A LiDAR snow removal method and system integrating intensity and spatiotemporal geometric features

Abstract

The present invention provides a method and system for removing snow from a laser radar that integrates intensity and spatiotemporal geometric features, which relates to the technical field of autonomous driving environment perception. The method removes snow from point cloud data by combining the reflection intensity information of the intensity neighborhood number and the spatial feature information of the spatiotemporal neighborhood number, thereby improving the accuracy and adaptability of the snow removal algorithm, and effectively removing snow noise points in the laser radar point cloud data while reducing the loss of non-snow noise points in the point cloud data, thereby reducing the damage to the environmental information in the point cloud data. It can adapt to snowfall of various densities.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving environment perception technology, and more specifically, to a laser radar snow removal method and system that integrates intensity and spatiotemporal geometric features. Background Art

[0002] Autonomous driving technology uses computers to drive and control vehicles. During autonomous driving, the computer's perception of the surrounding environment is the basis for making driving decisions, so reliable environmental perception data is particularly important. Currently, autonomous vehicles generally use LiDAR (Light Detection and Ranging) for environmental perception. LiDAR obtains point cloud data by scanning the surrounding environment. In severe weather conditions such as heavy snow and blizzards with dense snow distribution, there is currently no good method to effectively remove snow noise from the point cloud data obtained by the LiDAR sensor and better preserve environmental characteristics.

[0003] Publication number: CN113281782A, publication date 2021-08-20, a lidar snow point filtering method based on an unmanned vehicle, the invention roughly classifies the point cloud reflection intensity according to the threshold of the lidar point cloud reflection intensity solution, and then performs fine classification according to the number of neighbors based on field search; however, this scheme only considers the correlation between reflection intensity and different materials, and solves the threshold based on this premise. The influence of measurement distance and incident angle is not considered in the solution process; and only the point cloud distribution in ordinary space is used for further filtering, and the dimension of continuous time is not considered. The threshold solution results and further filtering results are inaccurate. Summary of the invention

[0004] In order to overcome the above-mentioned technical problems, the present invention provides a LiDAR snow removal method and system which can effectively identify and remove snow noise in point cloud data and better retain the fusion strength and spatiotemporal geometric characteristics of environmental characteristics.

[0005] The technical solution of the present invention is as follows:

[0006] A method and system for removing snow by laser radar that integrates intensity and spatiotemporal geometric features, the method comprising the steps of:

[0007] S1, obtaining actual point cloud data of the laser radar, filtering the actual point cloud data, and obtaining candidate snow noise points;

[0008] S2, according to the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data, calculate the number of intensity neighborhoods of the candidate snow noise point in the point cloud frame at the current moment;

[0009] S3, performing dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames by principal component analysis to extract a width-time image, wherein the adjacent point cloud frames refer to several frames before and after the point cloud frame with the candidate snow noise point;

[0010] S4, based on the motion law characteristics of snowflakes and the width-time image, calculating the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image;

[0011] S5, based on the entropy weight method, the comprehensive score is calculated using the number of intensity neighbors and the number of spatiotemporal neighbors, and then the candidate snow noise points are finally classified into real snow noise points and non-noise points according to the comprehensive score;

[0012] S6. Remove real snow noise points to achieve filtering, and restore non-noise points to ordinary point cloud points to achieve repair.

[0013] This technical solution proposes a LiDAR snow removal method and system that integrates intensity and spatiotemporal geometric features, combines the reflection intensity information of the intensity neighborhood number and the spatial feature information of the spatiotemporal neighborhood number to remove snow from the point cloud data, improves the accuracy and adaptability of the snow removal algorithm, and effectively removes snow noise points in the LiDAR point cloud data while reducing the loss of non-snow noise points in the point cloud data, thereby reducing the damage to the environmental information in the point cloud data. It can adapt to snowfall of various densities.

[0014] Further, in step S1, the candidate snow noise points are obtained by filtering the actual point cloud data according to the dynamic intensity threshold curve of snowflakes using an adaptive intensity filter, thereby obtaining the candidate snow noise points. The method for obtaining the dynamic intensity threshold curve of snowflakes is as follows: based on the intensity principle of the laser radar, the laser radar intensity curve of the laser pulse under ideal conditions is obtained; based on the physical model of snowflakes and the point cloud intensity characteristics of snowflakes, the dynamic intensity threshold curve of snowflakes is calculated according to the laser radar intensity curve;

[0015] The formula for the laser radar intensity curve is:

[0016]

[0017] Where I is the reflection intensity, R o is the distance between the ideal reference point and the laser radar, I o is the reflection intensity of the ideal reference point, and R is the distance between the target object and the lidar.

