A fusion positioning method based on multi-sensor in dynamic scenes

Abstract

The present invention relates to a fusion positioning method based on multiple sensors in a dynamic scene, the method comprising: S1, obtaining a dynamic event point cloud collected by a dynamic vision sensor camera and environmental point cloud data collected by a laser radar; S2, processing the dynamic event point cloud and filtering out noise events to obtain a dynamic object image; S3, identifying the dynamic object in the dynamic object image, and selecting the dynamic object area; S4, mapping the dynamic object area to the environmental point cloud data, removing the dynamic object point cloud, and obtaining a static environmental point cloud; S5, registering the static environmental point cloud with map features to obtain positioning information. Compared with the prior art, the positioning accuracy and robustness of the present invention are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a multi-sensor fusion positioning method in a dynamic scene. Background Art

[0002] Autonomous driving vehicles operate in an unsupervised state. The autonomous driving brain can replace the driver, simulate human observation (sensor system), thinking (driving brain) and operation (planning and control), and then control the wheels to complete the action task. The main functions involved in autonomous driving include self-driving positioning, moving target detection, route detection, path planning, vehicle tracking control and other autonomous driving functions.

[0003] In the vehicle's autonomous navigation system, accurate positioning is an important part of completing path planning, trajectory tracking control, and navigation to the target location. At present, LiDAR and cameras have been widely used for environmental perception. These artificial intelligence sensors have been used to extract and identify environmental targets, and apply landmark matching to achieve autonomous positioning of vehicles. However, there are inevitably many moving objects in the actual environment, such as pedestrians, cars, bicycles, etc. And the appearance of moving objects is incorrectly matched with map features, which will lead to a serious decrease in positioning accuracy. Therefore, improving the positioning accuracy of LiDAR and camera fusion in dynamic scenes has become an important research branch. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a multi-sensor fusion positioning method in dynamic scenes.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A multi-sensor fusion positioning method in a dynamic scene, the method comprising:

[0007] S1, obtaining dynamic event point cloud collected by the dynamic vision sensor camera and environmental point cloud data collected by the laser radar;

[0008] S2, processing the dynamic event point cloud and filtering out noise events to obtain a dynamic object image;

[0009] S3, identifying the dynamic object in the dynamic object image and selecting the dynamic object area;

[0010] S4, mapping the dynamic object area to the environmental point cloud data, removing the dynamic object point cloud to obtain the static environmental point cloud;

[0011] S5. Align the static environment point cloud with the map features to obtain positioning information.

[0012] Preferably, step S2 specifically includes:

[0013] S21, projecting the dynamic event point clouds at different sampling moments within the sampling time period onto the same image plane to obtain a dynamic event point cloud image;

[0014] S22, performing filtering processing based on the event point cloud thickness of each pixel in the dynamic event point cloud image to filter out noise point clouds;

[0015] S23, binarizing the dynamic event point cloud image after filtering out the noise point cloud to obtain a dynamic object image.

[0016] Preferably, step S22 is specifically:

[0017] S221, counting the number of event point clouds at each pixel in the dynamic event point cloud image, recorded as the point cloud thickness n of the i-th pixel i , i = 0, 1, ..., L-1, L is the total number of pixels;

[0018] S222, calculate the probability P that the number of event point clouds of the i-th pixel accounts for the total number of event points i And the average event point cloud thickness μ T :

[0019]

[0020]

[0021] n=n1+n2+…+n L-1

[0022] S223, sort the point cloud thicknesses of the L pixels from large to small, To find the point cloud thickness threshold n for distinguishing noise point cloud and dynamic object point cloud T ,in,

[0023]

[0024]

[0025]

[0026] j represents the jth pixel after the point cloud thickness is sorted from large to small, w0(n T ) represents the probability of the dynamic object area, w1(n T ) represents the probability of the noise region, μ0(n T ) represents the average point cloud thickness of the dynamic object area, μ1(n T ) represents the average point cloud thickness in the noise area, is the point cloud thickness variance of the dynamic object area and the noise area;

[0027] S224, the point cloud thickness in the dynamic event point cloud image is less than n T The points are regarded as noise points and the noise points are filtered out.

