Method and system for operation detection of autonomous vehicle

Abstract

A method and system for operation detection of an unmanned vehicle. The method comprises: acquiring LiDAR data when an autonomous vehicle travels on an urban road surface (S001); on the basis of density anomaly values and structural anomaly values at target measuring points in an original point cloud, constructing a screening model for ghost measuring points, and setting a threshold to distinguish between the ghost measuring points and normal measuring points (S002); performing k-means clustering on the normal measuring points to obtain environment regions, and on the basis of the differences between the environment regions to which different measuring points belong, acquiring environmental regulation factors during ghost measuring point clustering (S003); optimizing a clustering model for the ghost measuring points by using ghost features, spatial Euclidean distances and the environmental regulation factors, and then obtaining all ghost regions by means of k-means clustering (S004); and sliding and translating each ghost region in point cloud data, and constructing a loss function on the basis of the matching relationship between the ghost regions and the environment regions, so as to eliminate pseudo measuring points, thereby implementing point cloud registration (S005). The loss of valid information can be avoided, so that point cloud data, which is collected by means of LiDAR, is of higher quality, and an autonomous vehicle has a higher sensing accuracy and operates in a more stable state.

Description

A method and system for detecting operation of an unmanned vehicle Technical Field

[0001] The present invention relates to the field of laser radar data processing technology, and in particular to a method and system for detecting the operation of an unmanned vehicle. Background Art

[0002] Unlike traditional cars, self-driving cars can autonomously plan routes, identify targets, and navigate based on their perceived environment. They use a variety of sensors, including lidar, cameras, radar, and ultrasonic sensors, to capture data about roads, vehicles, pedestrians, and other obstacles. This sensor data is processed and analyzed by the vehicle's internal computer system, which then generates driving decisions and controls the vehicle to perform actions such as acceleration, braking, and steering.

[0003] LiDAR signals provide self-driving cars with highly accurate environmental perception and are an essential and key technology in autonomous driving systems. During operation, LiDAR emits short pulses of laser light and records the time it takes for the beam to reflect back. After receiving and processing the signals, they are converted into digital data, providing highly accurate measurements of the target. LiDAR generates three-dimensional point cloud data based on the measurement signals. This point cloud represents the geometric shape and positional distribution of the target object in space, enabling the vehicle to accurately detect obstacles, locate, and navigate. However, when self-driving cars operate in various environments, such as urban areas, rural areas, and highways, they are inevitably subject to significant interference, which can distort LiDAR data. This can lead to overly sensitive commands when the control system experiences perception errors, making autonomous driving technology unstable.

[0004] Summary of the Invention

[0005] The present invention provides a method and system for detecting the operation of an unmanned vehicle to solve the problem that existing unmanned vehicles are inevitably subject to a large amount of interference when traveling in various environments such as cities, suburbs, and highways, resulting in distortion of lidar data. When the control system has perception errors, allergic commands are likely to occur, making the unmanned driving technology unstable.

[0006] The present invention provides a method and system for detecting the operation of an unmanned vehicle using the following technical solutions:

[0007] An embodiment of the present invention provides a method for detecting the operation of an unmanned vehicle, the method comprising the following steps:

[0008] Obtaining raw point cloud data from a laser radar, wherein the raw point cloud data from the laser radar includes a plurality of measurement points and a spatial Euclidean distance between each measurement point and a laser radar emission position, and using the spatial Euclidean distance as a measurement value of each measurement point;

[0009] Each measurement point in the original point cloud data is used as a target measurement point, and several adjacent measurement points of the target measurement point are used as neighboring measurement points of the target measurement point. The density anomaly value of the target measurement point is obtained according to the average spatial Euclidean distance between the neighboring measurement points of each target measurement point in the original point cloud data and the target measurement point. The structural anomaly value of the target measurement point is obtained according to the information entropy of the difference between the measurement values ​​of the neighboring measurement points and the target measurement point and the entropy limit of the polarization of the difference between the measurement values ​​of the neighboring measurement points and the target measurement point. The afterimage characteristic value of each target measurement point is obtained according to the density anomaly value and the structural anomaly value of the target measurement point. A threshold is set for the afterimage characteristic value to screen out afterimage measurement points and normal measurement points.

[0010] Normal measuring points are clustered to obtain multiple environmental regions. Then, the environmental adjustment factors in the clustering process of afterimage measuring points are obtained based on the absolute values ​​of the differences in the average measured values ​​of the environmental regions to which different measuring points belong and the absolute values ​​of the differences in the standard deviations of the measured values ​​of the neighboring measuring points of different measuring points. The clustering distance measurement model of the afterimage measuring points is optimized based on the environmental adjustment factors and the afterimage characteristic values. The afterimage measuring points are clustered based on the clustering distance measurement model to obtain multiple afterimage regions.