[0018] Furthermore, the laser radar intensity curve and the dynamic intensity threshold curve of snowflakes are both calculated under ideal conditions, where the ideal conditions refer to the conditions where the laser pulse is vertically incident and completely reflected.

[0019] Furthermore, the method of filtering the actual point cloud data by the adaptive intensity filter in step S1 is:

[0020] For the point cloud data at a certain moment, we first calculate the distance from each point in the point cloud data to the laser radar, and then obtain the snowflake intensity threshold corresponding to the distance based on the value of the distance and the dynamic intensity threshold curve of the snowflake. If the intensity of the laser point is lower than the threshold, it will be marked as a candidate snow noise point.

[0021] Furthermore, the method for calculating the number of intensity neighborhoods of the candidate snow noise point in the point cloud image at the current moment in step S2 is:

[0022] Using the radius filtering method, for each candidate snow noise point, the number of intensity neighbors within the radius defined in its point cloud data is calculated.

[0023] Furthermore, the method for extracting the width-time image in step S3 is: performing dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames by principal component analysis, projecting the adjacent point cloud frames into a unified coordinate using the same transformation matrix to obtain a reduced dimensionality graph, combining multiple consecutive frames of reduced dimensionality graphs into three-dimensional data with width-height-time as dimensions, fixing the height, and extracting the width-time image.

[0024] Furthermore, the method for calculating the number of spatiotemporal neighborhoods in step S4 is:

[0025] On the width-time image, the neighborhood distribution of each candidate snow noise point in all directions is counted as the number of spatiotemporal neighbors of the candidate snow noise point in the width-time space.

[0026] Furthermore, in step S5, the comprehensive scoring method based on the entropy weight method is:

[0027] The number of intensity neighbors and the number of spatiotemporal neighbors are fused by the entropy weight method, the weight of the number of intensity neighbors and the weight of the number of spatiotemporal neighbors are calculated using the entropy information of the number of intensity neighbors and the number of spatiotemporal neighbors, and the weighted sum is performed using the weights to obtain a comprehensive score;

[0028] The weight calculation method is as follows: the number of candidate snow noise points is n, and the candidate snow noise points have two evaluation indicators: the number of intensity neighbors and the number of spatiotemporal neighbors. The candidate snow noise points are represented as an n*2 matrix F n*2 , where the first column is the number of intensity neighbors of n candidate snow noise points, and the second column is the number of spatiotemporal neighbors of n candidate snow noise points, f ij is the element in the i-th row and j-th column of the matrix F;

[0029] The evaluation index of each candidate snow noise point is normalized, and the formula is:

[0030]

[0031] Then calculate the corresponding entropy value. The entropy value calculation formula for the jth indicator is:

[0032]

[0033] The corresponding weight is calculated according to the calculated entropy value. The weight calculation formula of the jth indicator is as follows:

[0034]

[0035] Thus, the weights w1 and w2 of the candidate snow noise point as the real snow noise point are obtained, where w1 is the weight corresponding to the number of intensity neighbors of the candidate snow noise point, and w2 is the weight corresponding to the number of spatiotemporal neighbors of the candidate snow noise point.

[0036] A laser radar snow removal system integrating intensity and spatiotemporal geometric features, comprising: a laser radar, an adaptive intensity filter, an intensity neighborhood operation unit, a dimensionality reduction processing unit, a width-time image extraction unit, a spatiotemporal neighborhood extraction unit, a comprehensive score calculation unit, and a filtering and repairing unit;

[0037] The LiDAR snow removal system obtains the actual point cloud data of the LiDAR, and the adaptive intensity filter filters the actual point cloud data to obtain candidate snow noise points; the intensity neighborhood operation unit calculates the number of intensity neighbors of the candidate snow noise points in the point cloud frame at the current moment based on the radius filtering method and the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data; the dimensionality reduction processing unit performs dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames through the principal component analysis method, and the adjacent point cloud frames refer to the points before and after the point cloud frame with the candidate snow noise point. For several frames, the width-time image extraction unit extracts the width-time image by using the dimension reduction data; the spatiotemporal neighborhood extraction unit calculates the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image based on the motion law characteristics of snowflakes and the width-time image; the comprehensive score calculation unit calculates the comprehensive score by using the number of intensity neighbors and the number of spatiotemporal neighbors based on the entropy weight method, and then classifies the candidate snow noise points into real snow noise points and non-noise points according to the comprehensive score; the filtering and repairing unit removes the real snow noise points to achieve filtering, and restores the non-noise points to ordinary point cloud points to achieve repair.

