combining the window width w and the window height h.

[0071] for example Figure 2In the example, the grid column containing the largest number of valid grids in the left area is selected as the initial detection column and the sliding window shown in the green rectangular box is initialized, and the grid column containing the largest number of valid grids in the right area is selected as the initial detection column and the sliding window shown in the red rectangular box is initialized.

[0072] (4) For each sliding window, extract valid grids in the sliding window. When the number of valid grids in the sliding window reaches the minimum number of grids, update the position of the sliding window according to the valid grids in the sliding window, and perform the step of extracting valid grids in the sliding window again. When the number of valid grids in the sliding window does not reach the minimum number of grids, the updating of the sliding window is terminated. The minimum number of grids here is a user-defined setting.

[0073] When updating the position of a sliding window, the average value of the horizontal coordinates of all valid grids contained in the sliding window in the grid coordinate system is calculated as the horizontal coordinate of the window center of the updated sliding window, and the vertical coordinate of the window center of the current sliding window in the grid coordinate system is increased by the height h of the sliding window as the vertical coordinate of the window center of the updated sliding window.

[0074] For example, in Figure 2 When the two initialized sliding windows perform the first position update, the horizontal coordinates of the window centers of the two sliding windows are updated according to the positions of the valid grids in their respective sliding windows, and the vertical coordinates of the window centers of the two sliding windows are updated from h / 2 to 3h / 2, and the window size of the sliding window remains unchanged, such as Figure 3 As shown. Figure 3 The schematic diagram of updating and detecting the sliding window position several times based on Figure 4 shown.

[0075] All valid grids extracted by the same sliding window during the position update process are spliced ​​and merged to obtain the detection result of a lane. Since the high-precision point cloud data of the millimeter-wave radar is more flexible, it can provide accurate information about the lane, especially in bad weather and light changes. Therefore, this method shows excellent flexibility and reliability in dealing with various road conditions, and the sliding window local fitting method can handle nonlinear lanes with excellent performance. In fact, by introducing human interference factors such as moving obstacles and road signs, and conducting experiments at intersections with different geographical locations, time periods and traffic flows, this method can have good robustness.

[0076] The above is only a preferred embodiment of the present application, and the present application is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present application should be considered to be included in the protection scope of the present application.

Claims

1. A millimeter wave radar lane detection method based on optical flow method, characterized in that: The millimeter wave radar lane detection method comprises: The t-th frame point cloud data of the traffic scene is obtained by using the road-side millimeter-wave radar, and each cluster point in the first frame point cloud data is initialized as a key cluster point, and the initial value of the integer parameter t is 1; Starting from any t≥2, the optical flow method is used to determine the cluster points in the t-th frame point cloud data that have a matching relationship with the key cluster points in the t-1th frame point cloud data based on the change in cluster point position, and the angle between the optical flow vector formed by each pair of matching key cluster points and the horizontal line is calculated; The t-th frame point cloud data is rasterized, and the grids whose number of target key cluster points reaches the number threshold are set as valid grids, and the grids whose number of target key cluster points does not reach the number threshold are set as invalid grids, so as to obtain a binary bitmap of the t-th frame point cloud data; wherein the target key cluster point is a key cluster point whose angle between the optical flow vector and the horizontal line formed in the t-th frame point cloud data is within a predetermined angle range; The sliding window is used to locally fit and splice the valid grids in the binary bitmap of the t-th frame point cloud data to obtain the lane detection result.

