High-precision vehicle speed measurement method integrating lidar and vision technology

Abstract

This application relates to the field of image point cloud matching technology, specifically to a high-precision vehicle speed measurement method that integrates lidar and vision technology. The method includes: determining false eigenvalues ​​to divide clusters into false clusters and true clusters; filtering out point cloud distortion clusters from all true clusters based on the first eigenvalues ​​of all true clusters; recording all pixels with different coordinates between the suspected vehicle area corresponding to each point cloud distortion cluster and its adjacent vehicle area as distortion points, and filtering out discrete pixels from all distorted points in the two-dimensional depth image; and eliminating data points corresponding to all discrete pixels in the target point cloud and all data points in all false clusters to obtain a feature point cloud for vehicle speed measurement. This application addresses the impact of highly reflective objects on the accuracy of vehicle point cloud data and improves the accuracy of vehicle speed detection.

Description

Technical Field

[0001] The present application relates to the field of image point cloud matching technology, and specifically to a high-precision vehicle speed measurement method that integrates lidar and vision technologies. Background Art

[0002] In the field of transportation, measuring the speed of vehicles on the road is an important means to ensure traffic safety. LiDAR is widely used in the field of vehicle speed measurement due to its advantages such as high precision, little environmental interference, and no interference from visible light. Therefore, the vehicle's speed can be determined based on the center of gravity coordinates of the vehicle obtained from the point cloud data of the vehicle on the road collected by the three-dimensional LiDAR.

[0003] However, if the vehicle speed is detected only by lidar point cloud data, false targets will appear in the point cloud data collected by the lidar due to objects with high reflectivity such as license plates, headlights and traffic signs on the road. In addition, the vehicle position in the collected point cloud data will be distorted because the vehicle is located between highly reflective objects, which will affect the calculation results of the center of gravity coordinates of the vehicle in the point cloud data, thereby reducing the detection accuracy of the vehicle speed. Therefore, image fusion is needed for auxiliary judgment. Summary of the Invention

[0004] In order to solve the above technical problems, this application provides a high-precision vehicle speed measurement method that integrates lidar and vision technology to solve existing problems.

[0005] The high-precision vehicle speed measurement method of this application that integrates lidar and vision technology adopts the following technical solutions:

[0006] One embodiment of the present application provides a high-precision vehicle speed measurement method that integrates lidar and vision technology, the method comprising the following steps:

[0007] Acquire the target road's 3D point cloud, environment point cloud, and road image in real time;

[0008] The target point cloud is obtained by removing the data points corresponding to the environmental point cloud in the three-dimensional point cloud at each moment and all data points in its neighborhood. The target point cloud is converted into a two-dimensional depth image of the same size as the road image, all data points in the target point cloud are clustered, and the image area corresponding to each cluster is segmented in the two-dimensional depth image and recorded as the suspected vehicle area. The vehicle segmentation algorithm is used to extract all vehicle image areas in the road image at each moment. The distance between the geometric center of each suspected vehicle area and all vehicle image areas at each moment is analyzed, and the adjacent vehicle area of ​​each suspected vehicle area is screened out from all vehicle image areas. The first eigenvalue of each cluster is determined based on all pixels with different coordinates between each suspected vehicle area and its adjacent vehicle area.

[0009] By analyzing the similarity of the contours between each suspected vehicle area and its adjacent vehicle areas, the second eigenvalue of each cluster is determined, and combined with the first eigenvalue, the false eigenvalue of each cluster is determined to obtain a false segmentation threshold; based on the false segmentation threshold, all clusters in the target point cloud are divided into false clusters and real clusters; based on the first eigenvalues ​​of all real clusters, point cloud distorted clusters are filtered out from all real clusters; all pixels with different coordinates between the suspected vehicle area corresponding to each point cloud distorted cluster and its adjacent vehicle area are recorded as distorted points, and discrete pixels among all distorted points in the two-dimensional depth image are filtered out;

[0010] The data points corresponding to all discrete pixels in the target point cloud and all data points in all false clusters are eliminated to obtain the feature point cloud for vehicle speed measurement.

[0011] Preferably, the metric distance in the clustering process of all data points in the target point cloud is the Euclidean distance between the data points.

