An empty parking space recognition method, device and apparatus of an autonomous driving system

Abstract

The present application relates to the technical field of data processing, in particular to a kind of empty parking space identification method, equipment and device of automatic driving system, comprising: obtaining the point cloud data and millimeter wave radar data and feature data set of each time;According to the density of point cloud data in feature data set and the spatial feature of feature data set, obtain the abnormal data set of initial historical time;According to the abnormal performance of each point cloud data in feature data set, combined with the difference of millimeter wave radar data, the interference performance degree is obtained;In the abnormal data set of initial historical time corresponding to the same spatial position, the change of interference performance degree is tracked to obtain the noise influence coefficient of current time;According to the noise influence coefficient, determine the filter kernel size, carry out denoising processing to the point cloud data of current time, and the data after denoising is used for empty parking space identification to the vehicle to be analyzed.The present application improves the environmental adaptability of denoising algorithm by dynamically adjusting filter kernel size.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method, device, and apparatus for identifying available parking spaces in an autonomous driving system. Background Technology

[0002] Vacant parking space identification is one of the core technologies for achieving autonomous parking in autonomous driving systems. Traditional parking space identification mainly relies on single-mode perception from onboard cameras or millimeter-wave radar, detecting the position of parking lines or obstacles through image processing or echo signal analysis. Existing methods use point cloud data for vacant parking space identification. For example, existing methods for parking space identification based on point cloud data are disclosed in "Methods for Localizing Autonomous Vehicles in Underground Parking Lots" and "A Parking Control Method, Device, and System".

[0003] However, such methods rely on single-modal data for identification, resulting in poor environmental adaptability. During automatic parking, various interference factors and complex parking environments can cause a decrease in recognition accuracy and may lead to automatic parking malfunctions. Therefore, to reduce the impact of environmental factors on automatic parking space identification, it is necessary to remove environmental noise from the collected radar data. However, existing data denoising methods generally use fixed filter sizes, which cannot account for the influence of interference factors in the automatic parking environment, resulting in poor denoising effects and consequently affecting the accuracy of identifying available parking spaces during automatic parking. Summary of the Invention

[0004] To address the problem that existing data denoising methods employ fixed filter sizes, resulting in poor denoising performance, this invention aims to provide a method, device, and apparatus for identifying available parking spaces in an autonomous driving system. The specific technical solution adopted is as follows:

[0005] In a first aspect, the present invention provides a method for identifying available parking spaces in an autonomous driving system, comprising:

[0006] Based on the vehicle-mounted lidar and millimeter-wave radar of the vehicle to be analyzed, point cloud data and millimeter-wave radar data at each moment are acquired, wherein the moment includes the current moment and several historical moments.

[0007] Clustering is performed on the point cloud data at each time step to obtain the feature dataset for each time step; based on the density of the point cloud data in the feature dataset and the spatial characteristics of the feature dataset, the anomaly dataset for the initial historical time step is obtained.

[0008] Based on the abnormal behavior of each point cloud data in the feature dataset, and combined with the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset, the degree of interference performance of each feature dataset at each time moment is obtained.

[0009] By tracking the change in the degree of interference over time in the same spatial location corresponding to the abnormal dataset at the initial historical moment, the noise impact coefficient at the current moment can be obtained.

[0010] The filter kernel size is determined based on the noise impact coefficient, and the point cloud data at the current moment is denoised. The denoised data is then used to identify available parking spaces for the vehicle to be analyzed.

[0011] Preferably, obtaining the anomaly dataset for the initial historical time based on the density of point cloud data in the feature dataset and the spatial characteristics of the feature dataset specifically includes:

[0012] For any given feature dataset, the first feature coefficient of the feature dataset is determined based on the density value of the point cloud data per unit volume in the feature dataset; the volume of all point cloud data in the feature dataset in space is used as the second feature coefficient of the feature dataset.

[0013] Based on the negative correlation coefficient of the second feature coefficient and the first feature coefficient, the abnormal performance degree of any feature dataset is determined; based on the abnormal performance degree of each feature dataset at each historical moment, anomaly analysis is performed on the feature dataset to obtain the abnormal dataset at the initial historical moment.

[0014] Preferably, the step of performing anomaly analysis on the feature dataset based on the anomaly performance of each feature dataset at each historical moment to obtain the anomaly dataset at the initial historical moment specifically includes:

[0015] The feature datasets corresponding to anomalies exceeding a preset anomaly threshold are denoted as suspected anomaly feature sets; the first historical moment corresponding to all suspected anomaly feature sets is taken as the initial historical moment, and each suspected anomaly feature set at the initial historical moment is the anomaly dataset at the initial historical moment.

[0016] Preferably, the step of obtaining the interference performance level of each feature dataset at each time moment based on the abnormal performance of each point cloud data in the feature dataset, combined with the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset, specifically includes:

[0017] For any feature dataset, based on the position distribution of the center points of all point cloud data in the lidar coordinate system, the center points of the feature dataset are mapped from the lidar coordinate system to the millimeter-wave radar coordinate system to obtain the first dimension data points corresponding to the feature dataset.

[0018] Based on each millimeter-wave radar data at each moment, determine the second-dimensional data point in the millimeter-wave radar coordinate system at each moment;

[0019] At the same time, the spatial distance between the second-dimensional data point that is closest to the first-dimensional data point is obtained, which is used as the degree of data difference in the feature dataset at the corresponding time.

