Laser radar anti-environmental interference identification method and system based on machine learning

Abstract

The invention provides a laser radar anti-environmental interference identification method and system based on machine learning. The method comprises the following steps: respectively acquiring echoes of environmental interference and a target, performing feature value extraction and data marking on an acquired one-dimensional high-resolution range profile, performing dimensionality reduction on manually extracted feature values, and performing model training by using a support vector machine algorithm on the basis of classified data feature values to obtain an optimal classifier model. And the model is applied to a laser radar real-time detection algorithm to realize real-time resolution of the target and interference. According to the method, on the basis of a low-cost one-dimensional laser radar, effective distinguishing of natural environment interference can be achieved through software upgrading, and iteration can be continuously optimized and upgraded through data set expansion. The data training process is carried out in the background, and the method has the advantages of high speed and high real-time performance in practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of interference identification, and in particular to a laser radar anti-environmental interference identification method and system based on machine learning. Background Art

[0002] Pulse LiDAR is mainly based on the principle of microwave radar ranging, namely time-of-flight technology (TOF). LiDAR emits a laser signal, which is reflected by the target and returns to the receiving port. Based on the time interval between sending and receiving pulses, the distance between the radar and the target can be inferred. Compared with microwave radar, LiDAR has the advantages of good directionality and is not easily affected by electromagnetic interference. However, LiDAR is also easily affected by natural environment interference, such as clouds, fog, rain, snow and other natural weather, as well as smoke, which will affect the detection performance of LiDAR.

[0003] Traditional one-dimensional laser radar is low-cost and small in size, but because it can only collect distance and amplitude information, it usually distinguishes between physical targets and environmental interference such as clouds and fog through the corresponding relationship between amplitude and distance. In addition, if the distance resolution of the laser radar is improved, the target can be identified from the waveform characteristics. However, due to the limitations of the laser emitter and detector capabilities, the echo morphology of environmental interference may also be close to the target echo. The use of artificial experience values ​​to distinguish waveform characteristics is inevitably prone to large errors, and the empirical feature value resolution method is difficult to adapt to the expanded data set. If the echo of the target and the cloud and fog environment cannot be distinguished, it is easy to misjudge the distance to the target.

[0004] In addition, in order to improve the anti-interference performance of lidar, coherent lidar, lidar imaging radar and other solutions can also be used based on the traditional one-dimensional lidar. However, these methods require the use of complex laser transmitting and receiving systems, which have the disadvantages of high cost and large size. Summary of the invention

[0005] The purpose of the present invention is to provide a laser radar anti-environmental interference identification method and system based on machine learning. To solve the problems described in the background technology, the specific technical solution is as follows:

[0006] In a first aspect, the present invention provides a laser radar anti-environmental interference identification method based on machine learning, the method comprising:

[0007] Classify and mark the interference data and target data, and perform feature value extraction, normalization, and dimensionality reduction processing to determine the reduced dimension feature value and marked data;

[0008] Based on the reduced dimension feature values ​​and the labeled data, an optimal classification hyperplane in the machine learning model is determined by training in the machine learning model; the classification hyperplane is a function that classifies the reduced dimension feature values ​​according to a rule, the rule being that the minimum distance interval from the points in the data set to the hyperplane is the largest, and the optimal classification hyperplane is the optimal solution among the multiple classification hyperplanes;

[0009] Based on the principal component matrix and the optimal classification hyperplane, the received lidar echo data is classified and predicted to identify the detection results of the lidar, and the detection results include: identification as a target and identification as interference; the principal component matrix is ​​a matrix of the reduced dimensionality eigenvalue vectorization, which is used to perform dimensionality reduction processing on the lidar echo data.

[0010] Furthermore, the interference data and the target data are classified and marked, and feature value extraction, normalization, and dimension reduction are performed to determine the dimension reduction feature value and the marked data, including:

[0011] Classify and mark the collected interference data and target data and extract characteristic values ​​to determine the characteristic value data of the one-dimensional range image;

[0012] Normalizing the characteristic value data of the one-dimensional range image to determine normalized data;

[0013] The normalized data is subjected to dimensionality reduction processing to determine dimensionality reduction feature values ​​and labeled data.