[0028] Preferably, step S23 performs binarization processing according to the following formula to obtain a dynamic object image:

[0029]

[0030] Among them, f(x,y) represents the grayscale value of the pixel with coordinates (x,y), and n(x,y) represents the point cloud thickness of the pixel with coordinates (x,y).

[0031] Preferably, after acquiring the dynamic object image in step S23, dilated convolution processing is also performed, and during the dilated convolution processing, the grayscale value of each pixel is replaced by the minimum grayscale value in the convolution kernel.

[0032] Preferably, step S3 uses a YOLO-small network to identify dynamic objects in the dynamic object image.

[0033] Preferably, step S5 is specifically:

[0034] S51, establishing a static environment point cloud error model, and using the static environment point cloud error model to perform error evaluation on the static environment point cloud;

[0035] S52. Determine the error weight of each point based on the error evaluation result of the static environment point cloud, and align the static environment point cloud with the map features based on the error weight to obtain positioning information.

[0036] Preferably, the static environment point cloud error model includes a uniform distribution error model and a Gaussian distribution error model. The uniform distribution error model is used to calculate the uniform distribution error of each point in the static environment point cloud, and the Gaussian distribution error model is used to calculate the Gaussian distribution error of each point in the static environment point cloud.

[0037] Preferably, the error weight of each point in the static environment point cloud in step S52 is:

[0038]

[0039]

[0040]

[0041] in, is the error weight of the i-th point in the static environment point cloud, is the weight corresponding to the uniform distribution error model of the i-th point in the static environment point cloud, is the weight corresponding to the Gaussian distribution error model of the i-th point in the static environment point cloud, PP i even is the information entropy uniformly distributed at the i-th point in the static environment point cloud, P i even is the relative probability distribution information entropy of the i-th point in the current static environment point cloud in the uniform distribution error model; PP i gaussian is the information entropy of the Gaussian distribution of the i-th point in the static environment point cloud, P i gaussian is the relative probability distribution information entropy of the i-th point of the current static environment point cloud in the Gaussian distribution error model.

[0042] Preferably, P i even and P i gaussian They are:

[0043] P i even =ln[λ 1i λ 2i λ 3i d i 3 ]

[0044]

[0045] λ 1i =θd i ,λ 2i =φd i ,λ 2i =hd i

[0046] Among them, θ is the horizontal angular resolution of the laser radar, φ is the vertical angular resolution of the laser radar, h is the air dielectric constant, and d i is the scanning distance of the i-th point in the static environment point cloud.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] (1) Due to the interference of moving targets in dynamic scenes, there are positioning and registration errors in laser point clouds. At the same time, there are sparsity and multi-scale characteristics between the multi-beams of laser radar, which causes spatial distribution errors in the features extracted from the point cloud, leading to point cloud registration and positioning errors. Therefore, the present invention performs fusion positioning based on the dynamic event point cloud collected by the dynamic visual sensor camera and the environmental point cloud data collected by the laser radar, and uses the dynamic event point cloud to propose dynamic objects in the environmental point cloud data, thereby improving the accuracy and robustness of laser positioning.

[0049] (2) The present invention establishes a static environment point cloud error model to perform error evaluation on the static environment point cloud, optimizes the matching weight of features, and improves positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a general flow chart of a multi-sensor fusion positioning method in a dynamic scene according to the present invention;

[0051] Figure 2 This is an architecture diagram of a multi-sensor fusion positioning method in a dynamic scene according to the present invention;

[0052] Figure 3 Schematic diagram of event point recognition of a bicycle by a dynamic vision sensor camera in an embodiment of the present invention.

[0053] Figure 4 It is a schematic diagram of processing a dynamic event point cloud and filtering out noise events to obtain a dynamic object image in an embodiment of the present invention;

[0054] Figure 5 This is a network diagram of the ResN structure in the YOLO-small network;

[0055] Figure 6 This is a network diagram of the feature layer and Neck layer in the YOLO-small network;

[0056] Figure 7 Schematic diagram of characteristic error distribution of static LiDAR points;

[0057] Figure 8 Schematic diagram of stereo error distribution for uniform distribution and Gaussian distribution;

[0058] Fig. 9 Schematic diagram of error model for uniform distribution and Gaussian distribution;

[0059] Fig.10 The distribution probability curves are uniform distribution and Gaussian distribution. DETAILED DESCRIPTION

[0060] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Note that the following embodiments are merely illustrative in nature, and the present invention is not intended to limit its applicable objects or uses, and the present invention is not limited to the following embodiments.