[0011] The afterimage area is slid and translated in the original point cloud data. A loss function is constructed based on the mean square error of the overlapping measurement points between the afterimage area and the ambient area during the sliding translation process and the proportion of the number of overlapping measurement points. The pseudo measurement points in the afterimage area are obtained according to the loss function and the pseudo measurement points are eliminated.

[0012] Furthermore, the specific calculation method of taking each measurement point in the original point cloud data as a target measurement point, taking several adjacent measurement points of the target measurement point as the neighboring measurement points of the target measurement point, obtaining the density outlier value of the target measurement point based on the average spatial Euclidean distance between the neighboring measurement points of each target measurement point in the original point cloud data and the target measurement point, and obtaining the structural outlier value of the target measurement point based on the information entropy of the difference between the measurement values ​​of the neighboring measurement points of the target measurement point and the target measurement point and the entropy limit of the polarization of the difference between the measurement values ​​of the neighboring measurement points of the target measurement point and the target measurement point is as follows:

[0013] Each measurement point in the original point cloud data is regarded as the target measurement point. L adjacent measurement points are counted from near to far according to the Euclidean distance from the target measurement point. These points are called the neighboring measurement points of the target measurement point.

[0014] Among them, ω1 represents the density anomaly value of the target measurement point, o represents the oth target measurement point among all target measurement points, N represents the total number of measurement points of the point cloud data collected by the unmanned vehicle at the current moment, i represents the ith neighboring measurement point of the oth target measurement point, L is the number of neighboring measurement points of the target measurement point, L o Represents the number of neighboring measurement points of the oth target measurement point among all target measurement points, d iRepresents the spatial Euclidean distance between the i-th measuring point in the neighborhood of the target measuring point and the target measuring point;

[0015] ω2 represents the structural anomaly value of the target measurement point, h i represents the difference between the measured values ​​of the target point and the i-th neighboring point, and v represents any type of h i Value, the class represents h i The value is the same, R is the total number of neighboring measurement points of the target measurement point i Number of value classes, G v (h i ) represents the vth class h i The number of values, log22 represents the entropy limit of the polarization of the difference between the neighboring measurement points of the target measurement point and the measurement value of the target measurement point.

[0016] Furthermore, the afterimage characteristic value of each target measuring point is obtained according to the density anomaly value and the structural anomaly value of the target measuring point, a threshold is set for the afterimage characteristic value, and the specific calculation method for screening the afterimage measuring points and normal measuring points is as follows: P = ω1 × ω2

[0017] Among them, ω1 represents the density anomaly value of the target measurement point, ω2 represents the structural anomaly value of the target measurement point, and P represents the afterimage characteristic value of the target measurement point;

[0018] A ghost measurement point threshold is preset. When the ghost characteristic value of the target measurement point is greater than or equal to the threshold, the target measurement point is a ghost measurement point. When the ghost characteristic value of the target measurement point is less than the threshold, the target measurement point is a normal measurement point.

[0019] Furthermore, the clustering of normal measurement points to obtain multiple environmental areas includes the following specific steps:

[0020] Mark normal measuring points as Class A measuring points;

[0021] First, the k-means clustering algorithm is used to cluster all Class A measurement points. The clustering result is multiple Class A measurement point clusters. Then, the boundary function is used to obtain the boundary measurement points of each cluster to obtain multiple environmental areas.

[0022] Furthermore, the environmental adjustment factor in the afterimage measurement point clustering process is obtained according to the absolute value of the difference between the average measurement values ​​of the environmental areas to which different measurement points belong and the absolute value of the difference between the standard deviations of the measurement values ​​of the neighboring measurement points of different measurement points. The specific calculation method is as follows:

[0023] Mark the afterimage measurement point as a type B measurement point; ω(a, b) = exp[-(|σ′ a -σ′ b |×|μ″ a -μ″ b |)]

[0024] Where a and b represent any two Class B measurement points, ω(a, b) represents the environmental adjustment factors of the two measurement points a and b, and μ″ a , μ″ b Respectively represent the average measurement values ​​of the environmental areas to which the a and b measurement points belong, σ' a ,σ' b Represent the standard deviation of the measured values ​​of the neighboring measuring points of the a and b measuring points respectively.

[0025] Furthermore, the cluster distance measurement model for optimizing the afterimage measurement points according to the environmental adjustment factor and the afterimage characteristic value includes the following specific calculation method:

[0026] Mark the afterimage measurement points as Class B measurement points, where a and b represent any two Class B measurement points respectively;

[0027] Among them, sim(a,b) represents the cluster distance metric between two measurement points a and b, P a 、P b They represent the afterimage eigenvalues ​​of the a and b measuring points respectively, and d(a,b) represents the spatial Euclidean distance between the a and b measuring points.

[0028] Furthermore, clustering the afterimage measurement points according to the cluster distance measurement model to obtain multiple afterimage areas includes the following specific steps:

[0029] The clustering model is used as the distance metric in the k-means clustering algorithm. The optimal k value of the B-type measurement points is obtained according to the elbow method. The k value is input into the k-means clustering algorithm, and the B-type measurement points are clustered to obtain the clustering results of the B-type measurement points. The boundary function is used to obtain multiple cluster boundary measurement points in the clustering results, and multiple afterimage areas are obtained.