[0014] Furthermore, before classifying and marking the interference data and the target data and performing feature value extraction, normalization, and dimension reduction, the method further includes:

[0015] Collect environmental echoes in a complex environment to determine the interference data, wherein the environmental echoes include echoes of clouds, fog, rain, snow, and smoke;

[0016] Target echoes in different scenarios are collected to determine the target data.

[0017] Further, the training in the machine learning model based on the dimension reduction feature value and the labeled data to determine the optimal classification hyperplane in the machine learning model includes:

[0018] Determine the purpose of training the support vector machine model in the machine learning, that is, the objective function of the training;

[0019] Adding relaxation term parameters to the constraint conditions and adding penalty terms to the objective function;

[0020] Mapping the reduced-dimensional feature values ​​and the labeled data into a high-dimensional space through a kernel function;

[0021] Training is performed in the high-dimensional spatial data environment to determine the optimal classification hyperplane.

[0022] Furthermore, the classification prediction of the received laser radar echo data based on the principal component matrix and the optimal classification hyperplane includes:

[0023] Perform eigenvalue processing on the laser radar echo data collected in real time to determine the laser eigenvalue;

[0024] Performing dimensionality reduction processing on the laser eigenvalues ​​through the principal component matrix to obtain laser dimensionality reduction data;

[0025] Normalizing the laser dimension reduction data to determine a data matrix;

[0026] The support vector and kernel function are used to determine the distance from the data matrix to the optimal classification hyperplane, and classification prediction is performed based on the distance. If the distance is positive, it is identified as a target, and if the distance is negative, it is identified as interference.

[0027] In a second aspect, the present invention also provides a laser radar anti-environmental interference identification system based on machine learning, the system comprising: a pre-training processing module, a training module, and a prediction module;

[0028] The pre-training processing module is used to classify and mark the interference data and the target data and perform feature value extraction, normalization, and dimension reduction processing to determine the dimension reduction feature value and mark the data;

[0029] The training module is used to train and determine the optimal classification hyperplane in the machine learning model based on the reduced dimension feature value and the labeled data; the classification hyperplane is a function that classifies the reduced dimension feature value according to a rule, and the rule is that the minimum distance interval from the point in the data set to the hyperplane is the largest, and the optimal classification hyperplane is the optimal solution among multiple classification hyperplanes;

[0030] The prediction module is used to classify and predict the received lidar echo data based on the principal component matrix and the optimal classification hyperplane, and identify the detection results of the lidar, and the detection results include: identification as a target and identification as interference; the principal component matrix is ​​the matrix of the dimensionality reduction eigenvalue vectorization, which is used to perform dimensionality reduction processing on the lidar echo data.

[0031] In another embodiment of the second aspect, the pre-training processing module is specifically used to:

[0032] Classify and mark the collected interference data and target data and extract characteristic values ​​to determine the characteristic value data of the one-dimensional range image;

[0033] Normalizing the characteristic value data of the one-dimensional range image to determine normalized data;

[0034] The normalized data is subjected to dimensionality reduction processing to determine dimensionality reduction feature values ​​and labeled data.

[0035] In another embodiment of the second aspect, the system further includes: an acquisition module;

[0036] The acquisition module is used to collect environmental echoes in a complex environment to determine the interference data, wherein the environmental echoes include echoes of clouds, fog, rain, snow, and smoke; and collect target echoes in different scenes to determine the target data.

[0037] In another embodiment of the second aspect, the training module is specifically used to:

[0038] Determine the purpose of training the support vector machine model in the machine learning, that is, the objective function of the training;

[0039] Adding relaxation term parameters to the constraint conditions and adding penalty terms to the objective function;

[0040] Mapping the reduced-dimensional feature values ​​and the labeled data into a high-dimensional space through a kernel function;

[0041] Training is performed in the high-dimensional spatial data environment to determine the optimal classification hyperplane.

[0042] In another embodiment of the second aspect, the prediction module is specifically used to:

[0043] Perform eigenvalue processing on the laser radar echo data collected in real time to determine the laser eigenvalue;

[0044] Performing dimensionality reduction processing on the laser eigenvalues ​​through the principal component matrix to obtain laser dimensionality reduction data;

[0045] Normalizing the laser dimension reduction data to determine a data matrix;

[0046] The support vector and kernel function are used to determine the distance from the data matrix to the optimal classification hyperplane, and classification prediction is performed based on the distance. If the distance is positive, it is identified as a target, and if the distance is negative, it is identified as interference.