[0038] Further, the radar intensity curve generating unit obtains the laser radar intensity curve of the laser pulse under ideal conditions based on the intensity principle of the laser radar; the dynamic intensity threshold curve generating unit calculates the dynamic intensity threshold curve of the snowflake based on the physical model of the snowflake and the point cloud intensity characteristics of the snowflake according to the laser radar intensity curve;

[0039] The formula for the laser radar intensity curve is:

[0040]

[0041] Where I is the reflection intensity, R o is the distance between the ideal reference point and the laser radar, I o is the reflection intensity of the ideal reference point, and R is the distance between the target object and the lidar.

[0042] This technical solution proposes a LiDAR snow removal method and system that integrates intensity and spatiotemporal geometric features, combines the reflection intensity information of the intensity neighborhood number and the spatial feature information of the spatiotemporal neighborhood number to remove snow from the point cloud data, improves the accuracy and adaptability of the snow removal algorithm, and effectively removes snow noise points in the LiDAR point cloud data while reducing the loss of non-snow noise points in the point cloud data, thereby reducing the damage to the environmental information in the point cloud data. It can adapt to snowfall of various densities. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flowchart of the LiDAR snow removal method that integrates intensity and spatiotemporal geometry features;

[0044] Figure 2 Schematic diagram for extracting width-time image;

[0045] Figure 3 Schematic diagram of the snowflake physics model. DETAILED DESCRIPTION

[0046] In order to clearly illustrate the laser radar snow removal method and system that integrates intensity and spatiotemporal geometric characteristics of the present invention, the present invention is further described in conjunction with embodiments and drawings, but this should not limit the scope of protection of the present invention.

[0047] Example 1

[0048] A LiDAR snow removal method that integrates intensity and spatiotemporal geometric features. The flowchart is as follows: Figure 1 As shown, the steps include:

[0049] S1, obtaining actual point cloud data of the laser radar, filtering the actual point cloud data, and obtaining candidate snow noise points;

[0050] S2, according to the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data, calculate the number of intensity neighborhoods of the candidate snow noise point in the point cloud frame at the current moment;

[0051] S3, performing dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames by principal component analysis to extract a width-time image, wherein the adjacent point cloud frames refer to several frames before and after the point cloud frame with the candidate snow noise point;

[0052] S4, based on the motion law characteristics of snowflakes and the width-time image, calculating the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image;

[0053] S5, based on the entropy weight method, the comprehensive score is calculated using the number of intensity neighbors and the number of spatiotemporal neighbors, and then the candidate snow noise points are finally classified into real snow noise points and non-noise points according to the comprehensive score;

[0054] S6. Remove real snow noise points to achieve filtering, and restore non-noise points to ordinary point cloud points to achieve repair.

[0055] In practical applications, the steps of calculating the number of intensity neighbors and the number of spatiotemporal neighbors are not in order and can be processed and obtained simultaneously.

[0056] This embodiment combines the reflection intensity information of the intensity neighborhood number and the spatial feature information of the spatiotemporal neighborhood number to perform snow removal processing on the point cloud data, thereby improving the accuracy and adaptability of the snow removal algorithm, effectively removing snow noise points in the lidar point cloud data, and reducing the loss of non-snow noise points in the point cloud data, thereby reducing the damage to the environmental information in the point cloud data. It can adapt to snowfall of various densities.

[0057] Example 2

[0058] A LiDAR snow removal method that integrates intensity and spatiotemporal geometric features. The flowchart is as follows: Figure 1 As shown, the steps include:

[0059] S1, obtaining actual point cloud data of the laser radar, filtering the actual point cloud data using an adaptive intensity filter, and obtaining candidate snow noise points;

[0060] The method for filtering point cloud data by the adaptive intensity filter is:

[0061] For the point cloud data at a certain moment, we first calculate the distance from each point in the point cloud data to the laser radar, and then obtain the snowflake intensity threshold corresponding to the distance based on the value of the distance and the dynamic intensity threshold curve of the snowflake. If the intensity of the laser point is lower than the threshold, it will be marked as a candidate snow noise point.

[0062] The intensity filter screens the point cloud data points based on the intensity of each point in the point cloud data to obtain candidate snow noise points.

