A method for suppressing false echoes of laser radar in a high reflection environment

Abstract

The application discloses a method for suppressing false echoes of a laser radar in a high-reflectivity environment, comprising the steps of collecting environment point cloud data and environment image data by using a laser radar and a binocular camera, respectively, and performing downsampling and analysis on a current frame of the laser radar; judging the current frame and a previous frame based on a time difference, and determining whether to continue collecting or performing fusion judgment according to a judgment result; and fusing point cloud data of the laser radar and the binocular camera to suppress false echoes. The application combines data of the laser radar and the binocular camera, suppresses false echoes while retaining key features, realizes smooth transition of a downsampling rate by using a Gaussian weight model in a corner point area, maximizes feature retention, removes points in the radar point cloud that are inconsistent with binocular point cloud through spatial consistency verification of the radar point cloud and the binocular point cloud, and finally outputs high-quality environment perception point cloud, thereby effectively suppressing radar false echoes and sensor noise while retaining real geometric features.

Description

Technical Field

[0001] The present invention relates to the field of intelligent vehicle environment perception technology, and in particular to a method for suppressing false echoes of a laser radar in a highly reflective environment. Background Art

[0002] With the rapid development of artificial intelligence and intelligent transportation, environmental perception technology plays a vital role in the autonomous driving of intelligent vehicles. In order to accurately perceive the environment around the vehicle, sensor fusion technology has gradually become an important method to solve the limitations of single sensors. Among them, lidar has been widely used in environmental perception due to its high precision and strong anti-interference ability. However, in complex indoor and outdoor environments, lidar still faces some difficult-to-solve problems.

[0003] Existing LiDAR false echo suppression methods primarily rely on gain reduction and rain and snow interference suppression. While gain reduction is a common method, it can lead to feature loss in rainy and snowy conditions. Rain and snow interference suppression, on the other hand, is ineffective in adverse weather conditions. This is particularly true in complex indoor and outdoor scenarios with highly reflective surfaces. Rain and snow interference suppression is ineffective in these situations, while gain reduction inevitably results in feature loss. The higher the reflectivity, the more features are lost by this suppression method. In some cases, such as indoor positioning and navigation for robots, feature loss exceeding a certain threshold is unacceptable. Therefore, effectively suppressing false echoes from multiple radar reflections while maintaining feature integrity remains a pressing technical challenge. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for suppressing false echoes of a laser radar in a high-reflection environment, so as to solve the problems raised in the above-mentioned background technology.

[0005] The specific technical solution provided by the present invention is as follows: A method for suppressing false echoes of a laser radar in a highly reflective environment, comprising the following steps:

[0006] Step S1: Use the laser radar and binocular camera to collect environmental point cloud data and environmental image data respectively, and downsample and analyze the current frame of the laser radar;

[0007] Preferably, step S11: using the laser radar to collect the original point cloud data, traversing the point cloud through the threshold filter model, and calculating the filtered laser radar current frame ;

[0008] Step S12: Use binocular cameras to collect environmental image data, and use SGBM to perform three-dimensional reconstruction based on the distance between the binocular cameras to obtain the current frame based on the camera collection. ;

[0009] Step S13: Use SUSAN algorithm to analyze the current frame Extract corner points to form a corner point set ;

[0010] Step S14: Check the current frame of the laser radar Perform downsampling processing to obtain the downsampled point cloud frame ;

[0011] Step S15: downsample the point cloud frame Perform preprocessing to obtain the preprocessed point cloud frame .

[0012] Step S2: judging the current frame and the previous frame based on the time difference, and deciding whether to continue acquisition or perform fusion judgment according to the judgment result;

[0013] Preferably, by calculating the time difference of consecutive frames , , is the timestamp of the current frame, is the timestamp of the previous frame; if the time difference between the current frame and the previous frame When the time difference between the current frame and the previous frame is less than the set time difference threshold, return to the acquisition step. When the time difference is greater than or equal to the set threshold, the features are fused.

[0014] Step S3: Fuse the point cloud data of the lidar and binocular camera to suppress false echoes.