[0020] The product of the negative correlation coefficient of the degree of data difference in the feature dataset and the degree of abnormal performance is taken as the degree of interference performance of any feature dataset.

[0021] Preferably, the step of tracking the change in the degree of interference over time at the same spatial location corresponding to the abnormal dataset at the initial historical moment, to obtain the noise impact coefficient at the current moment, specifically includes:

[0022] Based on the location of the center point of the abnormal dataset at the initial historical moment, the feature dataset at each historical moment is matched to obtain the matching dataset corresponding to each abnormal dataset at each historical moment and the current moment.

[0023] The degree of influence of each anomalous dataset is obtained by analyzing the changes in the degree of interference in each anomalous dataset and all corresponding matching datasets.

[0024] The normalized result of the mean of the influence of all outlier datasets at the initial historical moment is used as the noise influence coefficient at the current moment.

[0025] Preferably, the step of matching the feature dataset of each historical time with the center point of the abnormal dataset at the initial historical time to obtain the matching dataset corresponding to each abnormal dataset at each historical time and the current time specifically includes:

[0026] For any anomalous dataset at the initial historical moment, the position of the center point of all point cloud data in the anomalous dataset in the lidar coordinate system is obtained as the target point position;

[0027] At each historical moment after the initial historical moment and at the current moment, obtain the feature dataset of the point cloud data closest to the target point location, and obtain the matching dataset corresponding to the abnormal data at each historical moment and the current moment.

[0028] Preferably, the step of determining the influence level of each anomalous dataset based on the changes in the interference performance of each anomalous dataset and all corresponding matching datasets specifically includes:

[0029] The perturbation performance of each anomalous dataset at the initial historical moment and the perturbation performance of the corresponding matching dataset are arranged in chronological order to form a perturbation performance sequence for each anomalous dataset.

[0030] In any abnormal dataset, any two adjacent data values ​​in the perturbation sequence are denoted as the first data value and the second data value, respectively.

[0031] When the absolute value of the difference between the first data value and the second data value is less than or equal to a preset first threshold, the first preset value is used as the change influencing factor between the first data value and the second data value.

[0032] When the absolute value of the difference between the first data value and the second data value is greater than the preset second threshold, the second preset value is used as the change influencing factor between the first data value and the second data value.

[0033] Wherein, the first preset value is greater than the second preset value, and the first threshold is less than the second threshold;

[0034] The degree of influence of any one of the abnormal datasets is obtained by summing the values ​​of all the changing influencing factors in the disturbance performance sequence of the abnormal dataset.

[0035] Preferably, determining the filter kernel size based on the noise impact coefficient specifically includes:

[0036] The integer result of the product of the noise impact coefficient at the current moment and the preset initial filter kernel size is used as the filter kernel size.

[0037] In a second aspect, the present invention provides an vacant parking space identification device for an autonomous driving system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor. When the computer program is executed by the processor, it implements the steps of an vacant parking space identification method for an autonomous driving system.

[0038] Thirdly, the present invention provides a parking space identification device for an autonomous driving system, the device being used to implement the steps of a parking space identification method for an autonomous driving system, the parking space identification device for the autonomous driving system comprising:

[0039] The data acquisition module is used to acquire point cloud data and millimeter-wave radar data at each moment based on the vehicle-mounted lidar and millimeter-wave radar of the vehicle to be analyzed, wherein the moment includes the current moment and several historical moments.

[0040] The feature recognition module is used to cluster the point cloud data at each time moment to obtain the feature dataset at each time moment; and to obtain the anomaly dataset of the initial historical time moment based on the density of the point cloud data in the feature dataset and the spatial characteristics of the feature dataset.

[0041] The interference analysis module is used to obtain the degree of interference performance of each feature dataset at each time moment based on the abnormal performance of each point cloud data in the feature dataset and the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset.

[0042] The noise impact module is used to track the changes in the degree of interference of the feature datasets of each adjacent historical time before the current time, under the same spatial location corresponding to the abnormal datasets at the initial historical time, and to obtain the noise impact coefficient at the current time.

[0043] The parking space recognition module is used to determine the filter kernel size based on the noise impact coefficient, and to denoise the point cloud data at the current moment. The denoised data is used to identify available parking spaces for the vehicle to be analyzed.

[0044] The embodiments of the present invention have at least the following beneficial effects:

[0045] This invention first collects data from two different dimensions, using point cloud data from LiDAR as the primary analysis data, combined with other millimeter-wave radar data to analyze noise interference, providing a data foundation for subsequent analysis. Then, feature datasets are obtained through clustering, characterizing the set of point cloud data containing suspected obstacles, and anomaly analysis is used to determine the earliest historical moment of anomaly and its corresponding anomalous dataset. Further, the spatial state of suspected obstacles corresponding to each feature dataset in the environment is analyzed, i.e., the degree of interference. Then, the continuous changes in the interference performance of the anomalous dataset at the initial historical moment are tracked to obtain the noise impact coefficient at the current moment, reflecting the extent to which the point cloud data is affected by the environment at the current moment. Finally, the filter kernel size is adaptively obtained to denoise the point cloud data. By dynamically adjusting the filter kernel size, the limitations of fixed filter parameters are avoided, improving the environmental adaptability of the denoising algorithm. This allows for automatic adjustment of the filtering degree in complex environments, effectively improving the robustness and accuracy of vacant parking space identification. Attached Figure Description

[0046] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of the steps of an autonomous driving system vacant parking space identification method provided by the present invention;

[0048] Figure 2 This is a flowchart of the steps of the method for obtaining the abnormal dataset at the initial historical moment provided by the present invention;

[0049] Figure 3 This is a flowchart of the steps of the method for obtaining the degree of interference performance provided by the present invention;

[0050] Figure 4 This is a flowchart of the steps for obtaining the noise influence coefficient at the current moment provided by the present invention;

[0051] Figure 5 This is a structural block diagram of an available parking space identification device for an autonomous driving system provided by the present invention. Detailed Implementation

[0052] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a method, device, and apparatus for identifying available parking spaces in an autonomous driving system according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0053] Unless otherwise defined, 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 invention pertains.