[0012] Preferably, the adjacent vehicle region of each suspected vehicle region is a vehicle image region corresponding to the minimum value of the geometric center distance between each suspected vehicle region and all vehicle image regions at each moment.

[0013] Preferably, the first eigenvalue of each cluster is: the proportion of the number of all pixels with different coordinates between the suspected vehicle area corresponding to each cluster and its adjacent vehicle area to the total number of pixels in the suspected vehicle area.

[0014] Preferably, the second eigenvalue of each cluster is a Jadekar similarity coefficient of the contours between the suspected vehicle region corresponding to each cluster and its adjacent vehicle region.

[0015] Preferably, the false eigenvalue of each cluster is the ratio of the normalized value of the first eigenvalue to the normalized value of the second eigenvalue of each cluster.

[0016] Preferably, obtaining the false segmentation threshold includes:

[0017] According to the target point cloud acquisition method, a preset number of historical target point clouds of the target roads are obtained, and the false point clouds corresponding to the high-reflectivity objects in each historical target point cloud are calibrated using a vision-led method; for all historical target point clouds, according to the false eigenvalue calculation method, the false eigenvalues ​​of all clusters in each historical target point cloud are obtained;

[0018] The false eigenvalues ​​of all clusters of all historical target point clouds are used as the input of the support vector machine, wherein the label of the false eigenvalue of the cluster belonging to the false point cloud is set to 1, and the label of the false eigenvalue of the cluster not belonging to the false point cloud is set to 0. The F1-Score is used as the evaluation index of the support vector machine to train the support vector machine, and the false eigenvalue corresponding to the segmentation plane of the trained support vector machine is used as the false segmentation threshold.

[0019] Preferably, the false clusters and true clusters are respectively clusters whose false eigenvalues ​​in the target point cloud at each moment are greater than a false segmentation threshold and clusters whose false eigenvalues ​​are less than or equal to the false segmentation threshold.

[0020] Preferably, the point cloud distorted cluster is a real cluster whose first eigenvalue is greater than 0.

[0021] Preferably, screening out discrete pixel points from all distorted points in the two-dimensional depth image includes:

[0022] All distorted points in the two-dimensional depth image are used as input to OpenCV, wherein the judgment length in OpenCV is set to a preset value of pixels, and the discrete pixel points in all distorted points are output.

[0023] This application has at least the following beneficial effects:

[0024] This application first removes the environmental point cloud and filters it to obtain the target point cloud and converts it into a two-dimensional depth image of the same size as the road image; then, it makes full use of the precise boundary information of vehicle segmentation in the road image, combines the lidar point cloud data, and calculates the first eigenvalue of each cluster by analyzing the geometric center distance and pixel coordinate difference between the suspected vehicle area and the adjacent real vehicle area to evaluate its spatial overlap, which helps to identify and distinguish the real vehicle point cloud, the distorted point cloud caused by multipath reflection interference, and the false point cloud caused by highly reflective objects. The image information provides a key basis for judging the authenticity of the point cloud, improves the quality of the point cloud data, and lays the foundation for subsequent accurate speed measurement; further, this application passes By analyzing the contour similarity (i.e., the second eigenvalue) and spatial overlap (i.e., the first eigenvalue) between the suspected vehicle area corresponding to the point cloud and its adjacent vehicle area on the image, combined with the threshold trained by machine learning, the shape information provided by the image is effectively utilized, thereby effectively distinguishing the false point clouds collected by the lidar, such as interference from highly reflective objects; in addition, by identifying the distorted point clouds of real vehicles caused by multipath reflection and eliminating discrete distorted points and false point cloud data, the collaborative analysis of images and point clouds ensures that the point cloud data finally used for speed measurement is closer to the real shape and position of the vehicle; finally, speed measurement is performed based on the processed high-quality point cloud data, and the point cloud data fused with image information significantly improves the accuracy of vehicle speed measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0026] Figure 1 A flowchart of the steps of a high-precision vehicle speed measurement method that integrates lidar and vision technology, provided in one embodiment of the present application;

[0027] Figure 2 A schematic diagram of the cluster division process provided in one embodiment of the present application. DETAILED DESCRIPTION

[0028] To further illustrate the technical means and effectiveness of this application to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effectiveness of the high-precision vehicle speed measurement method that integrates lidar and vision technology, as proposed in this application. In the following description, references to different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0029] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0030] The specific scheme of the high-precision vehicle speed measurement method that integrates lidar and vision technology provided by this application is described in detail below with reference to the accompanying drawings.