[0047] The beneficial effects of the present invention are as follows:

[0048] The laser radar anti-environmental interference identification method and system based on machine learning provided by the present invention have the following advantages:

[0049] 1. Without increasing the hardware cost, the support vector machine machine learning algorithm is used to identify interference from the echoes collected by the one-dimensional lidar.

[0050] Support vector machines are used to solve binary classification problems. For any set of data, there are two types of echo data, target and interference. The support vector machine maps the data to points in space, and there are countless intervals that can separate the two types of data (target and interference). By finding one of the intervals, the minimum distance from the points on both sides of the interval to the interval is maximized, that is, finding the interval that can maximally distinguish the two types of data. Figure 1 As shown, for two groups of point sets, intervals A, B, C, and D can all distinguish them, and interval C is the interval that can separate the two types of data with the maximum distance.

[0051] The lidar cloud and fog interference threshold calculation method based on support vector machine can distinguish targets and interference in lidar echoes only through software algorithms without making any hardware changes, greatly reducing labor costs and improving the lidar's target recognition capability and distance measurement accuracy.

[0052] 2. The recognition model can be optimized as the data set expands.

[0053] The anti-cloud and fog interference method based on support vector machine can optimize the algorithm as the data set is expanded. If the characteristics of the target change, or more interference feature data is collected, the newly expanded data can be distinguished on the original classification model. If the original model does not recognize the expanded data set well, it can be retrained based on the expanded data set to establish a new interference resolution model.

[0054] 3. Short online processing time and high real-time performance.

[0055] Although the anti-cloud and fog interference method based on support vector machine requires a large amount of test data to train the model, this process can be carried out offline. After the effective classification model is trained based on the test data through the support vector machine algorithm, the actual laser radar recognition application only needs to write the model parameters into the pre-set program. In the real-time detection process of the signal, the judgment result of the echo data at the current moment can be quickly obtained according to the model - target or interference, and distance detection can be performed. The processing time is only milliseconds or even microseconds, so real-time distance detection under anti-interference can be achieved.

[0056] 4. Reduce data processing complexity and human errors.

[0057] After manually extracting waveform eigenvalues, the principal component analysis (PCA) is used to reduce the dimension of the extracted eigenvalues, replace the original multidimensional variables with fewer comprehensive features, simplify the data structure, and integrate the original indicators into fewer principal components. Then, the contribution rate of these principal components is used as the weight to perform weighted average and construct a comprehensive evaluation function. In particular, the extracted features and weights are more objective and reasonable, avoiding the defects of manually extracting eigenvalues. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a schematic diagram for explaining the classification of the support vector machine of the present invention;

[0059] Figure 2 It is the one-dimensional range image of the laser radar of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and their corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this document.

[0061] The following combination Figure 1-2 , describes in detail the technical solutions provided by each embodiment of this specification. Figure 2 In the figure, the left side is the target echo image, and the right side is the interference echo image. Specific embodiment 1:

[0063] The echoes of environmental interference and targets are collected separately, and the feature values ​​of the collected one-dimensional high-resolution range image are extracted and data marked. The manually extracted feature values ​​are reduced in dimension. Based on this, the support vector machine algorithm is used for model training based on the classified data feature values ​​to obtain the optimal classifier model. This model is applied to the real-time detection algorithm of the laser radar to achieve real-time discrimination of targets and interference.

[0064] This method can effectively distinguish natural environment interference through software upgrades based on low-cost one-dimensional laser radar, and can be continuously optimized, upgraded and iterated through the expansion of data sets. The data training process is carried out in the background, and has the advantages of fast speed and high real-time performance in practical applications.