[0054] The specific implementation scenario addressed by this invention is as follows: A lidar system identifies vacant parking spaces by emitting laser beams and receiving reflected signals. First, the lidar performs a high-speed 3D scan of the target area, generating high-precision point cloud data that clearly presents the 3D outlines of parking space boundaries, adjacent vehicles, and other obstacles. Second, an intelligent algorithm analyzes the spatial distribution characteristics of the point cloud, automatically identifying the geometric shape of standard parking space lines (such as parallel lines, right angles, etc.) and detecting whether there are obstacle point cloud clusters within the parking space. For unmarked parking spaces, the available space is determined based on the distance between surrounding obstacles. The system integrates multiple frames of scan data, eliminates interference from temporary obstacles (such as pedestrians), and finally outputs the location and size information of vacant parking spaces.

[0055] The main objective of this invention is to filter the collected LiDAR point cloud data before identifying available parking spaces for automatic parking. The size of the filter kernel is adaptively determined by combining the feature analysis results of the LiDAR data and the feature analysis of the millimeter-wave radar data, using a dual-modal data joint analysis method, which can effectively improve the filtering accuracy of the data.

[0056] The following description, in conjunction with the accompanying drawings, details the specific scheme of the available parking space identification method, equipment, and apparatus provided by the present invention for an autonomous driving system.

[0057] Please see Figure 1 The diagram illustrates a flowchart of a method for identifying available parking spaces in an autonomous driving system according to an embodiment of the present invention. The method includes the following steps:

[0058] Step S100: Based on the vehicle-mounted lidar and millimeter-wave radar of the vehicle to be analyzed, acquire point cloud data and millimeter-wave radar data at each moment, wherein the moment includes the current moment and several historical moments.

[0059] First, at each moment, point cloud data is acquired using the vehicle-mounted LiDAR and millimeter-wave radar data are acquired using the vehicle-mounted millimeter-wave radar. These moments include the current moment and several historical moments. In this embodiment, the number of historical moments is set to 5, and the time interval between adjacent moments is equal, each being 1 second. Implementers can adjust this setting according to their specific implementation scenarios.

[0060] It should be noted that lidar provides high-precision 3D point clouds, while millimeter-wave radar excels at detecting the speed and distance of dynamic objects. Joint analysis allows for complementary advantages, improving the accuracy and robustness of obstacle recognition. In this embodiment, point cloud data refers to data acquired through lidar, which is a 3D point cloud in the lidar coordinate system. The lidar coordinate system is a 3D coordinate system with the lidar's location as the origin when the lidar is the target object.

[0061] Similarly, the millimeter-wave radar coordinate system refers to a three-dimensional coordinate system with the origin of the millimeter-wave radar as the location of the radar when the millimeter-wave radar is the detection object. The millimeter-wave radar data directly output by the millimeter-wave radar includes the straight-line distance and angle between the detected target and the millimeter-wave radar. The object coordinate position of the detected target in the millimeter-wave radar coordinate system can be directly determined through the millimeter-wave radar data.

[0062] It should be further explained that millimeter-wave radar transmits high-frequency electromagnetic waves. When these waves encounter an object, some of their energy is reflected back to the radar's receiving antenna. A mixer combines the transmitted wave and the echo into an intermediate-frequency signal, which is then analyzed using a Fourier transform to determine the target's distance, velocity, and angle. This is a well-known technology and will only be briefly introduced here without further elaboration.

[0063] Step S200: Cluster the point cloud data at each time point to obtain the feature dataset at each time point; obtain the anomaly dataset of the initial historical time point based on the density of the point cloud data in the feature dataset and the spatial characteristics of the feature dataset.

[0064] First, based on the point cloud data corresponding to the lidar, abnormal point clouds in space are identified.

[0065] The first step is to form a dense dataset of point cloud data at each moment, which reflects the three-dimensional outline of the object.

[0066] Specifically, the DBSCAN clustering algorithm is used to cluster the point cloud data at each time step to obtain the clustering result at each time step. The data set consisting of the point cloud data in each cluster in the clustering result at each time step is recorded as the feature dataset. In the clustering result at each time step, one cluster corresponds to one feature dataset.

[0067] It should be noted that because there are many objects in a parking lot in real space, such as trees, vehicles, pedestrians, fences, etc., their sizes also vary when they are identified. When point cloud interference occurs, it is mainly caused by smaller objects that are independent individuals in space. Therefore, the clustered point cloud distribution will generally be isolated.

[0068] The second step involves analyzing the density and shape distribution characteristics of the point cloud data in the feature dataset at each time step, performing anomaly analysis on each feature dataset, assessing the possible anomalies in each feature dataset, and identifying the first suspected obstacle to appear in the point cloud anomaly at any historical time step.

[0069] As a concrete example, such as Figure 2 As shown, the method for obtaining the abnormal dataset at the initial historical moment can be implemented by steps S201 to S204.