[0031] An embodiment of the present application provides a high-precision vehicle speed measurement method that integrates laser radar and vision technology. Specifically, the following high-precision vehicle speed measurement method that integrates laser radar and vision technology is provided. Please refer to Figure 1 , the method comprises the following steps:

[0032] Step S1: Acquire the three-dimensional point cloud of the target road, the environment point cloud, and the road image in real time.

[0033] This embodiment eliminates false target data and distorted data of vehicles on the road in the three-dimensional point cloud by analyzing the point cloud data of the target road collected by the lidar and the image of the target road collected by the camera, and then obtains the speed of the vehicle on the road based on the three-dimensional point cloud after the point cloud data elimination processing.

[0034] On the target road, a camera is installed at the location of the laser radar used to measure vehicle speed to collect road images of the road area monitored by the laser radar. The frame rate of the camera is consistent with the frame rate of the laser radar. Before using the camera to collect road images, the camera and the laser radar need to be synchronized in time and space to ensure that the camera and the laser radar can collect road images and three-dimensional point clouds at the same time and the same spatial position. In this embodiment, the collection frequency of the laser radar and the camera is 20 Hz, that is, 20 road images and three-dimensional point clouds are collected in one second respectively.

[0035] A real-time traffic vehicle monitoring system based on YOLOv8 is used to determine whether a vehicle appears in the road image captured by the camera. If a vehicle appears in the road image at the current moment, the three-dimensional point cloud and road image at the moment adjacent to the current moment are obtained from the lidar and camera respectively. Conversely, if no vehicle appears in the road image at the current moment, the three-dimensional point cloud captured by the lidar at the current moment is recorded as the environmental point cloud, and the historical environmental point cloud is updated at the same time.

[0036] Among them, the real-time traffic vehicle detection system based on YOLOv8 is a well-known technology, and the specific principles and processes of using it to determine whether there are vehicles in the road image are not described in detail.

[0037] Furthermore, in order to reduce noise interference in the three-dimensional point cloud, environmental point cloud and road image, this embodiment uses the three-dimensional point cloud, environmental point cloud and road point cloud as input of the filtering algorithm, and outputs the filtered three-dimensional point cloud, environmental point cloud and road point cloud for subsequent analysis. Among them, the road image is an image after grayscale processing. Grayscale processing is a well-known technology, and the specific process of grayscale conversion of the road image will not be repeated.

[0038] It should be noted that there are many commonly used filtering algorithms. In this embodiment, a median filtering algorithm is used to smooth the point cloud, i.e., the road image. In actual application, as other implementation methods, the implementer may also adopt other filtering methods such as the Gaussian filtering algorithm based on the specific situation. Regarding the selection of the filtering algorithm, this embodiment does not impose any special restrictions.

[0039] The median filtering algorithm is a well-known technology, and its specific principle will not be described in detail.

[0040] Step S2: remove the data points corresponding to the environmental point cloud at each moment in the three-dimensional point cloud and all data points in its neighborhood to obtain the target point cloud; convert the target point cloud into a two-dimensional depth image of the same size as the road image, cluster all data points in the target point cloud, and segment the image area corresponding to each cluster in the two-dimensional depth image, which is recorded as the suspected vehicle area; use the vehicle segmentation algorithm to extract all vehicle image areas in the road image at each moment; analyze the distance between the geometric center of each suspected vehicle area and all vehicle image areas at each moment, and filter out the adjacent vehicle area of ​​each suspected vehicle area from all vehicle image areas; determine the first eigenvalue of each cluster based on all pixel points with different coordinates between each suspected vehicle area and its adjacent vehicle area.

[0041] Since lidar is very sensitive to high-intensity echoes reflected by high-reflectivity targets, in addition to a set of real point cloud data of the target object in the original real coordinate system, a set of false target point cloud data of similar size and shape may also appear in other locations near it.