[0015] Preferably, step S31: the pre-processed point cloud frame As the current frame, take the previous n frames for correlation processing;

[0016] Step S32: The current frame Based on the current frame captured by the camera Fusion is performed to retain points in the radar point cloud that are consistent with the binocular camera point cloud, suppressing false echoes from inconsistent sensors;

[0017] Step S33: From the current frame captured by the camera Extract all points of the face closest to the origin and form a set containing all points ;

[0018] Step S34: For the current frame All points in the radar point cloud are judged and the points on the same ray but with a distance less than the set are removed. The point with the minimum distance is used to remove the false echo point;

[0019] Step S35: fusion of the current frame Filtering and smoothing are performed to obtain the final environmental perception result.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) The present invention overcomes the limitations of a single sensor in the existing technology by introducing the fusion of a binocular camera and a lidar multi-sensor, thereby improving the comprehensiveness and accuracy of environmental perception. In complex indoor and outdoor scenes, the suppression method combined with the time factor can more effectively identify and locate targets, solving the problem of feature loss in high-reflection environments caused by traditional gain reduction and rain and snow interference suppression methods.

[0022] (2) The present invention uses a binocular camera to acquire images and convert them into point cloud data frames. Combined with lidar data, it effectively removes false echoes through coordinate matching and analysis, avoiding the feature loss caused by reducing gain in traditional methods, while maintaining good penetration in rainy and snowy weather.

[0023] (3) The present invention improves the efficiency and accuracy of point cloud dedistortion, ground point segmentation, and lidar degradation detection through an improved algorithm in the data preprocessing stage, providing a more reliable basis for subsequent feature extraction and matching.

[0024] (4) The present invention achieves refined processing of highly reflective environments by establishing the relationship between a point and its adjacent pixel positions, effectively solving the problem of insufficient efficiency and accuracy in calculating perspective transformation relationships in the prior art, and improving the accuracy of sensor calibration.

[0025] (5) The method of the present invention not only ensures the suppression of false echoes but also maintains the original features to the greatest extent, which not only improves the accuracy of environmental perception but also provides more reliable technical support for application scenarios such as robot indoor positioning and navigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flowchart of the steps of a method for suppressing false echoes of a lidar in a high-reflection environment provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention are within the scope of protection of the present invention.

[0028] Example 1:

[0029] like Figure 1 As shown, the method for suppressing false echoes of a laser radar in a high-reflection environment described in an embodiment of the present invention includes the following steps:

[0030] Step S1: Use the laser radar and binocular camera to collect environmental point cloud data and environmental image data respectively, and downsample and analyze the current frame of the laser radar;

[0031] S11: Use the lidar to collect raw point cloud data, use the threshold filter model to traverse the point cloud, and calculate the filtered lidar current frame ;

[0032] In this embodiment, the original point cloud data is collected by laser radar. In complex environments, false echoes are often accompanied by abnormal intensity, but the real features including corner points usually have stable intensity. Therefore, the threshold filter model is used to traverse the point cloud and calculate the filtered laser radar current frame. , the calculation formula is:

[0033]

[0034] in, is the original lidar point cloud data, is the point contained in the original lidar point cloud data, each point , for point The reflection intensity, , The reflection intensity threshold is set, which is a smaller value preset based on the parameters and noise model of the laser radar. In this embodiment, it is set to 0.1. The first filtering is performed based on the reflection intensity threshold, that is, the reflection intensity threshold is used to determine whether the reflection intensity of each point is within the set threshold range. When it is lower than the intensity threshold, the point is regarded as noise and removed. While removing low-intensity noise points, high-reflection features are retained, solving the problem of feature loss under highly reflective surfaces. At the same time, because the reflection intensity threshold in this embodiment is a preset smaller value, close to the sensor noise, it avoids discarding potential features too early.