[0070] Step S201: For any feature dataset, determine the first feature coefficient of the feature dataset based on the density value of the point cloud data per unit volume in the feature dataset.

[0071] Specifically, in this embodiment, the preset unit volume is 1. Taking the i-th feature dataset at the t-th historical time as an example, we obtain the density value of all point cloud data in the i-th feature dataset at the t-th historical time within each unit volume. The average density value within all unit volumes of the i-th feature dataset at the t-th historical time is used as the first feature coefficient of the i-th feature dataset at the t-th historical time. The first feature coefficient characterizes the average density distribution of the point cloud data in the feature dataset, reflecting the degree of point cloud aggregation of the corresponding object in the feature dataset. High density may correspond to small objects, such as tree branches or traffic cones.

[0072] Step S202: The volume of all point cloud data in the feature dataset in space is used as the second feature coefficient of the feature dataset.

[0073] As a concrete example, taking any feature dataset as an example, the volume can be calculated based on the number of all unit volumes contained in the i-th feature dataset at the t-th historical time. The method for calculating the volume is a well-known technique and will not be described in detail here.

[0074] The second characteristic coefficient characterizes the shape distribution of the point cloud within the feature dataset, reflecting the spatial occupancy of the corresponding object in the feature dataset. The small volume further corroborates the small size of the obstacle.

[0075] Step S203: Based on the negative correlation coefficient of the second feature coefficient and the first feature coefficient, determine the degree of abnormality of any feature dataset.

[0076] As a concrete example, the result of normalizing the ratio of the first feature coefficient to the second feature coefficient is taken as the anomaly performance degree of any feature dataset. The normalization method is a well-known technique; minimization normalization can be used, and will not be discussed further here.

[0077] The quantification process of the anomaly performance of each feature dataset comprehensively considers the point cloud density features and volume features within the dataset, providing a data foundation for subsequent judgment of interference anomalies of small objects. In other words, the quantification process of anomaly performance conforms to the characteristics of interference from small objects and can effectively identify small obstacles that are easily overlooked by traditional methods.

[0078] Step S204: Based on the degree of abnormality of each feature dataset at each historical moment, perform anomaly analysis on the feature dataset to obtain the anomaly dataset at the initial historical moment.

[0079] Specifically, feature datasets with an anomaly intensity greater than a preset anomaly threshold are categorized as suspected anomaly feature sets. In this embodiment, the anomaly threshold is set to 0.6, but the implementer can set it according to the specific implementation scenario. The higher the anomaly intensity value of each feature dataset at each historical moment, the more the density and volume characteristics of the corresponding feature dataset match the characteristics of small objects, and the more likely the actual object represented by the corresponding feature dataset is to interfere with the parking space recognition process of automatic parking.

[0080] Therefore, the abnormality performance can be used to filter out datasets with point cloud anomalies, that is, suspected anomaly feature sets. Treating isolated small object clusters as anomalous point clouds can effectively identify small obstacles in parking lots that cause interference, such as tree branches, ice cream cones, etc.

[0081] Furthermore, the first historical moment corresponding to all suspected abnormal feature sets is taken as the initial historical moment, and each suspected abnormal feature set of the initial historical moment is the abnormal dataset of the initial historical moment.

[0082] The suspected anomaly feature set represents the point cloud data set corresponding to small objects that may be subject to interference. By tracking the suspected anomaly feature sets at all historical moments within the current time window, the first small object to exhibit interference is identified—that is, the first historical moment containing the suspected anomaly feature set—as the initial historical moment. The initial historical moment represents the time node where the interference first occurred. Furthermore, by locating each potentially anomaly position at the initial historical moment and tracking it at every moment after the initial historical moment up to the current moment, the short-term changes of the anomalous obstacle can be analyzed, providing data support for subsequent data filtering processes.

[0083] It should be noted that this embodiment only performs feature analysis on cases where abnormal datasets can be detected in historical moments prior to the current moment. If abnormal datasets from the initial historical moments cannot be obtained, it means that there are no interfering factors at the current moment. Therefore, it is not necessary to adaptively determine the filter kernel size. The filtering operation of the point cloud data can be performed directly based on empirical values ​​to remove the influence of basic noise.

[0084] Step S300: Based on the abnormal performance of each point cloud data in the feature dataset, and combined with the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset, the interference performance level of each feature dataset at each time moment is obtained.

[0085] In real-world applications, LiDAR is not very sensitive to smaller objects. For example, during automatic parking, if cars are parked in both parking spaces, LiDAR, in conjunction with millimeter-wave radar, can easily identify obstacles without issuing incorrect commands. However, when smaller objects are present, such as twigs or traffic cones, the accuracy decreases. This is because the laser beam diffuses with distance, diluting the reflected signal from distant objects. Furthermore, suspended particles in rain or fog scatter the laser, causing a sharp drop in the signal-to-noise ratio of the echo from small objects. Therefore, any abnormal point cloud should be monitored.

[0086] Furthermore, to overcome the drawbacks of lidar, multi-sensor fusion analysis is employed in practice, including lidar, millimeter-wave radar, and visual cameras. This multi-sensor fusion technology simultaneously analyzes obstacles in the surrounding environment. Therefore, to improve the accuracy of the identification results, when lidar identifies abnormal point clouds, calibration is performed based on their spatial location, thus combining this with data collected by millimeter-wave radar for joint analysis.