[0042] Therefore, in order to reduce the impact of false target point clouds appearing in the three-dimensional point cloud caused by objects with high reflectivity such as license plates and lights of vehicles on the target road, as well as traffic signs and road guardrails on the road, on the subsequent vehicle speed detection results, this embodiment obtains the target point cloud by removing the data points corresponding to the environmental point cloud in the three-dimensional point cloud at each moment and all data points in its neighborhood; converts the target point cloud into a two-dimensional depth image of the same size as the road image, clusters all data points in the target point cloud, and segments the image area corresponding to each cluster in the two-dimensional depth image, which is recorded as the suspected vehicle area; uses the vehicle segmentation algorithm to extract all vehicle image areas in the road image at each moment; analyzes the distance between the geometric center of each suspected vehicle area and all vehicle image areas at each moment, and filters out the adjacent vehicle areas of each suspected vehicle area from all vehicle image areas, specifically:

[0043] First, the data points corresponding to the environmental point cloud at each moment in the three-dimensional point cloud and all data points in its neighborhood are removed to obtain the target point cloud, wherein the process of dividing the data point neighborhood is as follows: as an implementation method, in this embodiment, each data point is taken as the center and a circular area with a neighborhood radius twice the unit length of the coordinate system where the three-dimensional point cloud is located is used as the neighborhood of each data point. In actual application, as other implementation methods, the implementer can also set the neighborhood division method according to the specific situation. This embodiment does not impose any special restrictions.

[0044] Furthermore, a statistical outlier removal (SOR) filter is used to filter the three-dimensional point cloud to remove isolated points in the target point cloud, and a filtered target point cloud is obtained for subsequent analysis.

[0045] Among them, statistical outlier removal (SOR) is a well-known technology, and the specific process of removing isolated points in the target point cloud using the statistical filter will not be described in detail.

[0046] Furthermore, in this embodiment, the target point cloud is converted into a two-dimensional depth image of the same size as the road image. Since the camera and the lidar are synchronized in time and space, the two-dimensional depth image obtained by converting the target point cloud and the road image captured by the camera not only have the same image size, but also the pixels with the same coordinates in the two images correspond to the same actual spatial position.

[0047] Furthermore, this embodiment uses the target point cloud as the input of the clustering algorithm, wherein the metric distance is set to the Euclidean distance between data points in the target point cloud. This embodiment uses the elbow method to determine the number of clusters. Finally, all clusters in the target point cloud are output to characterize the area corresponding to the suspected vehicle object in the target point cloud.

[0048] It should be noted that there are many commonly used clustering algorithms. In this embodiment, the k-means clustering algorithm is used to cluster data points. In actual application, as other implementation methods, implementers can also use other clustering methods such as density clustering algorithms based on specific circumstances. Regarding the selection of clustering methods, this embodiment does not impose any special restrictions.

[0049] Among them, the k-means clustering algorithm, the elbow method and the calculation method of the Euclidean distance are all well-known technologies. The process of clustering data points using the k-means clustering algorithm, the process of obtaining the number of clusters using the elbow method and the calculation process of the Euclidean distance are not repeated here.

[0050] Furthermore, this embodiment records the image region corresponding to each cluster segmented in the two-dimensional depth image as a suspected vehicle region. Furthermore, in this embodiment, a YOLO-based vehicle segmentation algorithm for traffic scenes is used to extract all vehicle image regions in the road image at each time. The YOLO-based vehicle segmentation algorithm for traffic scenes is a well-known technique, and its use to extract all vehicle image regions in the road image at each time is also a well-known technique. The specific process of extracting the vehicle image regions is not further described.

[0051] Furthermore, this embodiment screens out the adjacent vehicle regions of each suspected vehicle region from all vehicle image regions by analyzing the distance between the geometric centers of each suspected vehicle region and all vehicle image regions at each moment, specifically:

[0052] As an implementation method, in this embodiment, the distance between the geometric centers of each suspected vehicle area and all vehicle image areas at each moment is calculated, and the vehicle image area corresponding to the minimum distance is used as the adjacent vehicle area of ​​each suspected vehicle area at each moment.

[0053] Furthermore, this embodiment determines the first eigenvalue of each cluster based on all pixels with different coordinates between each suspected vehicle area and its adjacent vehicle areas, which is used to evaluate the degree of spatial overlap between the suspected vehicle area and the real vehicle area identified by the image.