[0065] Step 1: Data Collection

[0066] LiDAR collects echoes from environments such as clouds, fog, rain, snow, smoke, etc. in a variety of complex scenes, as well as echoes from targets of different colors and shapes in a variety of scenes, such as vehicles and houses. The echo data is stored with the same sampling frequency, number of sampling points, and the same data format. The laser echo data is sampled at N points to obtain a one-dimensional range image, recorded as x n ,n=1~N。

[0067] Step 2: Feature extraction and data labeling

[0068] The collected interference data and target data are classified and marked, and characteristic values ​​are extracted. The characteristic values ​​include but are not limited to the integral of the one-dimensional range image, the positive and negative peak values ​​of the range image, the location of the peak value, the effective data length of the range image, the mean and standard deviation of the range image, the weighted center of the range image, the entropy of the range image, etc.

[0069] The integral calculation method of the one-dimensional range image is:

[0070]

[0071] The calculation method of the positive peak value of the one-dimensional range image is:

[0072] P 2 =max[x n ], n = 1 ~ N;

[0073] And record the sampling point number corresponding to the positive peak value, which is recorded as the characteristic value P 3 ;

[0074] The negative peak value calculation method of the one-dimensional range image is:

[0075] P 4 =max[-x n ], n = 1 ~ N;

[0076] And record the sampling point number corresponding to the negative peak value, which is recorded as the characteristic value P 5 ;

[0077] The effective data length calculation method of one-dimensional range image is: record the first value > P 2 The sampling point number of / 2 is t st , record the first <P after the peak 2 The sampling point number of / 2 is t end , eigenvalue P 6 =t end -t st +1;

[0078] The mean calculation method of the one-dimensional range image is:

[0079]

[0080] The standard deviation of the one-dimensional range image is calculated as:

[0081]

[0082] The weighted center calculation method of the one-dimensional range image is:

[0083]

[0084] The entropy calculation method of one-dimensional distance image is:

[0085]

[0086] Step 3: Eigenvalue Normalization

[0087] After converting the one-dimensional distance image into multiple eigenvalues, the eigenvalues ​​need to be normalized because the magnitude of the calculated results of each eigenvalue is quite different. The large numerical differences between different data may lead to poor model training results. i For example, the feature value in the data set is P i_k , k ranges from 1 to M. The normalized eigenvalue Q i_k The calculation method is:

[0088]

[0089] Step 4: Eigenvalue Dimensionality Reduction

[0090] After normalizing the eigenvalues, principal component analysis is used to reduce the dimensionality of the eigenvalues.

[0091] The principle of principal component analysis is to replace the original multidimensional variables with fewer comprehensive variables to simplify the data structure, integrate the original indicators into fewer principal components, and then use the contribution rates of these principal components as weights for weighted average to construct a comprehensive evaluation function. As a multi-index analysis method, in the comprehensive evaluation function, the weight of each principal component is its contribution rate, which reflects the proportion of the information content of the original data contained in the principal component to the total information content. This determination of the weight is objective and reasonable, which overcomes the defects of artificially determining the weight in some evaluation methods and is easy to implement on a computer.

[0092] There are M data in the data set, and K eigenvalues ​​are extracted from each data to form a data matrix X 0 . For the data matrix X 0 Decentralize, that is, subtract the matrix mean to obtain the matrix X.

[0093] Calculate the covariance matrix The eigenvalue λ 1 ≥λ 2 ≥…≥λp ≥0 and the corresponding eigenvector u 1 ,u 2 ,…,u p , and arrange the eigenvectors into a matrix from top to bottom according to the corresponding eigenvalues;

[0094] Calculate the cumulative contribution rate (where d is the number of principal components and p is the number of total variables), then determine whether the condition is met The smallest d, usually c 0 =85%;

[0095] Take the first d rows of eigenvectors to form the matrix P;

[0096] The data is reduced in dimension by Y=PX.

[0097] By using principal component analysis to reduce data dimensionality, the correlation between principal components is reduced while retaining the original amount of information. This can effectively reduce the redundancy between artificially determined eigenvalue data, help reduce the complexity of model training, and improve training accuracy.

[0098] Step 5: Model training to obtain model parameters

[0099] The model is trained using the eigenvalues ​​and data labels after dimensionality reduction obtained in the fourth step.

[0100] For n training points x i , each training point has an index y i The training set is: S = {(x 1 ,y 1 ),(x 2 ,y 2 ),…,(x n ,y n )}. i ∈R N is the input variable, whose components are eigenvalues; y i ∈Y={1,-1}, is the data labeling result (target or interference).