[0087] As a concrete example, such as Figure 3 As shown, the method for obtaining the degree of interference can be implemented by steps S301 to S304.

[0088] It should be noted that, in order to measure the consistency of anomalies in the point cloud identified by the LiDAR across multi-sensor data, the anomaly point cloud needs to be mapped to the same coordinate system as other dimensions for distance analysis. When the distance analysis results between the LiDAR dimension and the millimeter-wave radar dimension are relatively close, it indicates that the LiDAR and millimeter-wave radar have higher consistency in detecting the same object, effectively fusing multi-source data and reducing misjudgments. The specific process is as shown in steps S301 and S302, which realizes the spatial consistency calibration of the object position represented by the point cloud data in the LiDAR dimension and the object position represented by the millimeter-wave radar dimension.

[0089] Step S301: For any feature dataset, based on the position distribution of the center points of all point cloud data in the lidar coordinate system, map the center points of the feature dataset from the lidar coordinate system to the millimeter-wave radar coordinate system to obtain the first dimension data points corresponding to the feature dataset.

[0090] In this embodiment, the cluster center of all point cloud data in each feature dataset is taken as the center point of the corresponding feature dataset. The point cloud data represented by the center point characterizes the three-dimensional coordinate position of the center point of each feature dataset in the lidar coordinate system. Through coordinate transformation, the center point of each feature dataset can be transformed from the lidar coordinate system to the millimeter-wave radar coordinate system to obtain the first dimension data point corresponding to each feature dataset.

[0091] It should be noted that the center point represents the three-dimensional coordinate position of the point cloud data corresponding to the cluster center relative to the location of the lidar. After coordinate transformation, the first dimension data point represents the three-dimensional coordinate position of the point cloud data of the corresponding feature dataset cluster center relative to the location of the millimeter-wave radar at each time moment.

[0092] Step S302: Based on each millimeter-wave radar data at each time moment, determine the second-dimensional data point in the millimeter-wave radar coordinate system at each time moment.

[0093] Each millimeter-wave radar data point at each moment includes the straight-line distance and angle between the detected object and the millimeter-wave radar. Based on this millimeter-wave radar data, each millimeter-wave radar data point at each moment can be converted into three-dimensional coordinates in the millimeter-wave radar coordinate system, that is, the second-dimensional data point in the millimeter-wave radar coordinate system at each moment can be obtained.

[0094] It should be noted that there is a one-to-one correspondence between the second-dimensional data points and the millimeter-wave radar data. Therefore, each second-dimensional data point at any given time represents the three-dimensional coordinate position of each detected object relative to the location of the millimeter-wave radar. Furthermore, the timeframe in this step includes both historical and current moments.

[0095] Step S303: At the same time, obtain the spatial distance between the second-dimensional data point that is closest to the first-dimensional data point, as the degree of data difference of the feature dataset at the corresponding time.

[0096] As a concrete example, let's take the t-th historical moment as an example. For the first dimension data point of the i-th feature dataset at the t-th historical moment... Obtain the second-dimensional data point and the first-dimensional data point at the t-th historical time. The distance between them, and thus the minimum of all distances, is the degree of data difference of the i-th feature dataset at the t-th historical time. At the t-th historical time, it is related to the data points in the first dimension. The second-dimensional data point corresponding to the minimum value among all distances can be denoted as the second-dimensional data point. The distance can be determined by calculating the Euclidean distance based on the three-dimensional coordinates of the data points in the millimeter-wave radar coordinate system.

[0097] At the t-th historical moment, the first dimension data point With the second dimension data points The closer the distance between them, the higher the consistency of the data acquisition results in the lidar dimension and the millimeter-wave radar dimension, and the greater the possibility that they belong to the same object.

[0098] Step S304: The product of the negative correlation coefficient of the degree of data difference of the feature dataset and the degree of abnormal performance is taken as the degree of interference performance of any feature dataset.

[0099] As a specific example, this embodiment uses The function performs negative correlation processing on the degree of data difference. This represents an exponential function with base e. For exponents.

[0100] For any given moment, the higher the consistency between the data acquisition results of the i-th feature dataset at the t-th historical moment in both the LiDAR and millimeter-wave radar dimensions, the higher the probability of anomalies in this feature dataset, and the higher the likelihood that the object represented by this feature dataset exhibits interference. The degree of interference characterizes the spatial state of the corresponding feature dataset at each moment, reflecting the probability that the object represented by the LiDAR point cloud data exhibits interference after joint analysis by LiDAR and millimeter-wave radar at each moment (including historical moments and the current moment).

[0101] Step S400: Under the same spatial location corresponding to the abnormal dataset at the initial historical moment, track the change of the interference performance degree over time to obtain the noise impact coefficient at the current moment.

[0102] At the initial historical moment, each anomalous dataset is regarded as a source of interference. By analyzing the spatial state performance of the feature datasets at the same location as the interference source at each historical moment after the initial historical moment and at the current moment, the performance difference of the same interference source between adjacent moments is evaluated, persistent and non-persistent interference sources are distinguished, and the influence of the interference source is quantified.

[0103] As a concrete example, such as Figure 4 As shown, the method for obtaining the noise influence coefficient at the current moment can be implemented by steps S401 to S403.

[0104] Step S401: Based on the location of the center point of the abnormal dataset at the initial historical moment, match the feature dataset at each historical moment to obtain the matching dataset corresponding to each abnormal dataset at each historical moment and the current moment.