[0030] Furthermore, the afterimage area is slid and translated in the original point cloud data, and a loss function is constructed according to the mean square error of the overlapping measurement points between the afterimage area and the environment area during the sliding translation process and the proportion of the number of overlapping measurement points. The specific calculation method is as follows:

[0031] Slide and translate each afterimage area in the original point cloud data, and the sliding method is to traverse all positions in the original point cloud data;

[0032] Among them, E represents the loss function, p represents the pth residual image area, and q represents the qth environment area. Represents the measurement value of the εth measuring point in the pth afterimage area among all the overlapping measuring points of the pth afterimage area and the qth environmental area during the sliding process, represents the measurement value of the εth measurement point in the qth environmental area, J p represents the number of all overlapping measurement points between the pth afterimage area and the qth environment area, G p Represents the total number of B-type measurement points in the p-th afterimage area.

[0033] Furthermore, pseudo measurement points in the afterimage area are obtained according to the loss function, and the pseudo measurement points are eliminated to achieve point cloud registration. The specific steps include the following:

[0034] The afterimage area is slid on the original point cloud data. Each time the afterimage area slides, the loss function will obtain an output value. When the output value is minimum, the sliding is stopped. According to the position where the afterimage area stops on the original point cloud data, the reflection source area of ​​the afterimage area is obtained, and the original point cloud data in the reflection source area is obtained. The overlapping measurement points between the original point cloud data in the afterimage area after stopping sliding and the original point cloud data in the reflection source area are obtained. According to the overlapping measurement points, pseudo measurement points in the afterimage area are obtained, and the pseudo measurement points in the afterimage area are directly removed.

[0035] A self-driving car operation detection system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program executes and implements all the above methods.

[0036] The beneficial effects of the technical solution of the present invention are:

[0037] During the driving of an autonomous vehicle, the three-dimensional point cloud data collected by the LiDAR may exhibit blurred afterimages. Traditional filters and redundant sampling methods are unable to effectively remove point cloud afterimages. To address this technical problem, the present invention proposes an optimization method. First, a screening model for afterimage measurement points is constructed based on the density anomaly and structural anomaly characteristics at the target measurement points in the original point cloud. A threshold is then set to filter out afterimage measurement points. The afterimage measurement point screening model not only represents the number of measurement value classes within the afterimage region, but also effectively describes the distribution characteristics of the intersection and overlap of real and pseudo measurement points in the afterimage region. Then, based on the differences between the environmental regions to which different measurement points belong, an environmental adjustment factor is obtained during the clustering process of the afterimage measurement points. The clustering model of the afterimage measurement points is optimized based on this environmental adjustment factor to obtain the environmental region and the residual region. Finally, each afterimage region is slid and translated within the point cloud data. Based on the matching relationship between the afterimage region and the environmental region, a loss function is constructed to eliminate pseudo measurement points and achieve point cloud registration. The present invention does not require the use of filters to indiscriminately smooth point cloud data, thus avoiding the loss of effective information, and does not require multiple sampling and fusion to eliminate point cloud afterimages, thereby improving the quality of point cloud data collected by the lidar, improving the perception accuracy of driverless cars, and making the operating state more stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces 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 invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] FIG1 is a flowchart of the steps of a method for detecting operation of an unmanned vehicle according to the present invention. DETAILED DESCRIPTION

[0040] To further illustrate the technical means and effectiveness of the present invention in achieving its intended objectives, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effectiveness of a method and system for detecting the operation of an unmanned vehicle proposed in accordance with the present invention. 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.

[0041] 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 invention belongs.

[0042] The specific scheme of the unmanned vehicle operation detection method and system provided by the present invention is described in detail below with reference to the accompanying drawings.

[0043] Please refer to FIG1 , which shows a flowchart of a method for detecting the operation of an unmanned vehicle provided by one embodiment of the present invention. The method includes the following steps:

[0044] S001. Obtain lidar data of an autonomous vehicle driving on urban roads.

[0045] An unmanned driving experiment was conducted on urban roads, with the car traveling at a constant speed of 40Lm / h. During the driving process, the lidar three-dimensional point cloud data was continuously collected and recorded as raw point cloud data. Each measuring point in the raw point cloud data contains its three-dimensional coordinates (X, Y, Z). The spatial Euclidean distance of each measuring point from the lidar emission position can be directly obtained as the measurement value of the measuring point (in this embodiment, the measurement value is rounded to one decimal place).

[0046] Urban environments have many reflective areas and dense electromagnetic interference, resulting in high noise levels in LiDAR point cloud data. Although LiDARs are typically equipped with multiple transmit and receive channels, enabling redundant sampling and multiple measurements, or using filters to reduce noise levels, these methods are effective for dealing with numerically abnormal noise. However, in urban environments, the large number of specular reflections can cause multiple specular reflections when the laser strikes the target object, resulting in residual images or overlap in the point cloud data. Conventional redundant sampling and filters cannot solve this problem and may even cause it to expand.