[0035] S12: Synchronously use binocular cameras to collect environmental image data, and use SGBM to perform 3D reconstruction based on the distance between the binocular cameras to obtain the current frame based on camera collection ;

[0036] In this embodiment, the left and right environmental images are captured by binocular cameras, the captured left and right environmental images are input through the stereo matching algorithm SGBM, the disparity map is analyzed and output, the disparity map is optimized through the cost aggregation function and dynamic programming, and the binocular vision calculates the depth value based on the disparity And obtain the 3D coordinates after 3D reconstruction, specifically:

[0037]

[0038] in, is the camera focal length obtained from the binocular camera intrinsic parameter matrix, is the preset baseline distance of the binocular camera, is the disparity value calculated by stereo matching, is the 3D coordinate after 3D reconstruction, which constitutes the current frame acquired by the camera after conversion ,SGBM algorithm is highly robust in texture-rich areas, suitable for complex indoor and outdoor scenes, and solves the shortcomings of lidar in low-reflection areas.

[0039] S13: Use SUSAN algorithm to analyze the current frame Extract corner points to form a corner point set ;

[0040] In this embodiment, the brightness of the corner point changes significantly in the neighborhood, so the SUSAN algorithm is used to calculate the response value of each pixel. , the calculation formula is:

[0041]

[0042] in, is the pixel coordinate currently being processed, are the pixel coordinates in the neighborhood, Pixels The central area, Pixel The brightness value, Pixel The brightness value, is the brightness difference threshold set, Similarity measurement for smooth transition, detecting stable corners by brightness similarity, that is, using exponential function to evaluate neighborhood brightness consistency, when the pixel response value Below the set threshold Mark it as a corner point, output a binary corner point map, and get a corner point set containing all marked corner points , ;

[0043] S14: Check the current frame of the laser radar Perform downsampling processing to obtain the downsampled point cloud frame ;

[0044] In this embodiment, the corner point set The coordinates of all corner points in the image are converted from the coordinate system of the binocular camera to the coordinate system of the lidar to obtain the set of corner points in the lidar coordinate system. , and in the radar coordinate system, define the corner point set Any corner point is the center and the radius is The area range, for each laser radar current frame obtained Construct KD tree and calculate Point in To corner point set The distance to the nearest corner , the calculation formula is: , The point to query for the KD tree To corner point set The nearest corner point in judge Point in At a radius of area, when hour, Point in Located in the characteristic area of ​​the region, when hour, Point in In the edge area of ​​the region, dynamic sampling rate calculation is performed on the feature area and edge area respectively, specifically:

[0045]

[0046]

[0047] in, for point The influence weight of is the preset standard deviation parameter, for point To corner point set The distance to the nearest corner, For the current point Dynamic sampling rate, , is the preset sampling rate, that is, the points in the feature area At a radius of Within the region, the closer the feature area is to the center of the region, the higher the dynamic sampling rate is, and the closer the feature area is to the edge of the region, the lower the dynamic sampling rate is. The downsampling rate of the feature area within the area of ​​​​the radius is the lowest, and more points can be retained. The closer to the corner point, the lowest downsampling rate. After downsampling, the recognition features of the point cloud image are maximized. When the area is within the range of , downsampling is performed according to the preset sampling rate, and the recognition features of the point cloud are retained to the maximum extent after downsampling;

[0048] For example, for each point Use random downsampling to generate a random number and set the probability value of each point being retained. When the random number is less than the set probability value of being retained, the point is retained. , when the random number is greater than or equal to the set probability value of being retained, it is discarded, and the sampling rate of each point is adapted, and all the retained points form the downsampled point cloud frame .

[0049] Step S15: downsample the point cloud frame Perform preprocessing to obtain the preprocessed point cloud frame .

[0050] In this embodiment, preprocessing includes point cloud dedistortion, ground point segmentation and lidar degradation detection. Motion in a dynamic environment will cause point cloud distortion. Therefore, a motion compensation model is used to estimate the pose change of each point based on linear interpolation to correct the point position. The RANSAC algorithm is used to fit the plane using random sampling points, and the proportion of ground points is iteratively optimized to remove irrelevant points. The eigenvalue of the local point cloud covariance matrix is ​​calculated by using the eigenvalue analysis model, and then the eigenvalue is compared with the set threshold. When it is less than the set threshold, it indicates geometric degradation. Degradation detection avoids erroneous fusion when features are sparse.