[0105] Specifically, for any anomalous dataset at the initial historical moment, the position of the center point of all point cloud data in the anomalous dataset in the LiDAR coordinate system is obtained as the target point position. At each historical moment after the initial historical moment and at the current moment, the feature dataset containing the point cloud data closest to the target point position is obtained, thus obtaining the matching dataset corresponding to the anomalous data at each historical moment and the current moment.

[0106] In this approach, the point cloud data corresponding to the cluster center of each outlier dataset is used as the centroid of the outlier dataset. The Euclidean distance between different points in the point cloud data can be calculated using the 3D coordinates of each point cloud data point; this is a well-known technique and will not be discussed further here.

[0107] Thus, for each anomalous dataset at the initial historical moment, there is a corresponding matching dataset for each historical moment after the initial historical moment, and the current moment also corresponds to a matching dataset. The target point location represents the location of the interference source represented by the corresponding anomalous dataset, and each matching dataset represents the location of the nearest object to the corresponding interference source location at each time moment.

[0108] Step S402: Based on the changes in the degree of interference performance of each abnormal dataset and all corresponding matching datasets, the influence degree of each abnormal dataset is obtained.

[0109] The first step is to construct a perturbation sequence for each anomalous dataset by combining the perturbation level of each anomalous dataset at the initial historical moment with the perturbation level of the corresponding matching dataset in chronological order.

[0110] The second step is to perform a persistent feature analysis on any interference source at the initial historical moment. That is, to distinguish between persistent and non-persistent interference sources by combining the performance differences between adjacent moments, taking into account the time-varying characteristics of dynamic interference in the parking lot.

[0111] Specifically, any two adjacent data values ​​in the interference sequence of any abnormal dataset are denoted as the first data value and the second data value, respectively. When the absolute value of the difference between the first data value and the second data value is less than or equal to a preset first threshold, the first preset value is used as the change influence factor between the first data value and the second data value; when the absolute value of the difference between the first data value and the second data value is greater than a preset second threshold, the second preset value is used as the change influence factor between the first data value and the second data value; wherein, the first preset value is greater than the second preset value, and the first threshold is less than the second threshold.

[0112] The first data value and the second data value characterize the spatial performance of the matching dataset corresponding to the same interference source at adjacent times. In this embodiment, the historical time corresponding to the first data value is before the historical time corresponding to the second data value. The historical time corresponding to the first data value is recorded as the first historical time, and the historical time corresponding to the second data value is recorded as the second historical time.

[0113] When the absolute value of the difference between the first data value and the second data value is much smaller than the first data value or the second data value, it indicates that the spatial behavior of the interference source corresponding to the second historical moment and the first historical moment has not changed significantly over time. In this case, the interference source continues to exist in these two adjacent moments. As a specific example, in this embodiment, the first threshold is set to 0.01 and the first preset value is set to 1. This indicates that when the difference between the first data value and the second data value is extremely small, it means that the interference source still exists in the corresponding adjacent moments. That is, in the first historical moment and the second historical moment, the interference source is a persistent interference source. The greater the interference effect of a persistent interference source on the current moment, the greater the impact of the interference. Therefore, the corresponding change influence factor is set to 1.

[0114] When the absolute value of the difference between the first data value and the second data value is close to the first data value, it indicates that the spatial performance of the interference source at the second historical moment and the first historical moment has not changed significantly over time. In this case, the interference source does not persist between these two adjacent moments. As a specific example, in this embodiment, the second threshold is set to 90% of the first data value, and the first preset value is set to 0. This indicates that when the difference between the first data value and the second data value is extremely large and close to the first data value, it means that the interference source does not persist between the corresponding adjacent moments. That is, at the first historical moment and the second historical moment, the interference source is a non-persistent interference source. The non-persistent interference source has a small impact on the current moment, so the corresponding change influence factor is set to 0.

[0115] The third step is to calculate the cumulative value of all changing influencing factors in the interference performance sequence of the anomalous dataset to obtain the influence degree of any one of the anomalous datasets. The influence degree characterizes the strength of the interference source represented by each anomalous dataset at the initial historical time.

[0116] Step S403: The normalized result of the mean of the influence of all abnormal datasets at the initial historical time is used as the noise influence coefficient at the current time.

[0117] By analyzing the interference changes of each interference source at the initial historical moment over consecutive time intervals, the degree of interference from objects in the environment at the current moment is evaluated. This has a significant impact on denoising LiDAR point cloud data, requiring greater filtering to remove noise. Even at slow vehicle speeds, dynamic obstacles such as pedestrians and shopping carts may still move or disappear within a short period. Using a time window composed of multiple historical moments can effectively distinguish between strong and weak interference factors.

[0118] It should be noted that the normalization method is a well-known technique, and implementers can choose it according to the specific implementation scenario, without any restrictions.

[0119] Step S500: Determine the filter kernel size based on the noise impact coefficient, and perform denoising processing on the point cloud data at the current moment. The denoised data is used to identify available parking spaces for the vehicle to be analyzed.

[0120] The core objective of point cloud denoising is to preserve valid point clouds, such as parking lines and fixed obstacles, while filtering out noise, such as dynamic interference and sensor errors. Strengthening denoising for persistent interference sources aligns with the requirement of prioritizing stable interference in autonomous driving. The noise impact coefficient characterizes the influence of interference sources within a certain time window preceding the current analysis moment. A large noise impact coefficient indicates a significant amount of persistent interference in the environment at the current moment, suggesting a potentially high proportion of noise in the point cloud data, requiring stronger denoising efforts. Conversely, a small noise impact coefficient suggests a potentially low proportion of noise in the point cloud data at the current moment, necessitating reduced denoising to avoid losing details.