[0047] S002. Based on the density anomaly values ​​and structural anomaly values ​​at the target measuring points in the original point cloud, a screening model for afterimage measuring points is constructed, and a threshold is set to screen out afterimage measuring points and normal measuring points.

[0048] It should be noted that each measuring point in the afterimage area is called an afterimage measuring point, and the afterimage measuring points are further divided into pseudo measuring points and real measuring points.

[0049] Afterimages are areas where some false measurement points appear around real measurement points, creating a localized blur in the point cloud. Objects that create afterimages due to their reflective or transparent materials are called reflection sources. After the laser beam passes from the LiDAR to the transparent, reflective reflection source, it is reflected. Each reflection creates a false measurement point in the receiver.

[0050] Pseudo-measurement points formed by reflection from the same reflection source plane all have the same or similar reflection trajectories. Therefore, pseudo-measurement points appear in sheets in the laser dot matrix. This is also the reason why pseudo-measurement points always blur the dot matrix. When these pseudo-measurement points overlap with real measurement points, the afterimage area formed has the characteristics of increased measurement point density and polarized measurement values.

[0051] In this embodiment, each measurement point in the original point cloud data is regarded as a target measurement point. L adjacent measurement points are counted from near to far from the target measurement point, which are called neighboring measurement points of the target measurement point. In this embodiment, L is set to 8.

[0052] Then, based on the average spatial Euclidean distance between each target point and its neighboring points in the original point cloud data, the density anomaly value of each target point is obtained. Based on the information entropy of the difference between the measured values ​​of the target point and its neighboring points, and the polarized entropy limit of the difference between the measured values ​​of the target point and its neighboring points, the structural anomaly value of each target point is obtained. Specifically:

[0053] Among them, ω1 represents the density anomaly value at the target measurement point, o represents the oth target measurement point among all target measurement points, N represents the total number of measurement points of the original point cloud data collected by the unmanned vehicle at the current moment, i represents the i-th neighboring measurement point of the target measurement point, L is the number of neighboring measurement points of the target measurement point, and L o Represents the number of neighboring measurement points of the oth target measurement point among all target measurement points, d i Represents the spatial Euclidean distance between the i-th measuring point and the target measuring point;

[0054] Further, It represents the average spatial Euclidean distance between the L nearest neighboring measurement points of any target measurement point and the target measurement point. It is called the target measurement point's nearest neighbor spatial Euclidean distance. The smaller the value, the greater the density of measurement points here. It represents the average value of the Euclidean distances of the neighboring spaces of all target measuring points in the point cloud data. The Euclidean distance of the neighboring spaces of the target measuring point is compared with the average value of the Euclidean distances of the neighboring spaces of all target measuring points. The smaller the ratio, the higher the abnormal value of the density of neighboring measuring points at the target measuring point. It is to correct the logical relationship.

[0055] Among them, ω2 represents the structural anomaly value at the target measurement point, h i represents the difference between the measured values ​​of the i-th measuring point in the neighborhood of the target measuring point and the target measuring point, and v represents any type of h i Value, the class represents h i The value is the same, R is the total h of the L neighboring measurement points of the target measurement point i Number of value classes, G v (h i ) represents the vth class h i the number of values, Represents the vth class h i The value is at the target measuring point's neighboring measuring point h i The probability within the value set, log22 represents the logarithmic function with 2 as the base and 2 as the real number. It should be noted that log22 here represents h i The entropy limit when the total number of categories is 2.

[0056] It should be noted that is the information entropy of the difference between the target measuring point and the neighboring measuring point, log22 is the entropy limit of the polarization of the difference between the target measuring point and the neighboring measuring point, and the target measuring point L is the neighboring measuring point h i The information entropy of the value and the hypothesis h iThe entropy limit is used to calculate the ratio when the value category is 2. The closer the ratio is to 1, the more it means that there are two types of measurement values ​​within the L nearest neighbor range centered on the target measurement point, and the proportion of each measurement point is almost half and half, which is consistent with the polarized distribution characteristics of the measurement values ​​in the afterimage area. Therefore, the larger the structural outlier value.

[0057] h of the neighboring measuring point and the target measuring point i When the information entropy of the value approaches the entropy limit, it means that the measured values ​​in the neighboring measured points of the target measured point are bipolarized, and the distribution probabilities of the two types of measured values ​​are almost equal. Therefore, using the entropy limit as a method to extract the polarization characteristics of the measured values ​​of the neighboring measured points of the target measured point can not only show the number of measured value classes in the afterimage area, but also well describe the distribution characteristics of the intersection and overlap of real measured points and pseudo measured points in the afterimage area.

[0058] So far, the density outliers and structural outliers of all target measuring points are obtained.