[0101] The purpose of model training is to find a function g(x) and then define the following decision function to achieve the output:

[0102] f(x)=sgn[g(x)]

[0103] Where sgn(z) is the sign function, that is, it takes the value 1 when z ≥ 0, otherwise it takes the value -1. Define g(x) = ω T x+b is the classification plane, that is, the hyperplane. Determine ω and b so that the minimum distance interval from the points in the data set to the hyperplane is maximized.

[0104] Assuming that the hyperplane can correctly classify the training samples, for (x i ,y i )∈S, then:

[0105]

[0106] The minimum distance from a point in the data set to the hyperplane is:

[0107]

[0108] Therefore, the purpose of support vector machine model training is:

[0109]

[0110] In order to ensure that the algorithm is robust to incompletely separable data sets, a relaxation parameter ε is added to the constraint conditions. i ≥0, and to avoid ε i The value is too large, so the penalty term C is added to the objective function, and the support vector machine model optimization problem can be written as

[0111]

[0112] Convert the above optimization problem into a quadratic function extremum problem and define the Lagrangian function:

[0113]

[0114] where α i ≥0, β i ≥0 is the Lagrange multiplier. Let the partial derivative of L(ω,b,α,ε,β) with respect to ω, b, ε be 0, we get:

[0115]

[0116] This gives us the dual problem

[0117]

[0118] Since data is difficult to be linearly separable in low dimensions, it is necessary to use the kernel function k(x i ,x j )=φ(x i ) T φ(x j ) maps the data to a high-dimensional space. Kernel functions include but are not limited to linear kernels, polynomial kernels, Gaussian kernels, etc.

[0119] Therefore, the model solution problem can be written as

[0120]

[0121] Based on the convex optimization algorithm, α can be solved i , and the set of support vectors constituting the hyperplane {(x s ,y s )}, thus obtaining the optimal classification hyperplane g(x)=ω T x+b.

[0122] Step 6: Real-time interference classification

[0123] The principal component matrix P obtained in the fourth step and the optimal classification model coefficients obtained in the fifth step are written into the lidar recognition algorithm.

[0124] The real-time collected laser radar echo data x n , n = 1 ~ N, calculate K eigenvalues, use the principal component matrix P to reduce the dimension, and then normalize it to obtain the data matrix X S .

[0125] Using support vectors and kernel functions, calculate X S The distance D between the support vector and the hyperplane is classified and predicted by the distance value. If the distance is positive, it is identified as a target, and if the distance is negative, it is identified as interference.

[0126] The above technical solution solves the following technical problems:

[0127] 1. This application can solve the problem of laser radar resisting interference such as clouds and fog at low cost

[0128] Traditional one-dimensional laser radar can only collect the distance information and corresponding amplitude information of the echo. When the echo intensity at a certain distance point or a certain distance point exceeds the threshold level, it is considered that a target appears at that distance. Furthermore, the resolution of the one-dimensional radar echo can be improved by increasing the system bandwidth, so that the echo of the target can appear in multiple distance units. The echo signal strength and characteristics in each distance unit are different, thereby extracting more waveform features to distinguish the target from interference such as clouds, fog, rain, snow, smoke and dust in the natural environment.

[0129] In actual applications, due to the different irradiation angles of the laser radar beam on the target, the characteristics of the echo will be somewhat discrete, and the reflection characteristics of the target will also vary greatly. Therefore, the method of manually extracting feature values ​​and using empirical values ​​for threshold judgment will have large errors.

[0130] As attached Figure 2 As shown, (a) is the echo of a physical target, and (b) is the echo of a smoke. It is difficult to accurately distinguish the threshold between the physical target and the smoke interference based on the artificially extracted peak value, pulse width and other information.

[0131] By using the classification network for machine learning based on the manually extracted eigenvalues ​​and taking the eigenvalues ​​as indicators, the efficiency of interference recognition of lidar echo data can be effectively improved.