[0121] Specifically, the product of the noise impact coefficient at the current moment and the preset initial filter kernel size, rounded down, is used as the filter kernel size. In this embodiment, the preset initial filter kernel size is 20, which can be set by the implementer according to the specific implementation scenario. The rounding operation can be performed using an up-rounding function. The stronger the influence of the interference source during data analysis, the larger the Gaussian filter kernel size should be, resulting in more noise data being removed and more effective removal of interference factors. The weaker the influence of the interference source, the smaller the Gaussian filter kernel size should be, resulting in less noise data needing to be removed and more effective preservation of detailed data.

[0122] Furthermore, Gaussian filtering is applied to all point cloud data at the current moment using the filter kernel size to obtain denoised point cloud data, which is then used to identify available parking spaces. It should be noted that the denoising Gaussian kernel filtering method is a well-known technique and will not be described in detail here. Similarly, identifying available parking spaces using the denoised point cloud data is also a well-known technique and will not be elaborated upon further. Implementers can choose any method that achieves the desired result. This embodiment of the invention focuses on analyzing the process of filtering the collected data during automatic parking.

[0123] This embodiment achieves enhanced denoising in environments with strong interference and mild denoising in environments with weak interference by dynamically adjusting the filter kernel size. This avoids the limitations of fixed filter parameters and improves the environmental adaptability of the denoising algorithm. In other words, it automatically adjusts the filtering level in complex environments, strengthening filtering in cases of strong interference and preserving details in cases of weak interference, ultimately improving the robustness and accuracy of vacant parking space recognition.

[0124] In some embodiments, an vacant parking space identification device for an autonomous driving system is also provided, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the computer program is executed by the processor, it implements the steps of an vacant parking space identification method for an autonomous driving system.

[0125] like Figure 5 As shown, in some embodiments, an vacant parking space identification device for an autonomous driving system is also provided. This device is used to implement the steps of an autonomous driving system vacant parking space identification method. The vacant parking space identification device for the autonomous driving system includes:

[0126] The data acquisition module is used to acquire point cloud data and millimeter-wave radar data at each moment based on the vehicle-mounted lidar and millimeter-wave radar of the vehicle to be analyzed, wherein the moment includes the current moment and several historical moments.

[0127] The feature recognition module is used to cluster the point cloud data at each time moment to obtain the feature dataset at each time moment; and to obtain the anomaly dataset of the initial historical time moment based on the density of the point cloud data in the feature dataset and the spatial characteristics of the feature dataset.

[0128] The interference analysis module is used to obtain the degree of interference performance of each feature dataset at each time moment based on the abnormal performance of each point cloud data in the feature dataset and the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset.

[0129] The noise impact module is used to track the changes in the degree of interference of the feature datasets of each adjacent historical time before the current time, under the same spatial location corresponding to the abnormal datasets at the initial historical time, and to obtain the noise impact coefficient at the current time.

[0130] The parking space recognition module is used to determine the filter kernel size based on the noise impact coefficient, and to denoise the point cloud data at the current moment. The denoised data is used to identify available parking spaces for the vehicle to be analyzed.

[0131] Since an embodiment of a method for identifying available parking spaces in an autonomous driving system has already been described in detail, it will not be repeated here.

[0132] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for identifying available parking spaces in an autonomous driving system, characterized in that, The method includes the following steps: Based on the vehicle-mounted lidar and millimeter-wave radar of the vehicle to be analyzed, point cloud data and millimeter-wave radar data at each moment are acquired, wherein the moment includes the current moment and several historical moments. Clustering is performed on the point cloud data at each time step to obtain the feature dataset for each time step; based on the density of the point cloud data in the feature dataset and the spatial characteristics of the feature dataset, the anomaly dataset for the initial historical time step is obtained. Based on the abnormal behavior of each point cloud data in the feature dataset, and combined with the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset, the degree of interference performance of each feature dataset at each time moment is obtained. By tracking the change in the degree of interference over time in the same spatial location corresponding to the abnormal dataset at the initial historical moment, the noise impact coefficient at the current moment can be obtained. The filter kernel size is determined based on the noise impact coefficient, and the point cloud data at the current moment is denoised. The denoised data is then used to identify available parking spaces for the vehicle to be analyzed.

2. The method for identifying available parking spaces in an autonomous driving system according to claim 1, characterized in that, The process of obtaining the initial historical time-based outlier dataset based on the density of point cloud data in the feature dataset and the spatial characteristics of the feature dataset specifically includes: For any given feature dataset, the first feature coefficient of the feature dataset is determined based on the density value of the point cloud data per unit volume in the feature dataset; the volume of all point cloud data in the feature dataset in space is used as the second feature coefficient of the feature dataset. Based on the negative correlation coefficient of the second feature coefficient and the first feature coefficient, the abnormal performance degree of any feature dataset is determined; based on the abnormal performance degree of each feature dataset at each historical moment, anomaly analysis is performed on the feature dataset to obtain the abnormal dataset at the initial historical moment.