[0054] At this point, this embodiment first removes the environmental point cloud and filters it to obtain the target point cloud and convert it into a two-dimensional depth image. Then, combined with the visual vehicle segmentation results, by analyzing the geometric center distance and pixel coordinate difference between the suspected vehicle area and the adjacent real vehicle area, the first eigenvalue of each cluster is calculated to evaluate its spatial overlap. This helps to identify and distinguish between real vehicle point clouds, distorted point clouds interfered by multipath reflections, and false point clouds caused by highly reflective objects, thereby improving the quality of point cloud data and laying the foundation for subsequent accurate speed measurement.

[0055] Step S3: By analyzing the similarity of the contours between each suspected vehicle area and its adjacent vehicle area, the second eigenvalue of each cluster is determined, and combined with the first eigenvalue, the false eigenvalue of each cluster is determined to obtain a false segmentation threshold; based on the false segmentation threshold, all clusters in the target point cloud are divided into false clusters and real clusters; based on the first eigenvalues ​​of all real clusters, point cloud distorted clusters are screened out from all real clusters; all pixel points with different coordinates between the image area corresponding to each point cloud distorted cluster and its adjacent vehicle area are recorded as distorted points, and discrete pixel points among all distorted points in the two-dimensional depth image are screened out.

[0056] During the process of collecting 3D point clouds of the target road using a LiDAR, if a vehicle on the target road is located between strong reflective objects, such as between a large truck and a road guardrail, the laser beam emitted by the LiDAR may first hit the large truck or the road guardrail. The metal body of the large truck or the reflective stickers on the body, as well as the reflective paint on the road guardrail, will reflect the laser beam, which will then hit the vehicle and finally reflect back to the LiDAR receiver. This multipath reflection will cause the echo signal received by the LiDAR to contain information from multiple reflection paths, causing interference with the vehicle's position information. This will cause the point cloud data corresponding to the vehicle in the collected 3D point cloud to exhibit lateral position deviation or jitter distortion. In turn, some point cloud data corresponding to the vehicle will appear at locations in the real world where there are no real object features. This will result in a certain non-overlapping area between the spatial region where the point cloud data corresponding to the vehicle is located and the spatial region where the vehicle is actually located. However, compared to the false point cloud in the collected 3D point cloud, the false point cloud will have a shape that is more similar to the vehicles adjacent to it in the spatial region.

[0057] Therefore, this embodiment determines the second eigenvalue of each cluster by analyzing the similarity between the contours of each suspected vehicle area and its adjacent vehicle areas, and determines the false eigenvalue of each cluster in combination with the first eigenvalue to obtain a false segmentation threshold. Based on the false segmentation threshold, all clusters in the target point cloud are divided into false clusters and real clusters to distinguish between false point clouds caused by high-reflectivity objects in the target point cloud and point clouds corresponding to real objects. Specifically:

[0058] In this embodiment, first, the second eigenvalue of each cluster is determined by evaluating the similarity between the contours of each suspected vehicle region and its adjacent vehicle regions. Specifically:

[0059] In this embodiment, the Jedkar similarity coefficient of the contour between the suspected vehicle area corresponding to each cluster and its adjacent vehicle area is used as the second eigenvalue of each cluster. If the second eigenvalue of the current cluster is larger, it means that the suspected vehicle corresponding to the current cluster and its adjacent vehicles are very similar in contour shape, which usually means that the point cloud in the current cluster is more likely to be point cloud data belonging to a real vehicle; conversely, if the second eigenvalue of the current cluster is smaller, it means that the suspected vehicle corresponding to the current cluster and its adjacent vehicles are very different in contour shape, which usually means that the point cloud in the current cluster is more likely to be false point cloud data, such as point cloud data generated by highly reflective objects such as license plates and headlights.

[0060] The Jadekar similarity coefficient is a method for calculating contour similarity, which is a well-known technology and its specific calculation process will not be described in detail.