[0061] Example

[0062] like Figure 1 , Figure 2 As shown, this embodiment provides a fusion positioning method based on multiple sensors in a dynamic scene, the method comprising:

[0063] S1, obtaining dynamic event point cloud collected by the dynamic vision sensor camera and environmental point cloud data collected by the laser radar;

[0064] S2, processing the dynamic event point cloud and filtering out noise events to obtain a dynamic object image;

[0065] S3, identifying the dynamic object in the dynamic object image and selecting the dynamic object area;

[0066] S4, mapping the dynamic object area to the environmental point cloud data, removing the dynamic object point cloud to obtain the static environmental point cloud;

[0067] S5. Align the static environment point cloud with the map features to obtain positioning information.

[0068] Step S2 is specifically as follows:

[0069] S21, projecting the dynamic event point clouds at different sampling moments within the sampling time period onto the same image plane to obtain a dynamic event point cloud image;

[0070] S22, performing filtering processing based on the event point cloud thickness of each pixel in the dynamic event point cloud image to filter out noise point clouds, wherein the filtering processing adopts the Otsu method;

[0071] S23, binarizing the dynamic event point cloud image after filtering out the noise point cloud to obtain a dynamic object image.

[0072] Step S22 is specifically as follows:

[0073] S221, counting the number of event point clouds at each pixel in the dynamic event point cloud image, recorded as the point cloud thickness n of the i-th pixel i , i = 0, 1, ..., L-1, L is the total number of pixels;

[0074] S222, calculate the probability P that the number of event point clouds of the i-th pixel accounts for the total number of event points i And the average event point cloud thickness μ T :

[0075]

[0076]

[0077] n=n1+n2+…+n L-1

[0078] S223, sort the point cloud thicknesses of the L pixels from large to small, To find the point cloud thickness threshold n for distinguishing noise point cloud and dynamic object point cloud T,in,

[0079]

[0080]

[0081]

[0082] j represents the jth pixel after the point cloud thickness is sorted from large to small, w0(n T ) represents the probability of the dynamic object area, w1(n T ) represents the probability of the noise region, μ0(n T ) represents the average point cloud thickness of the dynamic object area, μ1(n T ) represents the average point cloud thickness in the noise area, is the point cloud thickness variance of the dynamic object area and the noise area;

[0083] S224, the point cloud thickness in the dynamic event point cloud image is less than n T The points are regarded as noise points and the noise points are filtered out.

[0084] Step S23 performs binarization processing according to the following formula to obtain a dynamic object image:

[0085]

[0086] Among them, f(x,y) represents the grayscale value of the pixel with coordinates (x,y), and n(x,y) represents the point cloud thickness of the pixel with coordinates (x,y).

[0087] After acquiring the dynamic object image in step S23, a dilated convolution process is also performed, during which the grayscale value of each pixel is replaced by the minimum grayscale value in the convolution kernel.

[0088] Specifically, in this embodiment, Figure 3 As shown in the figure, the DVS camera captures a bicycle and projects the event point cloud onto the image plane. Affected by the environment, there are noise points on the image plane. And the image with noise points will seriously expand the area of ​​the moving object, which will bring challenges to the recognition of the moving object. The number of event points of the DVS camera is related to the moving speed and color change rate in a very short period of time. Therefore, the above Otsu method aims to calculate the judgment threshold and realize the filtering operation. As the moving speed of the object changes, the threshold n T The Otsu method of the present invention retains the event body features of the moving object and filters out the noise points. After that, the DVS image is dilated and convolved to enhance the main feature area and connect the scattered areas to expand the body area of ​​the moving object. The dynamic object image after filtering and dilation is as follows: Figure 4shown.

[0089] Step S3 uses the YOLO-small network to identify dynamic objects in the dynamic object image. The detection network consists of three parts: feature layer, neck layer and YOLO layer. The objects detected by the DVS camera include cars, pedestrians and bicycles.

[0090] Feature layer: Since the DVS camera has the advantage of only detecting moving objects, there are only two types of pixels (background static pixels and object moving pixels). Therefore, the feature extraction layer can be simplified. Design RES structure (such as Figure 5 ), and use it as the basic unit for extracting features from DVS images.