[0059] Then, a screening model for afterimage measurement points is constructed based on density outliers and structural outliers: P = ω1 × ω2

[0060] Where P represents the afterimage characteristic value of the target measurement point.

[0061] After extracting the afterimage features from all target measurement points, a threshold is directly set to filter out measurement points with higher afterimage feature values. In this embodiment, the empirical threshold is set to 0.7. When the afterimage feature value P of the target measurement point is greater than or equal to 0.7, the target measurement point is considered a afterimage measurement point. Conversely, when P is less than or equal to 0.7, the target measurement point is considered a normal measurement point.

[0062] At this point, all measuring points are divided into normal measuring points and afterimage measuring points.

[0063] S003. Perform k-means clustering on the normal measurement points to obtain environmental regions. Based on the differences between the environmental regions to which different measurement points belong, obtain the environmental adjustment factors in the clustering process of the afterimage measurement points.

[0064] Furthermore, we can obtain a set of all afterimage measurement points. Those with afterimage eigenvalues ​​less than or equal to the threshold are marked as Class B measurement points, while those with afterimage eigenvalues ​​greater than the threshold are marked as Class A measurement points. Note that the afterimage area is formed by the superposition of reflected pseudo-measurement points and real measurement points, so Class B measurement points include both real and pseudo-measurement points.

[0065] The marked A and B measurement points are divided into two data layers and processed separately.

[0066] First, a k-means clustering algorithm is used to cluster all Category A measurement points. These points represent different environmental units around the autonomous vehicle, such as trees, pedestrians, and obstacles. Each measurement point in each environmental unit forms a region. The purpose of clustering the two types of points separately is to group point cloud data from different regions into multiple clusters, allowing us to trace the reflection source in the afterimage area.

[0067] Since Class A measurement points are normal measurement point data, clustering of Class A measurement points directly uses the spatial Euclidean distance between measurement points as the clustering metric parameter. The optimal k value for the Class A measurement point set is obtained using the elbow method. This optimal k value is then input into the k-means clustering algorithm to cluster all Class A measurement points. The clustering result is multiple Class A measurement point clusters. The boundary function is then used to enclose the boundaries of each cluster, resulting in multiple environmental regions. The boundary function is used to search for boundary points in a 3D point cloud and is well known in the art, so its detailed description is omitted.

[0068] Furthermore, the B-type measurement points are clustered. The k-means clustering process generally uses the distance between data points as the clustering metric parameter. However, the afterimage area is where the real measurement points and pseudo measurement points overlap and are mixed. The measurement values ​​of the real measurement points are true, but the measurement values ​​of the pseudo measurement points are false. The displayed measurement values ​​do not correspond to their spatial positions. Therefore, the distance metric cannot be directly used to classify the B-type measurement points.

[0069] It should be noted that there are two problems in the clustering process of Class B measurement points:

[0070] The first is that there may be large distance differences between adjacent Class B measurement points in the same environmental area. This problem can be solved by introducing the afterimage feature, because the afterimage feature contains information about neighboring measurement points. The overlapping real measurement points and pseudo measurement points can be directly corrected by the afterimage feature to measure the clustering distance.

[0071] The second is to assume that when the afterimage region exists at the boundary of two environmental regions, whether an afterimage region should be divided into two clusters considering the subsequent point cloud registration process.

[0072] This implementation obtains the environmental adjustment factor to solve the above problem, specifically: ω(a, b) = exp[-(|σ′ a -σ′ b |×|μ″ a -μ″ b |)]

[0073] Where a and b represent any two Class B measurement points, ω(a, b) represents the environmental adjustment factors of the two measurement points a and b, and μ″ a , μ″ bRespectively represent the average measurement values ​​of the environmental areas to which the a and b measurement points belong, σ' a ,σ' b They represent the standard deviations of the measured values ​​of the neighboring measuring points of the a-th and b-th measuring points, respectively. exp() represents an exponential function with natural numbers as the base. This embodiment adopts the exp(-x) model to present the inverse proportional normalization processing. x is the model input. The implementer can set the inverse proportional normalization function according to the actual situation.

[0074] Among them, |σ' a -σ' b | represents the absolute value of the difference between the standard deviations of the measurements of the two neighboring measurement points a and b; |μ″ a -μ″ b | represents the absolute value of the difference between the average measurement values ​​of the environmental areas to which the two measurement points a and b belong. |σ' a -σ' b |×|μ″ a -μ″ b | represents the absolute value of the difference between the standard deviations of the measurements of the neighboring points a and b, multiplied by the absolute value of the difference between the average measurements of the environmental areas to which the two measurement points a and b belong respectively, and then the product is inversely normalized using the exp(-x) model.