[0061] Furthermore, this embodiment determines the false eigenvalue of each cluster based on the first eigenvalue and the second eigenvalue of each cluster, which is used to evaluate whether the point cloud data corresponding to the cluster is the point cloud data corresponding to objects with high reflectivity such as the license plate and headlights of vehicles on the target road, as well as traffic signs and road guardrails on the road. Specifically:

[0062] In this embodiment, the first eigenvalues ​​and second eigenvalues ​​of all clusters in the target point cloud at each moment are normalized respectively. There are many commonly used normalization methods. In this embodiment, the maximum and minimum value algorithms are used to normalize the first eigenvalues ​​and the second eigenvalues ​​respectively. In actual application, the implementer may also use other normalization methods such as z-score normalization method for normalization according to specific circumstances. This embodiment does not impose any special restrictions.

[0063] Furthermore, the ratio of the normalized value of the first eigenvalue to the normalized value of the second eigenvalue of each cluster is used as the false eigenvalue of each cluster.

[0064] According to the false eigenvalue of each cluster, it can be understood that if the first eigenvalue of the current cluster is larger, it means that the spatial position of the suspected vehicle corresponding to the current cluster and the nearby real vehicles do not overlap to a high degree. At the same time, if the second eigenvalue of the current cluster is smaller, it means that the contour difference between the suspected vehicle corresponding to the current cluster and the nearby real vehicles is greater, resulting in a larger false eigenvalue, and the more likely this cluster is to be a false point cloud, because it may not only be consistent with the real vehicle in spatial position, but also have a completely different shape from the real vehicle.

[0065] On the contrary, if the first eigenvalue of the current cluster is smaller, it means that the spatial position overlap between the suspected vehicle corresponding to the current cluster and the nearby real vehicle is very high, indicating that it is spatially closer to the real vehicle. At the same time, if the second eigenvalue of the current cluster is larger, it means that the contour similarity between the suspected vehicle corresponding to the current cluster and the nearby real vehicle is higher, and the shape is closer. The high spatial overlap and high shape similarity work together to make the false eigenvalue smaller. Therefore, it is more likely to be the point cloud data of the real vehicle, or the distorted data caused by the vehicle being between highly reflective objects. However, it is more likely to be a point cloud related to the real vehicle rather than a completely false target. This makes the subsequent calculation of the vehicle speed based on the point cloud center of gravity coordinates more reliable.

[0066] Furthermore, this embodiment divides all clusters in the target point cloud into false clusters and real clusters based on the false segmentation threshold, specifically:

[0067] In this embodiment, according to the target point cloud acquisition method, a preset number of historical target point clouds of target roads are obtained, and the false point clouds corresponding to high-reflectivity objects in each historical target point cloud are calibrated using a vision-led method. For all historical target point clouds, according to the false eigenvalue calculation method, the false eigenvalues ​​of all clusters in each historical target point cloud are obtained;

[0068] The false eigenvalues ​​of all clusters of all historical target point clouds are used as the input of the support vector machine, wherein the label of the false eigenvalue of the cluster belonging to the false point cloud is set to 1, and the label of the false eigenvalue of the cluster not belonging to the false point cloud is set to 0. The F1-Score is used as the evaluation index of the support vector machine to train the support vector machine, and the false eigenvalue corresponding to the segmentation plane of the trained support vector machine is used as the false segmentation threshold.

[0069] It should be noted that the value of the preset number is set manually. In this embodiment, the value of the preset number is 1000. In actual application, as other implementation methods, the implementer can also set it by himself based on the specific situation. This embodiment does not impose any special restrictions.

[0070] Among them, the specific training process of the support vector machine and F1-Score are well-known technologies, and the specific training process and the process of using F1-Score as the evaluation indicator of the support vector machine are not repeated here.

[0071] It should be noted that the vision-dominated method manually screens out false point clouds and non-false point clouds. The corresponding area of ​​the false point cloud in the two-dimensional depth image does not correspond to the same area in the road image, while the corresponding area of ​​the non-false point cloud does correspond to the same area in the road image.

[0072] Furthermore, based on the false feature values, all clusters in the target point cloud at each moment are divided into false clusters and real clusters, specifically:

[0073] In this embodiment, the clusters whose false feature values ​​in the target point cloud at each moment are greater than the false segmentation threshold are regarded as false clusters, and the clusters whose false feature values ​​in the target point cloud at each moment are less than or equal to the false segmentation threshold are regarded as real clusters. The false clusters are used to characterize the false point clouds in the target point cloud caused by objects with high reflectivity such as vehicle license plates, headlights, traffic signs and road guardrails on the road, and the real clusters are used to characterize the clusters composed of vehicle point cloud data.