[0091] Neck layer: Perform a single convolution operation on the deepest feature layer. After that, it is combined with other feature layers through upsampling transmission. In this way, the gradient vanishing problem caused by the deepening of the feature layer can be alleviated. The connection network between the feature layer and the Neck layer is as follows Figure 6 shown.

[0092] YOLO layer: DVS-based image detection has only one label (dynamic object). The YOLO layer has three pre-verification boxes, and each prediction box includes five parts: x offset, y offset, length, width, and confidence index. So the dimension of the last output layer is 3*(5+1)=3*6=18.

[0093] Loss function: The loss function consists of six parts: x-coordinate, y-coordinate, width, height, IOU error and prediction category of the prediction center.

[0094] Step S5 is specifically as follows:

[0095] S51, establishing a static environment point cloud error model, and using the static environment point cloud error model to perform error evaluation on the static environment point cloud;

[0096] S52. Determine the error weight of each point based on the error evaluation result of the static environment point cloud, and align the static environment point cloud with the map features based on the error weight to obtain positioning information.

[0097] Due to the sparsity of LiDAR points, there is a distribution deviation between the static features extracted and the actual features after deducting dynamic objects through LiDAR and camera recognition. The deviation is distributed along the width, height and depth directions as shown in Figure 7 As shown. Assume that the extracted features have a distributed error vector. The error vector is pe = [δx, δy, δz], δx is related to the horizontal angular resolution θ and the scanning distance d, δy is related to the vertical angular resolution φ and the scanning distance d, δz is related to the air medium h and the scanning distance d, δx = θd = λ1, δy = φd = λ2, δz = hd = λ3.

[0098] Information entropy is defined as the probability of occurrence of discrete random events. It is one of the most important theories in modern measurement. Therefore, information entropy can be introduced into the evaluation of feature distribution uncertainty. At the same time, it is a representation of the size or volume of distribution error. Information entropy and error entropy are expressed as:

[0099]

[0100] Where: P(X) represents the probability of event X.

[0101] The actual distribution law of point features will affect the calculation results of information entropy, so it is necessary to accurately establish the probability mathematical expression of the error model. If the error distribution follows a uniform distribution, the error volume is in the shape of a cube. If the error distribution follows a Gaussian distribution, the error volume is in the shape of an ellipse. The three-dimensional error distribution of uniform distribution and Gaussian distribution is as follows: Figure 8 As shown, Figure 8 (a) is the uniformly distributed stereo error distribution, where Figure 8 (b) is the stereo error distribution of Gaussian distribution. Fig. 9 Schematic diagram of error model for uniform distribution and Gaussian distribution. Fig. 9 (a) is a schematic diagram of the uniformly distributed error model, where Fig. 9 (b) is a schematic diagram of the Gaussian distribution error model. Fig.10 is the distribution probability curve of uniform distribution and Gaussian distribution, Fig.10 (a) is the distribution probability curve of uniform distribution, where Fig.10 (b) is the distribution probability curve of Gaussian distribution.

[0102] The distribution function of the uniform distribution is:

[0103]

[0104] The distribution function of the Gaussian distribution is:

[0105]

[0106] In the uniform distribution model, there is only one parameter, namely (x)~U(-0.5λ,0.5λ). In fact, all stereo points are within the range of the cube: [-0.5λ1,0.5λ1], [-0.5λ2,0.5λ2] and [-0.5λ3,0.5λ3]. Since the total probability integral of the error cube is 1, equation (1-1) can be obtained. Then, equation (1-2) can be further simplified. According to the definitions of information entropy and error entropy, the error entropy of the uniform distribution model can be calculated using equation (1-3). The calculation results of the error entropy show that the detection accuracy varies linearly with the scanning distance.

[0107]

[0108]

[0109]

[0110] S even =e P =λ1λ2λ3d 3 =1.84*10 -7 *d 3 (1-3)

[0111] Where f(x,y,z) represents the uniform distribution probability of a spatial point, P even The information entropy of the average probability distribution, S even Represents the error entropy of the mean probability distribution.