[0051] Step S2: judging the current frame and the previous frame based on the time difference, and deciding whether to continue acquisition or perform fusion judgment according to the judgment result;

[0052] In this embodiment, the time difference between consecutive frames is calculated , the calculation formula is: , is the timestamp of the current frame, is the timestamp of the previous frame; when the time difference between the current frame and the previous frame When the time difference between the current frame and the previous frame is less than the set time difference threshold, return to the acquisition step. When the time difference is greater than or equal to the set threshold, the features are fused.

[0053] Step S3: Fuse the radar and camera point cloud data to suppress false echoes.

[0054] Step S31: pre-processed point cloud frame As the current frame, take the previous n frames for correlation processing;

[0055] In this embodiment, since false echoes or noise are not consistent in consecutive frames, but real features are relatively stable, the present invention combines the NDT algorithm and the GICP algorithm to represent the point cloud as a Gaussian distribution, optimizes the pose transformation to maximize the likelihood, and adds covariance constraints. After matching, points that have no neighboring points in the historical frame are removed to improve the matching robustness. The first n frames of data are taken from the cache queue to form a cache set. ,in is the depth of the historical frame, , according to the depth of the historical frame of the radar movement speed Adjust the speed. The faster the speed, the greater the difference between the historical frame and the current frame. Therefore, the depth of the historical frame is adjusted. The smaller the selection;

[0056] For example, the NDT algorithm is used for rough matching, and the pre-processed point cloud frame As the target point cloud, the historical frame As the source point cloud, the target point cloud is divided into grids through the NDT algorithm. The point cloud in each grid is represented by a Gaussian distribution. The transformation parameters of the source point cloud are optimized to maximize the probability of projecting it onto the target grid, and an initial transformation matrix is ​​output. , the transformation matrix output by the NDT algorithm is used as the initial value of the GICP algorithm, the point cloud matching problem is transformed into the matching between probability distribution models, the optimal transformation is solved by minimizing the Mahalanobis distance between the two point clouds, and the optimized transformed matrix is ​​output , the historical frame After the transformation Matrix transformation to current frame In the coordinate system, get the aligned historical frame point cloud; traverse the current frame For each point in the transformed historical frame point cloud, find the neighboring points of each point. The search range is centered at any point with a radius of When there is no adjacent point in the historical frame, that is, there is no point in the region in the historical frame, the point is marked as a candidate transient point. Remove all points marked as candidate transient points in the current frame to obtain a new current frame , further processing multiple historical frames, performing registration and neighboring point judgment on each historical frame, and removing points when there are no points at the corresponding positions in all historical frames. In this embodiment, through the spatiotemporal association between historical frames and current frames, transient interference such as flying birds and fallen leaves and false echoes of dynamic objects are suppressed, while retaining the static environment structure.

[0057] Step S32: Further convert the new current frame Based on the current frame captured by the camera Perform fusion to retain points in the radar point cloud that are consistent with the camera point cloud and suppress false echoes from inconsistent sensors;

[0058] In this embodiment, all current frames collected by the camera are Convert the coordinate system of the binocular camera to the coordinate system of the lidar, and set the key parameters for compensating sensor calibration errors and measurement noise. , The absolute value of the calibration measurement error accuracy of the lidar, in the current frame converted to the lidar coordinate system In the current frame, search for a point m that matches the new current frame. The Euclidean distance of the points in is less than or equal to the key parameter , when there is at least one such point m, the point is retained, and when it is calculated that the point is consistent with the new current frame The Euclidean distance of the points in is greater than the key parameter When, from the current frame The point is removed from the image, and the spatial consistency verification between the radar point cloud and the binocular point cloud is performed to remove the points in the radar point cloud that are inconsistent with the binocular point cloud, retain the real environment points, and suppress the false echoes caused by inconsistent sensors.

[0059] Step S33: Based on the current frame captured by the camera Extract all points of the face closest to the origin and form a set containing all points ;

[0060] In this embodiment, in the camera coordinate system, the origin is the optical center of the camera. The surface closest to the origin is represented by selecting the point closest to the origin on each ray starting from the origin. The set of these points represents the forefront of all object surfaces in the scene. The point cloud is grouped according to the ray direction to form a direction vector pointing from the origin to the point. The direction space is discretized into several angle intervals. The point closest to the origin is selected as the point in each direction interval. An angle threshold for the direction interval is defined. When the angle between the direction vectors of two points relative to the origin is less than the threshold, the two points are in the same direction interval, that is, on the same ray.