[0075] It should be noted that |σ' a -σ' b | value represents whether the overlapping information of the neighboring measurement points of points a and b is consistent. When the type of the neighboring measurement point overlapping with measurement point a is similar to or the same as the type of the neighboring measurement point overlapping with measurement point b, |σ' a -σ' b |Close to 0. |μ″ a -μ″ b The smaller the value of |, the more likely that a and b belong to the same environmental area, or to two similar environmental areas with small differences in X, Y, and Z coordinates. In this case, even if a and b belong to different environmental areas, the differences between the environmental areas are small, and the transition is smooth and will not affect the point cloud registration. On the contrary, when |μ″ is large, the point cloud registration will not be affected. a -μ″ b When | is larger, it means that a and b belong to different environmental areas, and the difference in X, Y, and Z coordinates of the two environmental areas is greater. In this case, even if a and b belong to the same reflection source, the a and b measurement points cannot be classified into the same cluster because the transition difference between the two environmental areas a and b is large, which will affect the subsequent point cloud registration process.

[0076] S004. Use the afterimage characteristics, spatial Euclidean distance, and environmental adjustment factors to optimize the clustering model of the afterimage measurement points, and then use k-means clustering to obtain all the afterimage areas.

[0077] Furthermore, after obtaining the environmental adjustment ω(a, b), the clustering model is optimized according to the afterimage characteristics and the environmental adjustment factor, specifically:

[0078] Among them, sim(a,b) represents the cluster distance metric between two measurement points a and b, P a 、P b They represent the afterimage eigenvalues ​​of the a and b measuring points respectively, and d(a,b) represents the spatial Euclidean distance between the a and b measuring points.

[0079] Among them, P a -P b Represents the difference in the afterimage characteristic values ​​of the two measurement points a and b, The Euclidean norm represents the difference between the residual image eigenvalues ​​and the spatial Euclidean distance between measurement points a and b. This Euclidean norm is divided by the environmental adjustment factor ω(a, b). The smaller the environmental adjustment factor, the larger the Euclidean norm of the numerator will be adjusted; conversely, the larger the environmental adjustment factor, the smaller the Euclidean norm of the numerator will be adjusted. The environmental adjustment factor is used to adaptively adjust the measurement parameters between Class B measurement points under different environments, thereby interfering with the clustering results.

[0080] Furthermore, after obtaining the clustering distance measurement model of Class B measurement points, the optimal k value is obtained according to the elbow method, and k-means clustering is performed on the Class B measurement points to obtain the clustering results of Class B measurement points. The boundary function is used to enclose the boundaries of multiple clusters in the clustering results to obtain multiple afterimage areas.

[0081] S005. Slide and translate each afterimage area in the point cloud data, and construct a loss function based on the matching relationship between the afterimage area and the environment area, thereby eliminating pseudo measurement points and realizing point cloud registration.

[0082] The afterimage area is the area where the pseudo-measurement points and the real measurement points overlap, where the pseudo-measurement points are obtained by reflection from a real environment area. Therefore, nearly half of the measurement points in each afterimage area can be highly consistent with the environment area. Therefore, the afterimage area is slid and matched in the original point cloud data. The sliding method is to traverse all positions in the original point cloud data, and the pseudo-measurement points are screened according to the overlap of the afterimage area with the environment area in the original point cloud data during the sliding process. It is a well-known technology in this field to realize point cloud registration by sliding translation and rotation in the three-dimensional point cloud space, and the pseudo-measurement points cannot be rotated when the reflection is formed, so I will not go into details here.

[0083] Construct the loss function in the sliding matching process of the afterimage area, specifically:

[0084] Among them, p represents the pth afterimage area, q represents the qth environment area, Represents the measurement value of the εth measuring point in the pth afterimage area among all the overlapping measuring points of the pth afterimage area and the qth environmental area during the sliding process, represents the measurement value of the εth measurement point in the qth environmental area, J p represents the number of all overlapping measurement points between the pth afterimage area and the qth environment area, G p Represents the total number of Class B measurement points in the pth afterimage area. It should be noted that when the afterimage area does not overlap with any surrounding area during the sliding process, the surrounding area with the closest Euclidean distance to the afterimage area is selected to output the loss function.

[0085] in, Represents the mean square error of the measured values ​​between overlapping measurement points when the pth afterimage area slides on the qth environment area, The ratio of the difference between the measured values ​​of the εth overlapping measurement point in the p-th and q-th regions to the measured value of the εth overlapping measurement point in the q-th region is only a normalization calculation. represents the ratio of the number of overlapping measurement points in the p-th and q-th regions to the total number of measurement points in the p-th region, and the constant The absolute value of the difference.

[0086] Furthermore, when the measurement points in the pth afterimage area overlap with the measurement points in the qth environment area during the sliding translation process, the mean square error between the overlapping measurement points is calculated. The smaller the mean square error, the higher the matching degree. The number of overlapping measurement points needs to be close to the total number of measurement points in the afterimage area. Therefore, As a penalty term, the mean square error and the penalty term have the same dimension.

[0087] Furthermore, each time the afterimage area slides, the loss function will output an E value. When the E value is minimum, it means that the objective function converges. At this time, the position where the afterimage area stays in the original point cloud data is the reflection source of the afterimage area.