[0112] In the Gaussian distribution model, there is only one parameter, (x)~N(0,λ2). In fact, 97% of the lidar points fall within the range of the ellipsoid: [-λ1+μ,λ1+μ], [-λ2+μ,λ2+μ], and [-λ3+μ,λ3+μ].

[0113] make: Equation (1-4) can be further simplified to equation (1-5):

[0114]

[0115]

[0116] Where f(x,y,z) represents the Gaussian distribution probability of a spatial point, P gaussian Represents the information entropy of a Gaussian probability distribution.

[0117] Through the spherical surface integral and distribution integral methods, equation (1-5) can be transformed into equation (1-6). According to the definitions of information entropy and error entropy, the error entropy of Gaussian distribution can also be calculated using equation (1-7).

[0118]

[0119]

[0120] Among them, P gaussian Represents the information entropy of Gaussian probability distribution, S gaussian Represents the error entropy of the Gaussian probability distribution.

[0121] Based on the above, the static environment point cloud error model includes a uniform distribution error model and a Gaussian distribution error model. The uniform distribution error model is used to calculate the uniform distribution error of each point in the static environment point cloud, and the Gaussian distribution error model is used to calculate the Gaussian distribution error of each point in the static environment point cloud.

[0122] The error weight of each point in the static environment point cloud in step S52 is:

[0123]

[0124]

[0125]

[0126] in, is the error weight of the i-th point in the static environment point cloud, is the weight corresponding to the uniform distribution error model of the i-th point in the static environment point cloud, is the weight corresponding to the Gaussian distribution error model of the i-th point in the static environment point cloud, PP i even is the information entropy uniformly distributed at the i-th point in the static environment point cloud, P i even is the relative probability distribution information entropy of the i-th point in the current static environment point cloud in the uniform distribution error model; PP i gaussian is the information entropy of the Gaussian distribution of the i-th point in the static environment point cloud, P i gaussian is the relative probability distribution information entropy of the i-th point of the current static environment point cloud in the Gaussian distribution error model.

[0127] P i even and P i gaussian They are:

[0128] P i even =ln[λ 1i λ 2i λ 3i d i 3 ]

[0129]

[0130] λ 1i =θd i ,λ 2i =φd i ,λ 2i =hd i

[0131] Among them, θ is the horizontal angular resolution of the laser radar, φ is the vertical angular resolution of the laser radar, h is the air dielectric constant, and d i is the scanning distance of the i-th point in the static environment point cloud.

[0132] When the vehicle is driving autonomously, there is a map feature set G and a current driving feature set P. By rotating the shift matrix (R, H), the feature set P is recalculated as the feature set PT. The least squares distance error (Ematching) from PT to G is calculated according to the following formula:

[0133]

[0134] Where R represents the rotational attitude change matrix, and H represents the translation change matrix.

[0135] If the Ematching value is less than the predefined threshold, the current feature matches the map feature successfully.

[0136] The present invention integrates laser radar and dynamic vision sensor camera (DVS camera) to achieve robust positioning of vehicle autonomous navigation. The Otsu method is used to filter scattered noise events on the dynamic event point cloud image, and the expansion operation is performed to expand the dynamic area to solve the dispersion problem of the DVS image. The YOLO-small network is designed to identify, cluster and mark moving objects, and obtain the remaining static point cloud of the laser radar by deducting the moving objects. Then the static environment point cloud is evaluated and positioned, and the information entropy and error entropy of the model are solved by establishing a three-dimensional distribution error model of the laser radar point feature, optimizing the matching weight of the feature, and improving the positioning accuracy.

[0137] The above embodiments are merely examples and do not limit the scope of the present invention. These embodiments can also be implemented in various other ways, and various omissions, substitutions, and changes can be made without departing from the technical concept of the present invention.