[0061] For example, the current frame captured by the camera For each point in Represents the direction vector, and calculates the pitch angle of the direction vector and azimuth , specifically:

[0062]

[0063] Then press the direction space Divide into grids. For example, the azimuth angle is divided into 1 degree intervals, that is, 0°-360° is divided into 360 intervals, and the pitch angle is also divided into 1 degree intervals, and -90°-90° is divided into 180 intervals. Then the entire space is divided into 360×180 direction intervals, and each point is assigned to the corresponding direction interval. According to the pitch angle of each point and azimuth Calculate the interval index to which it belongs, traverse each direction interval, skip if there is no point in the interval, if there is only one point in the interval, directly add it to the nearest surface point set, if there are multiple points in the interval, compare the distances of these points, select the point with the smallest distance, add it to the nearest surface point set, and finally output the data containing the current frame The set of all nearest surface points in .

[0064] Step S34: For the current frame All points in the radar point cloud are judged and the points on the same ray but with a distance less than the set are removed. The point with the minimum distance is used to remove the false echo point;

[0065] In this embodiment, on a ray starting from the origin of the laser radar, the set of closest surface points of the binocular camera is Indicates the surface point of the object closest to the origin on the ray, and the set The points in the pitch angle and azimuth Binning, each box represents a ray direction, stores the nearest point and its distance in that direction, establishes a mapping from the angle box to the nearest point distance, and traverses the current frame For each point in , calculate the pitch angle of each point and azimuth And distance, discretize its angle into the corresponding angle box. When the angle box exists in the mapping from the angle box to the nearest point distance, it means there is a corresponding nearest surface point. Get the nearest surface point distance in that direction. When the points in the lidar point cloud are on the same ray and the distance from the origin is shorter than the nearest surface point set of the binocular camera If the corresponding point in is closer, the point is marked as a false echo and the corresponding point in the current frame is When the angle box does not exist, that is, there is no point reconstructed by the binocular camera in this direction, since the fields of view of the binocular camera and the lidar are not completely consistent, the points where the radar points are not in the camera's field of view are not processed in the present invention, that is, they are retained.

[0066] Step 405: fusion of the current frame Filtering and smoothing are performed to obtain the final environmental perception result.

[0067] In this embodiment, the current frame after fusion For each point in the image, the average distance to multiple nearest neighbors in the neighborhood is calculated, and the average distance and standard deviation of the entire point cloud are calculated. When the average distance is greater than the set threshold, the point is determined to be an outlier, and all outliers are removed. The surface is then smoothed using bilateral filtering to preserve the integrity of the geometric features. Finally, a high-quality environmental perception point cloud is output. While retaining the true geometric features, it effectively suppresses radar false echoes and sensor noise, providing a reliable data foundation for tasks such as positioning and navigation in autonomous driving.