[0063] The method for obtaining the dynamic intensity threshold curve of the snowflake is as follows: based on the intensity principle of the laser radar, the laser radar intensity curve of the laser pulse under ideal conditions is obtained; based on the physical model of the snowflake and the point cloud intensity characteristics of the snowflake, the dynamic intensity threshold curve of the snowflake is calculated according to the laser radar intensity curve;

[0064] The formula for the laser radar intensity curve is:

[0065]

[0066] Where I is the reflection intensity, R o is the distance between the ideal reference point and the laser radar, I o is the reflection intensity of the ideal reference point, and R is the distance between the target object and the lidar.

[0067] The above formula is determined based on the following principle: the echo power P of the laser pulse emitted by the laser beam transmitter after hitting the target object is inversely proportional to the square of the distance R. Since the reflection intensity I is proportional to the echo power P, the ideal reference point o is selected.

[0068] The laser radar refers to a multi-line laser radar sensor, which is used to obtain three-dimensional information of the surrounding environment. Laser pulses are emitted from different angles by rotating multiple laser beam transmitters placed in the vertical direction. After receiving the reflected laser, the detection distance from the sensor to the target point is calculated by the time difference from emission to reception, so that the three-dimensional information and reflection intensity information of the surrounding environment can be obtained and stored in the form of point cloud.

[0069] Based on the physical model of snowflakes and the point cloud intensity characteristics of snowflakes, according to the laser radar intensity curve, the dynamic intensity threshold curve of snowflakes under ideal conditions is calculated, and the ideal condition refers to the condition where the laser pulse is vertically incident and completely reflected;

[0070] The dynamic intensity threshold curve of the snowflake is calculated under the ideal condition of vertical incidence and complete reflection of the laser pulse. Figure 3 As shown in the figure, according to different snowfall levels such as heavy snow, moderate snow, and light snow, snowflakes are usually formed by the aggregation of 2 to 3 snow crystals, and the maximum radius of the snow crystal is 2.8 mm, so the diameter of the snowflake is agreed to be 1.12 cm. Snowflakes are spherical, so the incident angle of the laser pulse when it hits the snowflake is the same as the average incident angle of sunlight when it hits the earth, which is 45°.

[0071] S2, based on the radius filtering method, according to the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data, the number of intensity neighborhoods of the candidate snow noise point in the point cloud frame at the current moment is calculated;

[0072] The method for calculating the number of intensity neighbors of the candidate snow noise point in the point cloud image at the current moment is:

[0073] Using the radius filtering method, for each candidate snow noise point, the number of intensity neighbors within the radius defined in its point cloud data is calculated.

[0074] The radius filtering method is a dynamic radius filtering method that removes noise caused by snow. The neighboring points of each point cloud are searched in the point cloud dataset using the search radius specified by the user. If the number of searched points is greater than the set radius threshold, they are retained, otherwise they are removed. The search radius of the dynamic radius filtering method is dynamically adjusted based on the distance from the point cloud to the sensor.

[0075] S3, performing dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames by principal component analysis to extract a width-time image, wherein the adjacent point cloud frames refer to several frames before and after the point cloud frame with the candidate snow noise point;

[0076] The method for extracting the width-time image is as follows: performing dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames by the principal component analysis method, projecting the adjacent point cloud frames into a unified coordinate using the same transformation matrix to obtain a dimensionality reduction graph, combining multiple frames of the dimensionality reduction graph into three-dimensional data with width-height-time as the dimensions, fixing the height, and extracting the width-time image.

[0077] The dimension reduction is to use principal component analysis to reduce the dimension of the candidate snow noise point and its preceding and following adjacent frames.

[0078] Principal component analysis is a method for reducing the dimensionality of high-dimensional data. Based on this method, the original data is reduced in dimensionality and the environmental characteristic information contained in the original data is reduced.

[0079] S4, based on the motion law characteristics of snowflakes and the width-time image, calculating the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image;

[0080] The method for calculating the number of spatiotemporal neighbors is:

[0081] On the width-time image, the neighborhood distribution of each candidate snow noise point in each direction is counted as the number of spatiotemporal neighbors of the candidate snow noise point in the width-time space.

[0082] S5, based on the entropy weight method, the comprehensive score is calculated using the number of intensity neighbors and the number of spatiotemporal neighbors, and then the candidate snow noise points are finally classified into real snow noise points and non-noise points according to the comprehensive score;

[0083] The entropy weight method enables comprehensive evaluation to objectively and adaptively adjust the weights of various evaluation indicators according to different data. According to the basic principles of information theory, the smaller the information entropy of the sample data corresponding to the evaluation indicator, the more effective the impact on the comprehensive evaluation results will be. The entropy weight method calculates the entropy weight based on the data of the sample under each indicator, and eliminates the influence of indicators that have little impact on the evaluation results on the comprehensive evaluation by reducing the weight value corresponding to the indicator with a larger entropy value.