[0132] 2. This application can solve the problem of large amount of data required for waveform classification and high time cost of manual identification

[0133] Lidar echoes are sensitive to the target's illumination angle, target geometry and reflection characteristics. At the same time, the concentration of smoke and fog, rainfall and snowfall in the environmental interference will affect the radar echo. Therefore, in order to extract the target and interference features more accurately, a large amount of Lidar echo data with different illumination angles, different target types and different interference types needs to be accumulated. Manual processing is used to process and extract features from a large amount of data and determine the interference threshold. On the one hand, manual processing has large errors and it is difficult to extract the optimal threshold; on the other hand, manual processing of a large amount of data takes a very long time.

[0134] Using machine learning to classify waveforms can obtain stable and accurate classification results with fast training speed and low time cost.

[0135] 3. This application can solve the problem of long processing time and low real-time performance of complex recognition algorithms

[0136] LiDAR has high requirements for the real-time data processing in certain application scenarios. For example, the on-board ranging radar used in unmanned driving and the speed measurement LiDAR on the viaduct need to quickly distinguish the target and interference from the echo information within a few seconds or even milliseconds, and provide accurate distance information. LiDARs that use imaging and other systems usually require a higher echo signal-to-noise ratio and exposure time to obtain higher-dimensional echo information to achieve target and interference distinction. As a result, it takes a long time to obtain each processing result, which is not suitable for scenarios with high real-time requirements.

[0137] This application uses a trained model to perform real-time recognition on each set of echo data, with a fast processing speed that can be as fast as milliseconds or even microseconds.

[0138] 4. This application can solve the errors and data complexity introduced when manually extracting feature values

[0139] The most important point of the classification algorithm based on machine learning is the extraction of eigenvalues. The quality of eigenvalue extraction determines the accuracy of interference identification. The main purpose of eigenvalue extraction is to obtain the eigenvalues ​​that are unique to the target signal in the waveform and have the highest correlation with its inherent characteristics, and still have high recognition characteristics as the data projection space changes. There is usually a strong correlation between the artificially extracted waveform eigenvalues, and there will be a lot of repeated information between the eigenvalues. If multiple eigenvalues ​​extracted manually are directly used for analysis, it will cause large processing errors and increase the complexity of data processing.

[0140] This application reduces the redundancy of the feature value itself by means of data dimensionality reduction, thereby further improving the recognition efficiency.

[0141] By solving the above technical problems, the present invention has the following innovative features:

[0142] 1. Without increasing the hardware cost, the support vector machine machine learning algorithm is used to identify interference from the echoes collected by the one-dimensional laser radar

[0143] Support vector machines are used to solve binary classification problems. For any set of data, there are two types of echo data, target and interference. The support vector machine maps the data to points in space, and there are countless intervals that can separate the two types of data (target and interference). By finding one of the intervals, the minimum distance from the points on both sides of the interval to the interval is maximized, that is, finding the interval that can maximally distinguish the two types of data. Figure 1 As shown, for two groups of point sets, intervals A, B, C, and D can all distinguish them, and interval C is the interval that can separate the two types of data with the maximum distance.

[0144] The lidar cloud and fog interference threshold calculation method based on support vector machine can distinguish targets and interference in lidar echoes only through software algorithms without making any hardware changes, greatly reducing labor costs and improving the lidar's target recognition capability and distance measurement accuracy.

[0145] 2. The recognition model can be optimized as the data set expands

[0146] The anti-cloud and fog interference method based on support vector machine can optimize the algorithm as the data set is expanded. If the characteristics of the target change, or more interference feature data is collected, the newly expanded data can be distinguished on the original classification model. If the original model does not recognize the expanded data set well, it can be retrained based on the expanded data set to establish a new interference resolution model.

[0147] 3. Short online processing time and high real-time performance

[0148] Although the anti-cloud and fog interference method based on support vector machine requires a large amount of test data to train the model, this process can be carried out offline. After the effective classification model is trained based on the test data through the support vector machine algorithm, the actual laser radar recognition application only needs to write the model parameters into the pre-set program. In the real-time detection process of the signal, the judgment result of the echo data at the current moment can be quickly obtained according to the model - target or interference, and distance detection can be performed. The processing time is only milliseconds or even microseconds, so real-time distance detection under anti-interference can be achieved.