[0088] The convergence of the Loss function indicates that the matching degree of the overlapping measurement points in the afterimage area is the highest at this time. The overlapping measurement points are the pseudo measurement points that cause the point cloud afterimage. By directly removing the overlapping measurement points, the point cloud registration can be achieved and the point cloud afterimage data can be eliminated.

[0089] After eliminating point cloud afterimages, the data collected by the LiDAR about the autonomous vehicle's surroundings is more accurate. By inputting this point cloud data into the trained PointNet neural network, obstacles can be better identified while the autonomous vehicle is in motion. Eliminating point cloud afterimages can also prevent allergic commands during vehicle operation. It should be noted that this embodiment aims to optimize the collection quality of point cloud data from the LiDAR of autonomous vehicles, providing reliable data support for obstacle identification, environmental monitoring, and path planning, thereby making the operation of autonomous vehicles more stable. Specific obstacle identification methods and other related techniques are well-known technologies and means in the autonomous driving field and will not be elaborated upon in this invention.

[0090] This embodiment also provides an unmanned vehicle operation detection system, including:

[0091] The memory, processor, and computer program stored in the memory and executable on the processor are used to execute the method of steps S001 to S005 of this embodiment. The specific processing logic of the computer program has been described in detail above and will not be described again.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting the operation of an unmanned vehicle, characterized in that: The method comprises the following steps: Acquire original point cloud data of the laser radar, wherein the original point cloud data of the laser radar includes a plurality of measuring points and a spatial Euclidean distance between each measuring point and a transmitting position of the laser radar, and use the spatial Euclidean distance as a measurement value of each measuring point; Each measuring point in the original point cloud data is taken as a target measuring point, and several adjacent measuring points of the target measuring point are taken as neighbor measuring points of the target measuring point. The density anomaly value of the target measuring point is obtained according to the average spatial Euclidean distance between the neighbor measuring points of each target measuring point in the original point cloud data and the target measuring point. The structural anomaly value of the target measuring point is obtained according to the information entropy of the difference between the measured values ​​of the neighbor measuring points of the target measuring point and the target measuring point and the polarized entropy limit of the difference between the measured values ​​of the neighbor measuring points of the target measuring point and the target measuring point. The residual image characteristic value of each target measuring point is obtained according to the density anomaly value and the structural anomaly value of the target measuring point, a threshold is set for the residual image characteristic value, and residual image measuring points and normal measuring points are obtained by screening; The normal measuring points are clustered to obtain multiple environmental areas. Then, the environmental adjustment factors in the clustering process of the afterimage measuring points are obtained according to the absolute values ​​of the differences in the average measured values ​​of the environmental areas to which the different measuring points belong and the absolute values ​​of the differences in the standard deviations of the measured values ​​of the neighboring measuring points of the different measuring points. The clustering distance measurement model of the afterimage measuring points is optimized according to the environmental adjustment factors and the afterimage characteristic values. The afterimage measuring points are clustered according to the clustering distance measurement model to obtain multiple afterimage areas. The afterimage area is slid and translated in the original point cloud data, and a loss function is constructed according to the mean square error of the overlapping measurement points between the afterimage area and the environment area during the sliding translation process and the proportion of the number of overlapping measurement points. According to the loss function, pseudo measurement points in the afterimage area are obtained, the pseudo measurement points are eliminated, and the operation detection of the unmanned vehicle is performed based on the original point cloud data after eliminating the pseudo measurement points.

2. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The specific calculation method of taking each measuring point in the original point cloud data as a target measuring point, taking several adjacent measuring points of the target measuring point as neighbor measuring points of the target measuring point, obtaining the density anomaly value of the target measuring point according to the average spatial Euclidean distance between the neighbor measuring points of each target measuring point in the original point cloud data and the target measuring point, and obtaining the structural anomaly value of the target measuring point according to the information entropy of the difference between the measured values ​​of the neighbor measuring points of the target measuring point and the target measuring point and the entropy limit of the polarization of the difference between the measured values ​​of the neighbor measuring points of the target measuring point and the target measuring point is as follows: Each measuring point in the original point cloud data is taken as the target measuring point, and L adjacent measuring points are counted from near to far according to the spatial Euclidean distance from the target measuring point, which are called the neighboring measuring points of the target measuring point; Among them, ω1 represents the density outlier of the target measurement point, o represents the oth target measurement point among all target measurement points, and N represents The total number of measurement points of the point cloud data collected by the driverless car at the current moment, i represents the i-th neighboring measurement point of the o-th target measurement point, L is the number of neighboring measurement points of the target measurement point, and L o Represents the number of neighboring measurement points of the oth target measurement point among all target measurement points, d i Represents the spatial Euclidean distance between the i-th measuring point in the neighborhood of the target measuring point and the target measuring point; ω2 represents the structural anomaly value of the target measurement point, h i represents the difference between the measured values ​​of the target point and the i-th neighboring point, and v represents any type of h i value, the class represents h i The value is the same, R is the total number of neighboring measurement points of the target measurement point i Number of value classes, G v (h i ) represents the vth class h i The number of values, log22 represents the entropy limit of the polarization of the difference between the neighboring points of the target point and the measured value of the target point.

3. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The specific calculation method of obtaining the residual image characteristic value of each target measuring point according to the density abnormal value and the structural abnormal value of the target measuring point, setting a threshold for the residual image characteristic value, and screening the residual image measuring points and normal measuring points is as follows: P = ω1 × ω2 Among them, ω1 represents the density anomaly value of the target measurement point, ω2 represents the structural anomaly value of the target measurement point, and P represents the residual image characteristic value of the target measurement point; A threshold value of the afterimage measuring point is preset. When the afterimage characteristic value of the target measuring point is greater than or equal to the threshold value, the target measuring point is the afterimage measuring point. When the afterimage characteristic value of the target measuring point is less than the threshold value, the target measuring point is the normal measuring point.

4. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The clustering of normal measurement points to obtain multiple environmental areas includes the following specific steps: Mark normal measuring points as Class A measuring points; First, the k-means clustering algorithm is used to cluster all Class A measurement points. The clustering result is multiple Class A measurement point clusters. Then, the boundary function is used to obtain the boundary measurement points of each cluster to obtain multiple environmental areas.

5. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The environmental adjustment factor in the process of clustering the residual image measuring points is obtained according to the absolute value of the difference between the average measured values ​​of the environmental areas to which the different measuring points belong and the absolute value of the difference between the standard deviations of the measured values ​​of the neighboring measuring points of the different measuring points, including the specific calculation method as follows: Mark the afterimage measurement point as a Class B measurement point; ω(a,b)=exp[-(|σ′ a -s′ b |×|μ″ a -m″ b |)] Where a and b represent any two Class B measuring points, ω(a, b) represents the environmental adjustment factor of the two measuring points a and b, μ″ a , μ″ b Respectively represent the average measurement values ​​of the environmental areas to which the a and b measurement points belong, σ' a ,σ' b Respectively represent the standard deviation of the measured values ​​of the neighboring measuring points of the a and b measuring points.

6. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The cluster distance measurement model for optimizing the afterimage measurement points according to the environmental adjustment factor and the afterimage characteristic value includes the following specific calculation method: Mark the afterimage measurement points as Class B measurement points, and a and b represent any two Class B measurement points respectively; Among them, sim(a,b) represents the cluster distance metric between two measurement points a and b, P a , P b They represent the afterimage eigenvalues ​​of the a-th and b-th measuring points respectively, and d(a,b) represents the spatial Euclidean distance between the a-th and b-th measuring points.

7. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The method of clustering the afterimage measurement points according to the clustering distance measurement model to obtain multiple afterimage areas includes the following specific steps: The clustering model is used as the distance metric in the k-means clustering algorithm, and the optimal k value of the B-type measurement point is obtained according to the elbow method. The k value is input into the k-means clustering algorithm, and the B-type measurement points are clustered to obtain the clustering results of the B-type measurement points. The boundary function is used to obtain multiple cluster boundary measurement points in the clustering results to obtain multiple afterimage areas.

8. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The afterimage area is slid and translated in the original point cloud data, and a loss function is constructed according to the mean square error of overlapping measurement points between the afterimage area and the environment area during the sliding translation process and the proportion of the number of overlapping measurement points. The specific calculation method is as follows: Each residual image area is slid and translated in the original point cloud data, and the sliding method is to traverse all positions in the original point cloud data; Among them, E represents the loss function, p represents the pth residual image area, and q represents the qth environment area. Represents the measurement value of the εth measuring point in the pth afterimage area among all the overlapping measuring points between the pth afterimage area and the qth environmental area during the sliding process, represents the measured value of the εth measuring point in the qth environmental area, J p represents the number of all overlapping measurement points between the pth afterimage area and the qth environment area, G p Represents the total number of type B measurement points in the pth afterimage area.

9. The method for detecting the operation of an unmanned vehicle according to claim 1, characterized in that: The specific steps of obtaining the pseudo measurement points in the residual image area according to the loss function and eliminating the pseudo measurement points are as follows: Slide the afterimage area on the original point cloud data. Each time the afterimage area slides, the loss function will get an output value. Stop sliding when the output value is the smallest. According to the position where the afterimage area stays on the original point cloud data, get the reflection source area of ​​the afterimage area, get the original point cloud data in the reflection source area, and get the original point cloud in the afterimage area after stopping sliding. The overlapping measurement points between the original point cloud data and the reflection source area are obtained, and the pseudo measurement points in the afterimage area are obtained according to the overlapping measurement points, and the pseudo measurement points in the afterimage area are directly removed.

10. An unmanned vehicle operation detection system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the unmanned vehicle operation detection method as described in any one of claims 1 to 9 are implemented.

Images