[0084] The comprehensive scoring method based on the entropy weight method is:

[0085] The number of intensity neighbors and the number of spatiotemporal neighbors are fused by the entropy weight method, the weight of the number of intensity neighbors and the weight of the number of spatiotemporal neighbors are calculated using the entropy information of the number of intensity neighbors and the number of spatiotemporal neighbors, and the weighted sum is performed using the weights to obtain a comprehensive score;

[0086] The weight calculation method is as follows: the number of candidate snow noise points is n, and the candidate snow noise points have two evaluation indicators: the number of intensity neighbors and the number of spatiotemporal neighbors. The candidate snow noise points are represented as an n*2 matrix F n*2 , where the first column is the number of intensity neighbors of n candidate snow noise points, and the second column is the number of spatiotemporal neighbors of n candidate snow noise points, f ij is the element in the i-th row and j-th column of the matrix F;

[0087] The evaluation index of each candidate snow noise point is normalized, and the formula is:

[0088]

[0089] Then calculate the corresponding entropy value. The entropy value calculation formula for the jth indicator is:

[0090]

[0091] The corresponding weight is calculated according to the calculated entropy value. The weight calculation formula of the jth indicator is as follows:

[0092]

[0093] Thus, the weights w1 and w2 of the candidate snow noise point as the real snow noise point are obtained, w1 is the weight corresponding to the number of intensity neighbors of the candidate snow noise point, and w2 is the weight corresponding to the number of spatiotemporal neighbors of the candidate snow noise point;

[0094] The weights w1 and w2 are used to calculate the comprehensive score of the candidate snow noise points. The calculation formula is:

[0095]

[0096] According to the comprehensive score, the classification result of the candidate snow noise point is obtained as a real snow noise point or a non-noise point;

[0097] S6. Remove real snow noise points to achieve filtering, and restore non-noise points to ordinary point cloud points to achieve repair.

[0098] In practical applications, the steps of calculating the number of intensity neighbors and the number of spatiotemporal neighbors are not in order and can be processed and obtained simultaneously.

[0099] Example 3

[0100] A laser radar snow removal system integrating intensity and spatiotemporal geometric features, comprising: a laser radar, an adaptive intensity filter, an intensity neighborhood operation unit, a dimensionality reduction processing unit, a width-time image extraction unit, a spatiotemporal neighborhood extraction unit, a comprehensive score calculation unit, and a filtering and repairing unit;

[0101] The LiDAR snow removal system obtains the actual point cloud data of the LiDAR, and the adaptive intensity filter filters the actual point cloud data to obtain candidate snow noise points; the intensity neighborhood operation unit calculates the number of intensity neighbors of the candidate snow noise points in the point cloud frame at the current moment based on the radius filtering method and the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data; the dimensionality reduction processing unit performs dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames through the principal component analysis method, and the adjacent point cloud frames refer to the points before and after the point cloud frame with the candidate snow noise point. For several frames, the width-time image extraction unit extracts the width-time image by using the dimension reduction data; the spatiotemporal neighborhood extraction unit calculates the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image based on the motion law characteristics of snowflakes and the width-time image; the comprehensive score calculation unit calculates the comprehensive score by using the number of intensity neighbors and the number of spatiotemporal neighbors based on the entropy weight method, and then classifies the candidate snow noise points into real snow noise points and non-noise points according to the comprehensive score; the filtering and repairing unit removes the real snow noise points to achieve filtering, and restores the non-noise points to ordinary point cloud points to achieve repair.

[0102] It also includes a radar intensity curve generating unit and a dynamic intensity threshold curve generating unit;

[0103] The radar intensity curve generation unit obtains the laser radar intensity curve of the laser pulse under ideal conditions based on the intensity principle of the laser radar; the dynamic intensity threshold curve generation unit calculates the dynamic intensity threshold curve of the snowflake based on the physical model of the snowflake and the point cloud intensity characteristics of the snowflake according to the laser radar intensity curve;

[0104] The formula for the laser radar intensity curve is:

[0105]

[0106] Where I is the reflection intensity, R o is the distance between the ideal reference point and the laser radar, I o is the reflection intensity of the ideal reference point, and R is the distance between the target object and the lidar.