[0068] It should be noted that, in the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0069] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for suppressing false echoes of a laser radar in a highly reflective environment, characterized by: The following steps are involved: Step S1: Use the laser radar and binocular camera to collect environmental point cloud data and environmental image data respectively, and downsample and analyze the current frame of the laser radar; The step S1 further comprises the following steps: Step S11: Use the laser radar to collect the original point cloud data, traverse the point cloud through the threshold filter model, and calculate the filtered laser radar current frame ; Step S12: Use binocular cameras to collect environmental image data, and use SGBM to perform three-dimensional reconstruction based on the distance between the binocular cameras to obtain the current frame based on the camera collection. ; Step S13: Use SUSAN algorithm to analyze the current frame Extract corner points to form a corner point set ; Step S14: Check the current frame of the laser radar Perform downsampling processing to obtain the downsampled point cloud frame ; Step S15: downsample the point cloud frame Perform preprocessing to obtain the preprocessed point cloud frame ; The current frame of the laser radar The downsampling process includes: The coordinates of all corner points in the image are converted from the coordinate system of the binocular camera to the coordinate system of the lidar to obtain the set of corner points in the lidar coordinate system. , and in the radar coordinate system, define the corner point set Any corner point is the center and the radius is The area range, for each laser radar current frame obtained Construct KD tree and calculate Point in To corner point set The distance to the nearest corner , the calculation formula is: , The point to query for the KD tree To corner point set The nearest corner point in ; According to this distance judge Point in At a radius of area, when hour, Point in Located in the characteristic area of ​​the regional range, when hour, Point in In the edge area of ​​the region, dynamic sampling rate calculation is performed on the feature area and edge area respectively, specifically: in, for point The influence weight of is the preset standard deviation parameter, for point To corner point set The distance to the nearest corner, For the current point Dynamic sampling rate, , is the preset sampling rate, that is, the points in the feature area At a radius of In the region of radius , the dynamic sampling rate of the feature area closer to the center of the region is higher, and the dynamic sampling rate of the feature area closer to the edge of the region is lower, so that in the region of radius The downsampling rate of the feature area within the region is the lowest; For each point Use random downsampling to generate a random number and set the probability value of each point being retained. When the random number is less than the set probability value of being retained, the point is retained. , when the random number is greater than or equal to the set probability value of being retained, it is discarded, and the sampling rate of each point is adapted, and all the retained points form the downsampled point cloud frame ; Step S2: judging the current frame and the previous frame based on the time difference, and deciding whether to continue acquisition or perform fusion judgment according to the judgment result; Step S3: Fuse the point cloud data of the lidar and binocular camera to suppress false echoes.

2. The method for suppressing false echoes of a laser radar in a high-reflection environment according to claim 1, characterized in that: The 3D reconstruction using SGBM includes: capturing left and right environmental images through binocular cameras, inputting the captured left and right environmental images using the stereo matching algorithm SGBM, analyzing and outputting disparity maps, optimizing the disparity maps through cost aggregation functions and dynamic programming, and calculating depth values ​​based on the disparity of binocular vision. And obtain the 3D coordinates after 3D reconstruction, specifically: in, is the camera focal length obtained from the binocular camera intrinsic parameter matrix, is the preset baseline distance of the binocular camera, is the disparity value calculated by stereo matching, is the 3D coordinate after 3D reconstruction, which constitutes the current frame acquired by the camera after conversion .

3. The method for suppressing false echoes of a laser radar in a high-reflection environment according to claim 1, characterized in that: The corner point set is formed Includes: Using the SUSAN algorithm to calculate the response value of each pixel , the calculation formula is: in, is the pixel coordinate currently being processed, are the pixel coordinates in the neighborhood, Pixels The central area, Pixel The brightness value, Pixel The brightness value, is the brightness difference threshold set, Similarity measurement for smooth transition, detecting stable corners by brightness similarity, that is, using exponential function to evaluate neighborhood brightness consistency, when the pixel response value Below the set threshold Mark it as a corner point, output a binary corner point map, and get a corner point set containing all marked corner points , .

4. The method for suppressing false echoes of a laser radar in a high-reflection environment according to claim 3, characterized in that: The determination of the current frame and the previous frame based on the time difference specifically includes: calculating the time difference of consecutive frames , , is the timestamp of the current frame, is the timestamp of the previous frame; if the time difference between the current frame and the previous frame When the time difference between the current frame and the previous frame is less than the set time difference threshold, return to the acquisition step. When the time difference is greater than or equal to the set threshold, the features are fused.

5. The method for suppressing false echoes of a laser radar in a high-reflection environment according to claim 1, characterized in that: The step S3 specifically includes the following steps: Step S31: pre-processed point cloud frame As the current frame, take the previous n frames for correlation processing; Step S32: The current frame Based on the current frame captured by the camera Fusion is performed to retain points in the radar point cloud that are consistent with the binocular camera point cloud, suppressing false echoes from inconsistent sensors; Step S33: From the current frame captured by the camera Extract all points of the face closest to the origin and form a set containing all points ; Step S34: For the current frame All points in the radar point cloud are judged and the points on the same ray but with a distance less than the set are removed. The point with the minimum distance in the middle is used to remove the false echo point; Step S35: fusion of the current frame Filtering and smoothing are performed to obtain the final environmental perception result.