[0074] Preferably, the cluster division process diagram provided in this embodiment is as follows: Figure 2 shown.

[0075] Secondly, in order to reduce the impact of distortion of some point cloud data caused by vehicles on the target road being between highly reflective objects on the subsequent vehicle speed detection results, this embodiment filters out point cloud distortion clusters from all real clusters based on the first eigenvalues ​​of all real clusters; all pixels with different coordinates between the suspected vehicle area corresponding to each point cloud distortion cluster and its adjacent vehicle area are recorded as distorted points, and discrete pixels among all distorted points in the two-dimensional depth image are filtered out, specifically:

[0076] First, in this embodiment, based on the first eigenvalues ​​of all true clusters, point cloud distortion clusters are screened out from all true clusters. Specifically:

[0077] In this embodiment, a real cluster whose first eigenvalue is greater than 0 among all real clusters of the target point cloud at each moment is used as a point cloud distorted cluster to represent the real cluster containing distorted data in the target point cloud.

[0078] Furthermore, all pixels with different coordinates between the suspected vehicle area corresponding to each point cloud distortion cluster and its adjacent vehicle area are recorded as distortion points;

[0079] Furthermore, all distorted points in the two-dimensional depth image corresponding to the target point cloud at each moment are used as input to OpenCV, wherein the judgment length in OpenCV is set to a preset value of pixels, and the discrete pixel points in all distorted points are output.

[0080] It should be noted that the value of the preset value is set manually. In this embodiment, the value of the preset value is 3. In actual application, as other implementation methods, the implementer can also set it by himself based on the specific situation. This embodiment does not impose any special restrictions.

[0081] The process of using OpenCV to filter out discrete pixels from all distorted points is a well-known technique, and the specific principle and process will not be described in detail.

[0082] Thus, this embodiment effectively distinguishes between false point clouds collected by the lidar, such as interference from highly reflective objects, and real vehicle point clouds, by analyzing the contour similarity (i.e., the second eigenvalue) and spatial overlap (i.e., the first eigenvalue) between the point cloud and neighboring vehicles, combined with the threshold trained by machine learning. Furthermore, by identifying the distorted point clouds of real vehicles caused by multipath reflections and eliminating discrete distorted points and false point cloud data, it helps to improve the accuracy and reliability of vehicle speed measurement.

[0083] Step S4: Eliminate all data points corresponding to discrete pixels in the target point cloud and all data points in all false clusters to obtain a feature point cloud for vehicle speed measurement.

[0084] Based on step S3, false point clouds and distorted point clouds that affect vehicle speed measurement are screened out in the target point cloud of the target road at each moment. Therefore, this embodiment further obtains a feature point cloud for vehicle speed measurement by eliminating data points corresponding to all discrete pixel points in the target point cloud and all data points in all false clusters.

[0085] Furthermore, the feature point cloud used for vehicle speed measurement at all times within a preset time period is transmitted to the speed measurement unit of the lidar, and the speed detection results of each vehicle on the target road are output.

[0086] It should be noted that the value of the preset time length is set manually. In this embodiment, the value of the preset time length is 1s, that is, the vehicle speed is measured every 1 second. In actual application, as other implementation methods, the implementer can also set it by himself based on the specific situation. This embodiment does not impose any special restrictions.

[0087] Among them, the process of using laser radar to measure vehicle speed based on the feature point cloud between adjacent time periods is a well-known technology and will not be described in detail.

[0088] This embodiment integrates lidar and vision, removing ambient point clouds, filtering, and clustering. Combined with visual vehicle detection results, the system uses the spatial overlap and outline similarity between the point cloud and neighboring vehicles to assess the authenticity of each point cloud cluster. This effectively identifies and removes false point clouds caused by highly reflective objects, as well as point clouds distorted by multipath reflections from vehicles. Ultimately, speed measurement based on this processed, high-quality point cloud data improves the accuracy and reliability of vehicle speed measurement.