[0107] Example 4

[0108] A laser radar snow removal system integrating intensity and spatiotemporal geometric features, comprising: a laser radar, an adaptive intensity filter, an intensity neighborhood operation unit, a dimensionality reduction processing unit, a width-time image extraction unit, a spatiotemporal neighborhood extraction unit, a comprehensive score calculation unit, and a filtering and repairing unit;

[0109] The LiDAR snow removal system obtains the actual point cloud data of the LiDAR, and the adaptive intensity filter filters the actual point cloud data to obtain candidate snow noise points; the intensity neighborhood operation unit calculates the number of intensity neighbors of the candidate snow noise points in the point cloud frame at the current moment based on the radius filtering method and the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data; the dimensionality reduction processing unit performs dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames through the principal component analysis method, and the adjacent point cloud frames refer to the points before and after the point cloud frame with the candidate snow noise point. For several frames, the width-time image extraction unit extracts the width-time image by using the dimension reduction data; the spatiotemporal neighborhood extraction unit calculates the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image based on the motion law characteristics of snowflakes and the width-time image; the comprehensive score calculation unit calculates the comprehensive score by using the number of intensity neighbors and the number of spatiotemporal neighbors based on the entropy weight method, and then classifies the candidate snow noise points into real snow noise points and non-noise points according to the comprehensive score; the filtering and repairing unit removes the real snow noise points to achieve filtering, and restores the non-noise points to ordinary point cloud points to achieve repair.

[0110] It also includes a radar intensity curve generating unit and a dynamic intensity threshold curve generating unit;

[0111] The radar intensity curve generation unit obtains the laser radar intensity curve of the laser pulse under ideal conditions based on the intensity principle of the laser radar; the dynamic intensity threshold curve generation unit calculates the dynamic intensity threshold curve of the snowflake based on the physical model of the snowflake and the point cloud intensity characteristics of the snowflake according to the laser radar intensity curve;

[0112] The formula for the laser radar intensity curve is:

[0113]

[0114] Where I is the reflection intensity, R o is the distance between the ideal reference point and the laser radar, I o is the reflection intensity of the ideal reference point, and R is the distance between the target object and the lidar.

[0115] The laser radar intensity curve calculated by the radar intensity curve generating unit and the dynamic intensity threshold curve of the snowflake calculated by the dynamic intensity threshold curve generating unit are both calculated under ideal conditions, and the ideal condition refers to the condition where the laser pulse is vertically incident and completely reflected.

[0116] The method of filtering the actual point cloud data by the adaptive intensity filter is:

[0117] For the point cloud data at a certain moment, we first calculate the distance from each point in the point cloud data to the laser radar, and then obtain the snowflake intensity threshold corresponding to the distance based on the value of the distance and the dynamic intensity threshold curve of the snowflake. If the intensity of the laser point is lower than the threshold, it will be marked as a candidate snow noise point.

[0118] The method for the intensity neighborhood operation unit to calculate the number of intensity neighbors of the candidate snow noise point in the point cloud image at the current moment is:

[0119] Using the radius filtering method, for each candidate snow noise point, the number of intensity neighbors within the radius defined in its point cloud data is calculated.

[0120] The width-time image extraction unit extracts the width-time image by: performing dimensionality reduction on the point cloud frame with candidate snow noise points and its adjacent point cloud frames by principal component analysis, projecting the adjacent point cloud frames into a unified coordinate using the same transformation matrix to obtain a reduced dimensionality graph, combining multiple consecutive frames of reduced dimensionality graphs into three-dimensional data with width-height-time as dimensions, fixing the height, and extracting the width-time image.

[0121] The method for the spatiotemporal neighborhood extraction unit to calculate the number of spatiotemporal neighborhoods is:

[0122] On the width-time image, the neighborhood distribution of each candidate snow noise point in all directions is counted as the number of spatiotemporal neighbors of the candidate snow noise point in the width-time space.

[0123] The comprehensive score calculation unit calculates the comprehensive score based on the entropy weight method as follows:

[0124] The number of intensity neighbors and the number of spatiotemporal neighbors are fused by the entropy weight method, the weight of the number of intensity neighbors and the weight of the number of spatiotemporal neighbors are calculated using the entropy information of the number of intensity neighbors and the number of spatiotemporal neighbors, and the weighted sum is performed using the weights to obtain a comprehensive score;