3. The method for identifying available parking spaces in an autonomous driving system according to claim 2, characterized in that, The step of performing anomaly analysis on the feature dataset based on the anomaly performance of each feature dataset at each historical moment to obtain the anomaly dataset at the initial historical moment specifically includes: The feature datasets corresponding to anomalies exceeding a preset anomaly threshold are denoted as suspected anomaly feature sets; the first historical moment corresponding to all suspected anomaly feature sets is taken as the initial historical moment, and each suspected anomaly feature set at the initial historical moment is the anomaly dataset at the initial historical moment.

4. The method for identifying available parking spaces in an autonomous driving system according to claim 2, characterized in that, The step of obtaining the interference level of each feature dataset at each time moment based on the abnormal behavior of each point cloud data in the feature dataset, combined with the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset, specifically includes: For any feature dataset, based on the position distribution of the center points of all point cloud data in the lidar coordinate system, the center points of the feature dataset are mapped from the lidar coordinate system to the millimeter-wave radar coordinate system to obtain the first dimension data points corresponding to the feature dataset. Based on each millimeter-wave radar data at each moment, determine the second-dimensional data point in the millimeter-wave radar coordinate system at each moment; At the same time, the spatial distance between the second-dimensional data point that is closest to the first-dimensional data point is obtained, which is used as the degree of data difference in the feature dataset at the corresponding time. The product of the negative correlation coefficient of the degree of data difference in the feature dataset and the degree of abnormal performance is taken as the degree of interference performance of any feature dataset.

5. The method for identifying available parking spaces in an autonomous driving system according to claim 3, characterized in that, The process of tracking the change in the degree of interference over time at the same spatial location corresponding to the abnormal dataset at the initial historical moment, and obtaining the noise impact coefficient at the current moment, specifically includes: Based on the location of the center point of the abnormal dataset at the initial historical moment, the feature dataset at each historical moment is matched to obtain the matching dataset corresponding to each abnormal dataset at each historical moment and the current moment. The degree of influence of each anomalous dataset is obtained by analyzing the changes in the degree of interference in each anomalous dataset and all corresponding matching datasets. The normalized result of the mean of the influence of all outlier datasets at the initial historical moment is used as the noise influence coefficient at the current moment.

6. The method for identifying available parking spaces in an autonomous driving system according to claim 5, characterized in that, The step of matching the feature dataset of each historical time with the center point of the abnormal dataset at the initial historical time to obtain the matching dataset corresponding to each abnormal dataset at each historical time and the current time specifically includes: For any anomalous dataset at the initial historical moment, the position of the center point of all point cloud data in the anomalous dataset in the lidar coordinate system is obtained as the target point position; At each historical moment after the initial historical moment and at the current moment, obtain the feature dataset of the point cloud data closest to the target point location, and obtain the matching dataset corresponding to the abnormal data at each historical moment and the current moment.

7. The method for identifying available parking spaces in an autonomous driving system according to claim 5, characterized in that, The degree of influence of each abnormal dataset is obtained by analyzing the changes in the interference performance of each abnormal dataset and all corresponding matching datasets. Specifically, this includes: The perturbation performance of each anomalous dataset at the initial historical moment and the perturbation performance of the corresponding matching dataset are arranged in chronological order to form a perturbation performance sequence for each anomalous dataset. In any abnormal dataset, any two adjacent data values ​​in the perturbation sequence are denoted as the first data value and the second data value, respectively. When the absolute value of the difference between the first data value and the second data value is less than or equal to a preset first threshold, the first preset value is used as the change influencing factor between the first data value and the second data value. When the absolute value of the difference between the first data value and the second data value is greater than the preset second threshold, the second preset value is used as the change influencing factor between the first data value and the second data value. Wherein, the first preset value is greater than the second preset value, and the first threshold is less than the second threshold; The degree of influence of any one of the abnormal datasets is obtained by summing the values ​​of all the changing influencing factors in the disturbance performance sequence of the abnormal dataset.

8. The method for identifying available parking spaces in an autonomous driving system according to claim 1, characterized in that, The step of determining the filter kernel size based on the noise impact coefficient specifically includes: The integer result of the product of the noise impact coefficient at the current moment and the preset initial filter kernel size is used as the filter kernel size.

9. A parking space identification device for an autonomous driving system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the available parking space identification method for an autonomous driving system as described in any one of claims 1-8.

10. A parking space identification device for an autonomous driving system, characterized in that, The device is used to implement the steps of the available parking space identification method of an autonomous driving system as described in any one of claims 1-8, wherein the available parking space identification device of the autonomous driving system comprises: The data acquisition module is used to acquire point cloud data and millimeter-wave radar data at each moment based on the vehicle-mounted lidar and millimeter-wave radar of the vehicle to be analyzed, wherein the moment includes the current moment and several historical moments. The feature recognition module is used to cluster the point cloud data at each time moment to obtain the feature dataset at each time moment; and to obtain the anomaly dataset of the initial historical time moment based on the density of the point cloud data in the feature dataset and the spatial characteristics of the feature dataset. The interference analysis module is used to obtain the degree of interference performance of each feature dataset at each time moment based on the abnormal performance of each point cloud data in the feature dataset and the differences in the millimeter-wave radar data corresponding to each point cloud data in the feature dataset. The noise impact module is used to track the changes in the degree of interference of the feature datasets of each adjacent historical time before the current time, under the same spatial location corresponding to the abnormal datasets at the initial historical time, and to obtain the noise impact coefficient at the current time. The parking space recognition module is used to determine the filter kernel size based on the noise impact coefficient, and to denoise the point cloud data at the current moment. The denoised data is used to identify available parking spaces for the vehicle to be analyzed.

Images