Claims

1. A multi-sensor fusion positioning method in a dynamic scene, characterized in that: The method includes: S1, obtaining dynamic event point cloud collected by the dynamic vision sensor camera and environmental point cloud data collected by the laser radar; S2, processing the dynamic event point cloud and filtering out noise events to obtain a dynamic object image; S3, identifying the dynamic object in the dynamic object image and selecting the dynamic object area; S4, mapping the dynamic object area to the environmental point cloud data, removing the dynamic object point cloud to obtain the static environmental point cloud; S5, registering the static environment point cloud with the map features to obtain positioning information; Step S2 is specifically as follows: S21, projecting the dynamic event point clouds at different sampling moments within the sampling time period onto the same image plane to obtain a dynamic event point cloud image; S22, performing filtering processing based on the event point cloud thickness of each pixel in the dynamic event point cloud image to filter out noise point clouds; S23, binarizing the dynamic event point cloud image after filtering out the noise point cloud to obtain a dynamic object image; Step S22 is specifically as follows: S221, counting the number of event point clouds at each pixel in the dynamic event point cloud image, recorded as the point cloud thickness n of the i-th pixel i , i = 0, 1, ..., L-1, L is the total number of pixels; S222, calculate the probability P that the number of event point clouds of the i-th pixel accounts for the total number of event points i And the average event point cloud thickness μ T : n=n1+n2+…+n L-1 S223, sort the point cloud thicknesses of the L pixels from large to small, To find the point cloud thickness threshold n for distinguishing noise point cloud and dynamic object point cloud T ,in, j represents the jth pixel after the point cloud thickness is sorted from large to small, w0(n T ) represents the probability of the dynamic object area, w1(n T ) represents the probability of the noise region, μ0(n T ) represents the average point cloud thickness of the dynamic object area, μ1(n T ) represents the average point cloud thickness in the noise area, is the point cloud thickness variance of the dynamic object area and the noise area; S224, the point cloud thickness in the dynamic event point cloud image is less than n T The points are regarded as noise points and the noise points are filtered out.

2. According to the multi-sensor fusion positioning method in a dynamic scene according to claim 1, it is characterized in that: Step S23 performs binarization processing according to the following formula to obtain a dynamic object image: Among them, f(x,y) represents the grayscale value of the pixel with coordinates (x,y), and n(x,y) represents the point cloud thickness of the pixel with coordinates (x,y).

3. The multi-sensor fusion positioning method in a dynamic scene according to claim 2, characterized in that: After acquiring the dynamic object image in step S23, a dilated convolution process is also performed, during which the grayscale value of each pixel is replaced by the minimum grayscale value in the convolution kernel.

4. The multi-sensor fusion positioning method in a dynamic scene according to claim 1, characterized in that: Step S3 uses the YOLO-small network to identify dynamic objects in the dynamic object image.

5. The multi-sensor fusion positioning method in a dynamic scene according to claim 1, characterized in that: Step S5 is specifically as follows: S51, establishing a static environment point cloud error model, and using the static environment point cloud error model to perform error evaluation on the static environment point cloud; S52. Determine the error weight of each point based on the error evaluation result of the static environment point cloud, and align the static environment point cloud with the map features based on the error weight to obtain positioning information.

6. The multi-sensor fusion positioning method in a dynamic scene according to claim 5, characterized in that: The static environment point cloud error model includes a uniform distribution error model and a Gaussian distribution error model. The uniform distribution error model is used to calculate the uniform distribution error of each point in the static environment point cloud, and the Gaussian distribution error model is used to calculate the Gaussian distribution error of each point in the static environment point cloud.

7. The multi-sensor fusion positioning method in a dynamic scene according to claim 6, characterized in that: The error weight of each point in the static environment point cloud in step S52 is: in, is the error weight of the i-th point in the static environment point cloud, is the weight corresponding to the uniform distribution error model of the i-th point in the static environment point cloud, is the weight corresponding to the Gaussian distribution error model of the i-th point in the static environment point cloud, PP i even is the information entropy uniformly distributed at the i-th point in the static environment point cloud, P i even is the relative probability distribution information entropy of the i-th point in the current static environment point cloud in the uniform distribution error model; PP i gaussian is the information entropy of the Gaussian distribution of the i-th point in the static environment point cloud, P i gaussian is the relative probability distribution information entropy of the i-th point of the current static environment point cloud in the Gaussian distribution error model.

8. The multi-sensor fusion positioning method in a dynamic scene according to claim 7, characterized in that: P i even and P i gaussian They are: P i even =ln[λ 1i l 2i l 3i d i 3 ] l 1i =θd i ,l 2i =φd i ,l 2i =hd i Among them, θ is the horizontal angular resolution of the laser radar, φ is the vertical angular resolution of the laser radar, h is the air dielectric constant, and d i is the scanning distance of the i-th point in the static environment point cloud.

Images