[0125] The weight calculation method is as follows: the number of candidate snow noise points is n, and the candidate snow noise points have two evaluation indicators: the number of intensity neighbors and the number of spatiotemporal neighbors. The candidate snow noise points are represented as an n*2 matrix F n*2 , where the first column is the number of intensity neighbors of n candidate snow noise points, and the second column is the number of spatiotemporal neighbors of n candidate snow noise points, f ij is the element in the i-th row and j-th column of the matrix F;

[0126] The evaluation index of each candidate snow noise point is normalized, and the formula is:

[0127]

[0128] Then calculate the corresponding entropy value. The entropy value calculation formula for the jth indicator is:

[0129]

[0130] The corresponding weight is calculated according to the calculated entropy value. The weight calculation formula of the jth indicator is as follows:

[0131]

[0132] Thus, the weights w1 and w2 of the candidate snow noise point as the real snow noise point are obtained, where w1 is the weight corresponding to the number of intensity neighbors of the candidate snow noise point, and w2 is the weight corresponding to the number of spatiotemporal neighbors of the candidate snow noise point.

[0133] In practical applications, the steps of calculating the number of intensity neighbors and the number of spatiotemporal neighbors are not in order and can be processed and obtained simultaneously.

[0134] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that deviate from the spirit and principles of the present invention shall be equivalent replacement methods and shall be included in the protection scope of the present invention.

Claims

1. A LiDAR snow removal method integrating intensity and spatiotemporal geometric features, characterized in that: The method comprises the steps of: S1, obtaining actual point cloud data of the laser radar, filtering the actual point cloud data, and obtaining candidate snow noise points; S2, according to the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data, calculate the number of intensity neighborhoods of the candidate snow noise point in the point cloud frame at the current moment; S3, performing dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames by principal component analysis to extract a width-time image, wherein the adjacent point cloud frames refer to several frames before and after the point cloud frame with the candidate snow noise point; S4, based on the motion law characteristics of snowflakes and the width-time image, calculating the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image; S5, based on the entropy weight method, the comprehensive score is calculated using the number of intensity neighbors and the number of spatiotemporal neighbors, and then the candidate snow noise points are finally classified into real snow noise points and non-noise points according to the comprehensive score; S6. Remove real snow noise points to achieve filtering, and restore non-noise points to ordinary point cloud points to achieve repair.

2. The method for snow removal by laser radar integrating intensity and spatiotemporal geometric features according to claim 1, characterized in that: Step S1 obtains the candidate snow noise point by filtering the actual point cloud data according to the dynamic intensity threshold curve of snowflakes using an adaptive intensity filter, thereby obtaining the candidate snow noise point. The method for obtaining the dynamic intensity threshold curve of snowflakes is as follows: based on the intensity principle of laser radar, the laser radar intensity curve of the laser pulse under ideal conditions is obtained; based on the physical model of snowflakes and the point cloud intensity characteristics of snowflakes, the dynamic intensity threshold curve of snowflakes is calculated according to the laser radar intensity curve; The formula for the laser radar intensity curve is: Where I is the reflection intensity, R o is the distance between the ideal reference point and the laser radar, I o is the reflection intensity of the ideal reference point, and R is the distance between the target object and the lidar.

3. The method for snow removal by laser radar integrating intensity and spatiotemporal geometric features according to claim 2, characterized in that: The laser radar intensity curve and the dynamic intensity threshold curve of snowflakes are both calculated under ideal conditions, where the ideal conditions refer to the conditions where the laser pulse is vertically incident and completely reflected.

4. The method for snow removal by laser radar integrating intensity and spatiotemporal geometric features according to claim 2, characterized in that: The method of filtering the actual point cloud data by the adaptive intensity filter in step S1 is: For the point cloud data at a certain moment, we first calculate the distance from each point in the point cloud data to the laser radar, and then obtain the snowflake intensity threshold corresponding to the distance based on the value of the distance and the dynamic intensity threshold curve of the snowflake. If the intensity of the laser point is lower than the threshold, it will be marked as a candidate snow noise point.

5. The method for snow removal by laser radar integrating intensity and spatiotemporal geometric features according to claim 1, characterized in that: The method for calculating the number of intensity neighborhoods of the candidate snow noise point in the point cloud image at the current moment in step S2 is: Using the radius filtering method, for each candidate snow noise point, the number of intensity neighbors within the radius defined in its point cloud data is calculated.

6. The method for snow removal by laser radar integrating intensity and spatiotemporal geometric features according to claim 1, characterized in that: The method for extracting the width-time image described in step S3 is: reducing the dimension of the point cloud frame with candidate snow noise points and its adjacent point cloud frames by principal component analysis, projecting the adjacent point cloud frames into a unified coordinate using the same transformation matrix to obtain a reduced dimension graph, combining multiple frames of reduced dimension graphs into three-dimensional data with width-height-time as dimensions, fixing the height, and extracting the width-time image.