6. The method for suppressing false echoes of a laser radar in a high-reflection environment according to claim 5, characterized in that: The correlation processing includes: Take the first n frames of data from the cache queue to form a cache set , is the depth of the historical frame, , according to the depth of the historical frame of the radar movement speed Adjust the depth of the history frame at faster speeds The smaller the selection; Use NDT algorithm for rough matching and pre-process point cloud frame As the target point cloud, the historical frame As the source point cloud, the target point cloud is divided into grids through the NDT algorithm. The point cloud in each grid is represented by a Gaussian distribution. The transformation parameters of the source point cloud are optimized to maximize the probability of projecting it onto the target grid, and an initial transformation matrix is ​​output. , the transformation matrix output by the NDT algorithm is used as the initial value of the GICP algorithm, the point cloud matching problem is transformed into the matching between probability distribution models, the optimal transformation is solved by minimizing the Mahalanobis distance between the two point clouds, and the optimized transformed matrix is ​​output , the historical frame After the transformation Matrix transformation to current frame In the coordinate system, the aligned historical frame point cloud is obtained; Traverse the current frame For each point in the transformed historical frame point cloud, find the neighboring points of each point. The search range is centered at any point with a radius of When there is no adjacent point in the historical frame, that is, there is no point in the region in the historical frame, the point is marked as a candidate transient point. Remove all points marked as candidate transient points in the current frame to obtain a new current frame ; Multiple historical frames are processed, and each historical frame is registered and the neighboring points are judged. When all historical frames have no points at the corresponding positions, they are removed.

7. The method for suppressing false echoes of a laser radar in a high-reflection environment according to claim 6, characterized in that: The new current frame Based on the current frame captured by the camera Fusion includes: all the current frames collected by the camera Convert the coordinate system of the binocular camera to the coordinate system of the lidar, and set the key parameters for compensating sensor calibration errors and measurement noise. , The absolute value of the calibration measurement error accuracy of the lidar, in the current frame converted to the lidar coordinate system In the current frame, search for a point m that matches the new current frame. The Euclidean distance of the points in is less than or equal to the key parameter , when there is at least one such point m, the point is retained, and when it is calculated that the point is consistent with the new current frame The Euclidean distance of the points in is greater than the key parameter When, from the current frame Remove the point in the radar point cloud and verify the spatial consistency of the binocular point cloud to remove the points in the radar point cloud that are inconsistent with the binocular point cloud, retain the real environment points, and suppress the false echoes caused by inconsistent sensors; The collection Including: the current frame based on camera acquisition For each point in Represents the direction vector, and calculates the pitch angle of the direction vector and azimuth , specifically: Press the direction space Divide into grids, assign each point to the corresponding direction interval, and calculate the pitch angle of each point. and azimuth Calculate the interval index to which it belongs, traverse each direction interval, skip it when there is no point in the interval, and directly add it to the nearest surface point set when there is only one point in the interval. When there are multiple points in the interval, compare the distances of these points, select the point with the smallest distance, and add it to the nearest surface point set. Finally, output the data containing the current frame. The set of all nearest surface points in .

8. The method for suppressing false echoes of a laser radar in a high-reflection environment according to claim 7, characterized in that: The current frame Judging all points in the set includes: The points in the pitch angle and azimuth Binning, each box represents a ray direction, stores the nearest point and its distance in that direction, establishes a mapping from the angle box to the nearest point distance, and traverses the current frame For each point in , calculate its pitch angle and azimuth And distance, discretize its angle into the corresponding angle box. When the angle box exists in the mapping from the angle box to the nearest point distance, it means there is a corresponding nearest surface point. Get the nearest surface point distance in that direction. When the points in the lidar point cloud are on the same ray and the distance from the origin is shorter than the nearest surface point set of the binocular camera If the corresponding point in the current frame is closer, the point is marked as a false echo. Removed from , and retained when the angle box does not exist.

Images