[0149] 4. Reduce data processing complexity and human errors

[0150] After manually extracting waveform eigenvalues, the principal component analysis (PCA) is used to reduce the dimension of the extracted eigenvalues, replace the original multidimensional variables with fewer comprehensive features, simplify the data structure, and integrate the original indicators into fewer principal components. Then, the contribution rate of these principal components is used as the weight to perform weighted average and construct a comprehensive evaluation function. In particular, the extracted features and weights are more objective and reasonable, avoiding the defects of manually extracting eigenvalues. Specific embodiment 2:

[0152] The purpose of the present invention is to provide a laser radar anti-environmental interference identification method and system based on machine learning. To solve the problems described in the background technology, the specific technical solution is as follows:

[0153] In a first aspect, the present invention provides a laser radar anti-environmental interference identification method based on machine learning, the method comprising:

[0154] Classify and mark the interference data and target data, and perform feature value extraction, normalization, and dimensionality reduction processing to determine the reduced dimension feature value and marked data;

[0155] Based on the reduced dimension feature values ​​and the labeled data, an optimal classification hyperplane in the machine learning model is determined by training in the machine learning model; the classification hyperplane is a function that classifies the reduced dimension feature values ​​according to a rule, the rule being that the minimum distance interval from the points in the data set to the hyperplane is the largest, and the optimal classification hyperplane is the optimal solution among the multiple classification hyperplanes;

[0156] Based on the principal component matrix and the optimal classification hyperplane, the received lidar echo data is classified and predicted to identify the detection results of the lidar, and the detection results include: identification as a target and identification as interference; the principal component matrix is ​​a matrix of the reduced dimensionality eigenvalue vectorization, which is used to perform dimensionality reduction processing on the lidar echo data.

[0157] Furthermore, the interference data and the target data are classified and marked, and feature value extraction, normalization, and dimension reduction are performed to determine the dimension reduction feature value and the marked data, including:

[0158] Classify and mark the collected interference data and target data and extract characteristic values ​​to determine the characteristic value data of the one-dimensional range image;

[0159] Normalizing the characteristic value data of the one-dimensional range image to determine normalized data;

[0160] The normalized data is subjected to dimensionality reduction processing to determine dimensionality reduction feature values ​​and labeled data.

[0161] Furthermore, before classifying and marking the interference data and the target data and performing feature value extraction, normalization, and dimension reduction, the method further includes:

[0162] Collect environmental echoes in a complex environment to determine the interference data, wherein the environmental echoes include echoes of clouds, fog, rain, snow, and smoke;

[0163] Target echoes in different scenarios are collected to determine the target data.

[0164] Further, the training in the machine learning model based on the dimension reduction feature value and the labeled data to determine the optimal classification hyperplane in the machine learning model includes:

[0165] Determine the purpose of training the support vector machine model in the machine learning, that is, the objective function of the training;

[0166] Adding relaxation term parameters to the constraint conditions and adding penalty terms to the objective function;

[0167] Mapping the reduced-dimensional feature values ​​and the labeled data into a high-dimensional space through a kernel function;

[0168] Training is performed in the high-dimensional spatial data environment to determine the optimal classification hyperplane.

[0169] Furthermore, the classification prediction of the received laser radar echo data based on the principal component matrix and the optimal classification hyperplane includes:

[0170] Perform eigenvalue processing on the laser radar echo data collected in real time to determine the laser eigenvalue;

[0171] Performing dimensionality reduction processing on the laser eigenvalues ​​through the principal component matrix to obtain laser dimensionality reduction data;

[0172] Normalizing the laser dimension reduction data to determine a data matrix;

[0173] The support vector and kernel function are used to determine the distance from the data matrix to the optimal classification hyperplane, and classification prediction is performed based on the distance. If the distance is positive, it is identified as a target, and if the distance is negative, it is identified as interference.

[0174] In a second aspect, the present invention also provides a laser radar anti-environmental interference identification system based on machine learning, the system comprising: a pre-training processing module, a training module, and a prediction module;

[0175] The pre-training processing module is used to classify and mark the interference data and the target data and perform feature value extraction, normalization, and dimension reduction processing to determine the dimension reduction feature value and mark the data;

[0176] The training module is used to train and determine the optimal classification hyperplane in the machine learning model based on the reduced dimension feature value and the labeled data; the classification hyperplane is a function that classifies the reduced dimension feature value according to a rule, and the rule is that the minimum distance interval from the point in the data set to the hyperplane is the largest, and the optimal classification hyperplane is the optimal solution among multiple classification hyperplanes;

[0177] The prediction module is used to classify and predict the received lidar echo data based on the principal component matrix and the optimal classification hyperplane, and identify the detection results of the lidar, and the detection results include: identification as a target and identification as interference; the principal component matrix is ​​the matrix of the dimensionality reduction eigenvalue vectorization, which is used to perform dimensionality reduction processing on the lidar echo data.