7. The method for snow removal by laser radar integrating intensity and spatiotemporal geometric features according to claim 5, characterized in that: The method for calculating the number of spatiotemporal neighborhoods in step S4 is: On the width-time image, the neighborhood distribution of each candidate snow noise point in all directions is counted as the number of spatiotemporal neighbors of the candidate snow noise point in the width-time space.

8. The method for snow removal by laser radar integrating intensity and spatiotemporal geometric features according to claim 1, characterized in that: Step S5 calculates the comprehensive score based on the entropy weight method as follows: The number of intensity neighbors and the number of spatiotemporal neighbors are fused by the entropy weight method, the weight of the number of intensity neighbors and the weight of the number of spatiotemporal neighbors are calculated using the entropy information of the number of intensity neighbors and the number of spatiotemporal neighbors, and the weighted sum is performed using the weights to obtain a comprehensive score; The weight calculation method is as follows: the number of candidate snow noise points is n, and the candidate snow noise points have two evaluation indicators: the number of intensity neighbors and the number of spatiotemporal neighbors. The candidate snow noise points are represented as an n*2 matrix F n*2 , where the first column is the number of intensity neighbors of n candidate snow noise points, and the second column is the number of spatiotemporal neighbors of n candidate snow noise points, f ij is the element in the i-th row and j-th column of the matrix F; The evaluation index of each candidate snow noise point is normalized, and the formula is: Then calculate the corresponding entropy value. The entropy value calculation formula for the jth indicator is: The corresponding weight is calculated according to the calculated entropy value. The weight calculation formula of the jth indicator is as follows: Thus, the weights w1 and w2 of the candidate snow noise point as the real snow noise point are obtained, where w1 is the weight corresponding to the number of intensity neighbors of the candidate snow noise point, and w2 is the weight corresponding to the number of spatiotemporal neighbors of the candidate snow noise point.

9. A laser radar snow removal system integrating intensity and spatiotemporal geometric features, which implements the laser radar snow removal method integrating intensity and spatiotemporal geometric features as described in any one of claims 1 to 8, characterized in that: include: Laser radar, adaptive intensity filter, intensity neighborhood operation unit, dimension reduction processing unit, width-time image extraction unit, spatiotemporal neighborhood extraction unit, comprehensive score calculation unit, filtering and repairing unit; The LiDAR snow removal system obtains the actual point cloud data of the LiDAR, and the adaptive intensity filter filters the actual point cloud data to obtain candidate snow noise points; the intensity neighborhood operation unit calculates the number of intensity neighbors of the candidate snow noise points in the point cloud frame at the current moment based on the radius filtering method and the point cloud distribution density characteristics around the snowflakes in the actual single-frame point cloud frame of the point cloud data; the dimensionality reduction processing unit performs dimensionality reduction on the point cloud frame with the candidate snow noise point and its adjacent point cloud frames through the principal component analysis method, and the adjacent point cloud frames refer to the points before and after the point cloud frame with the candidate snow noise point. For several frames, the width-time image extraction unit extracts the width-time image by using the dimension reduction data; the spatiotemporal neighborhood extraction unit calculates the number of spatiotemporal neighbors of the candidate snow noise point on the width-time image based on the motion law characteristics of snowflakes and the width-time image; the comprehensive score calculation unit calculates the comprehensive score by using the number of intensity neighbors and the number of spatiotemporal neighbors based on the entropy weight method, and then classifies the candidate snow noise points into real snow noise points and non-noise points according to the comprehensive score; the filtering and repairing unit removes the real snow noise points to achieve filtering, and restores the non-noise points to ordinary point cloud points to achieve repair.

10. The laser radar snow removal system according to claim 9, characterized in that: It also includes a radar intensity curve generating unit and a dynamic intensity threshold curve generating unit; The radar intensity curve generation unit obtains the laser radar intensity curve of the laser pulse under ideal conditions based on the intensity principle of the laser radar; the dynamic intensity threshold curve generation unit calculates the dynamic intensity threshold curve of the snowflake based on the physical model of the snowflake and the point cloud intensity characteristics of the snowflake according to the laser radar intensity curve; The formula for the laser radar intensity curve is: Where I is the reflection intensity, R o is the distance between the ideal reference point and the laser radar, I o is the reflection intensity of the ideal reference point, and R is the distance between the target object and the lidar.

Images