2. The millimeter wave radar lane detection method according to claim 1, characterized in that: The cluster points in the t-th frame point cloud data that match the key cluster points in the t-1-th frame point cloud data are determined by using the optical flow method based on the change in cluster point position. The cluster points include: The error function E is constructed based on the change of cluster point position: Among them, (x t-1 ,y t-1 ) is the position of a key cluster point in the t-1th frame of point cloud data, (dx, dy) is the key cluster point (x t-1 ,y t-1 ) displacement; Ω t-1 is the set of all key cluster points in the t-1th frame point cloud data, C t-1 is the cluster point set composed of all cluster points in the t-1th frame point cloud data, C t is the cluster point set composed of all cluster points in the t-th frame point cloud data; G(C t-1 ,(x t-1 ,y t-1 )) is based on the cluster point set C t-1 The calculated position (x t-1 ,y t-1 ), G(C t ,(x t-1 +dx,y t-1 +dy)) is based on the cluster point set C t The calculated position (x t-1 +dx,y t-1 +dy) at the fitting pixel value; Based on the constructed error function E, the optical flow method is used to perform optical flow calculation to determine the optical flow vector composed of key cluster points with matching relationship.

3. The millimeter wave radar lane detection method according to claim 2, characterized in that: Calculating the fitting pixel value G(C, (α, β)) at the position (α, β) according to the distribution of cluster point positions in any cluster point set C includes: According to the distribution of cluster point positions in the cluster point set C, the candidate pixel values ​​at the position (α, β) and its multiple neighboring positions are calculated respectively; for any position, when there is a cluster point in the cluster point set C and the Euclidean distance between the position is less than the distance threshold, the candidate pixel value at the position is determined to be 255, otherwise the candidate pixel value at the position is determined to be 0; The candidate pixel values ​​at the position (α, β) and its multiple neighboring positions are weighted to obtain the fitting pixel value G(C, (α, β)) at the position (α, β).

4. The millimeter wave radar lane detection method according to claim 3, characterized in that: The candidate pixel values ​​at position (α, β) and its multiple neighboring positions are weighted to obtain the fitted pixel value G(C, (α, β)) at position (α, β): Among them, spυ(C, (α, β)) is the candidate pixel value at the position (α, β) calculated according to the distribution of cluster point positions in the cluster point set C. It is the position calculated based on the distribution of cluster points in the cluster point set C. The candidate pixel value at It is the position calculated based on the distribution of cluster points in the cluster point set C. The candidate pixel value at It is the position calculated based on the distribution of cluster points in the cluster point set C. Candidate pixel value at ; means rounding α upwards. represents rounding up of β; ω1, ω2, ω3 and ω4 are weights respectively.

5. The millimeter wave radar lane detection method according to claim 4, characterized in that: Among them, {α} represents taking the decimal part of α, and {β} represents taking the decimal part of β.

6. The millimeter wave radar lane detection method according to claim 1, characterized in that: The lane detection results obtained by locally fitting and splicing the valid grids in the binary bitmap of the t-th frame point cloud data using a sliding window include: Determine the number n of sliding windows and the window size of each sliding window; Divide the binary bitmap of the point cloud data of the tth frame into n regions along the row direction, and take the grid column with the largest number of valid grids in each region as the initial detection column of the current region; Initializing a sliding window at the initial detection column of each region according to the window size, wherein the window center of the sliding window is located at the initial detection column, and the lower edge of the sliding window coincides with the lower edge of the binary bitmap of the t-th frame point cloud data; For each sliding window, extract the valid grids in the sliding window. When the number of valid grids contained in the sliding window reaches the minimum number of grids, update the position of the sliding window according to the valid grids contained in the sliding window, and perform the step of extracting the valid grids in the sliding window again. All valid grids extracted from the same sliding window during the position update process are spliced ​​and merged to obtain the detection result of a lane.

7. The millimeter wave radar lane detection method according to claim 6, characterized in that: Updating the position of the sliding window according to the valid grid contained in the sliding window includes: The average value of the horizontal coordinates of all valid grids contained in the sliding window in the grid coordinate system is calculated as the horizontal coordinate of the window center of the updated sliding window, and the vertical coordinate of the window center of the current sliding window in the grid coordinate system is increased by the height of the sliding window as the vertical coordinate of the window center of the updated sliding window; wherein the grid coordinate system takes the lower left corner of the binary bitmap as the origin, the rightward direction along the row direction is the positive direction of the horizontal coordinate, and the upward direction along the column direction is the positive direction of the vertical coordinate.

Images