[0089] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0090] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0091] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them. Modifications to the technical solutions described in the aforementioned embodiments, or equivalent replacements of some of the technical features therein, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A high-precision vehicle speed measurement method integrating laser radar and vision technology, characterized in that: The method comprises the following steps: Acquire the target road's 3D point cloud, environment point cloud, and road image in real time; The target point cloud is obtained by removing the data points corresponding to the environmental point cloud in the three-dimensional point cloud at each moment and all data points in its neighborhood. The target point cloud is converted into a two-dimensional depth image of the same size as the road image, all data points in the target point cloud are clustered, and the image area corresponding to each cluster is segmented in the two-dimensional depth image and recorded as the suspected vehicle area. The vehicle segmentation algorithm is used to extract all vehicle image areas in the road image at each moment. The distance between the geometric center of each suspected vehicle area and all vehicle image areas at each moment is analyzed, and the adjacent vehicle area of ​​each suspected vehicle area is screened out from all vehicle image areas. The first eigenvalue of each cluster is determined based on all pixels with different coordinates between each suspected vehicle area and its adjacent vehicle area. By analyzing the similarity of the contours between each suspected vehicle area and its adjacent vehicle areas, the second eigenvalue of each cluster is determined, and combined with the first eigenvalue, the false eigenvalue of each cluster is determined to obtain a false segmentation threshold; based on the false segmentation threshold, all clusters in the target point cloud are divided into false clusters and real clusters; based on the first eigenvalues ​​of all real clusters, point cloud distorted clusters are filtered out from all real clusters; all pixels with different coordinates between the suspected vehicle area corresponding to each point cloud distorted cluster and its adjacent vehicle area are recorded as distorted points, and discrete pixels among all distorted points in the two-dimensional depth image are filtered out; Eliminate all data points corresponding to discrete pixels in the target point cloud and all data points in all false clusters to obtain the feature point cloud for vehicle speed measurement; The obtaining of the false segmentation threshold comprises: According to the target point cloud acquisition method, a preset number of historical target point clouds of the target roads are obtained, and the false point clouds corresponding to the high-reflectivity objects in each historical target point cloud are calibrated using a vision-led method; for all historical target point clouds, according to the false eigenvalue calculation method, the false eigenvalues ​​of all clusters in each historical target point cloud are obtained; The false eigenvalues ​​of all clusters of all historical target point clouds are used as the input of the support vector machine, wherein the label of the false eigenvalue of the cluster belonging to the false point cloud is set to 1, and the label of the false eigenvalue of the cluster not belonging to the false point cloud is set to 0. The F1-Score is used as the evaluation index of the support vector machine to train the support vector machine, and the false eigenvalue corresponding to the segmentation plane of the trained support vector machine is used as the false segmentation threshold.

2. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The metric distance in the clustering process of all data points in the target point cloud is the Euclidean distance between data points.

3. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The adjacent vehicle region of each suspected vehicle region is a vehicle image region corresponding to the minimum value of the geometric center distance between each suspected vehicle region and all vehicle image regions at each moment.

4. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The first characteristic value of each cluster is: the proportion of the number of all pixels with different coordinates between the suspected vehicle area corresponding to each cluster and its adjacent vehicle area to the total number of pixels in the suspected vehicle area.

5. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The second eigenvalue of each cluster is a Jedkar similarity coefficient of the contours between the suspected vehicle region corresponding to each cluster and its adjacent vehicle regions.

6. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The false eigenvalue of each cluster is the ratio of the normalized value of the first eigenvalue to the normalized value of the second eigenvalue of each cluster.

7. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The false clusters and true clusters are respectively clusters whose false feature values ​​in the target point cloud at each moment are greater than the false segmentation threshold and clusters whose false feature values ​​are less than or equal to the false segmentation threshold.

8. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The point cloud distorted cluster is a real cluster whose first eigenvalue is greater than 0.

9. The high-precision vehicle speed measurement method integrating laser radar and vision technology as claimed in claim 1, characterized in that: The step of filtering out discrete pixel points from all distorted points in the two-dimensional depth image includes: All distorted points in the two-dimensional depth image are used as input to OpenCV, wherein the judgment length in OpenCV is set to a preset value of pixels, and the discrete pixel points in all distorted points are output.

Images