[0178] In another embodiment of the second aspect, the pre-training processing module is specifically used to:

[0179] Classify and mark the collected interference data and target data and extract characteristic values ​​to determine the characteristic value data of the one-dimensional range image;

[0180] Normalizing the characteristic value data of the one-dimensional range image to determine normalized data;

[0181] The normalized data is subjected to dimensionality reduction processing to determine dimensionality reduction feature values ​​and labeled data.

[0182] In another embodiment of the second aspect, the system further includes: an acquisition module;

[0183] The acquisition module is used to collect environmental echoes in a complex environment to determine the interference data, wherein the environmental echoes include echoes of clouds, fog, rain, snow, and smoke; and collect target echoes in different scenes to determine the target data.

[0184] In another embodiment of the second aspect, the training module is specifically used to:

[0185] Determine the purpose of training the support vector machine model in the machine learning, that is, the objective function of the training;

[0186] Adding relaxation term parameters to the constraint conditions and adding penalty terms to the objective function;

[0187] Mapping the reduced-dimensional feature values ​​and the labeled data into a high-dimensional space through a kernel function;

[0188] Training is performed in the high-dimensional spatial data environment to determine the optimal classification hyperplane.

[0189] In another embodiment of the second aspect, the prediction module is specifically used to:

[0190] Perform eigenvalue processing on the laser radar echo data collected in real time to determine the laser eigenvalue;

[0191] Performing dimensionality reduction processing on the laser eigenvalues ​​through the principal component matrix to obtain laser dimensionality reduction data;

[0192] Normalizing the laser dimension reduction data to determine a data matrix;

[0193] The support vector and kernel function are used to determine the distance from the data matrix to the optimal classification hyperplane, and classification prediction is performed based on the distance. If the distance is positive, it is identified as a target, and if the distance is negative, it is identified as interference.

[0194] Example embodiments have been disclosed herein, and although specific terms are employed, they are used and should be interpreted only in a general illustrative sense and not for limiting purposes. In some embodiments, it will be apparent to those skilled in the art that, unless otherwise expressly noted, features, characteristics, and / or elements described in conjunction with a particular embodiment may be used alone or in combination with features, characteristics, and / or elements described in conjunction with other embodiments. Therefore, it will be understood by those skilled in the art that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A laser radar anti-environmental interference identification method based on machine learning, characterized in that: The method comprises: Classify and mark the interference data and target data, and perform feature value extraction, normalization, and dimensionality reduction processing to determine the reduced dimension feature value and marked data; Based on the reduced dimension feature values ​​and the labeled data, an optimal classification hyperplane in the machine learning model is determined by training in the machine learning model; the classification hyperplane is a function that classifies the reduced dimension feature values ​​according to a rule, the rule being that the minimum distance interval from the points in the data set to the hyperplane is the largest, and the optimal classification hyperplane is the optimal solution among the multiple classification hyperplanes; Based on the principal component matrix and the optimal classification hyperplane, the received lidar echo data is classified and predicted to identify the detection results of the lidar, and the detection results include: identification as a target and identification as interference; the principal component matrix is ​​a matrix of the reduced dimensionality eigenvalue vectorization, which is used to perform dimensionality reduction processing on the lidar echo data.

2. The laser radar anti-environmental interference identification method based on machine learning as claimed in claim 1, characterized in that: The interference data and the target data are classified and marked, and feature value extraction, normalization, and dimension reduction are performed to determine the dimension reduction feature value and the marked data, including: Classify and mark the collected interference data and target data and extract characteristic values ​​to determine the characteristic value data of the one-dimensional range image; Normalizing the characteristic value data of the one-dimensional range image to determine normalized data; The normalized data is subjected to dimensionality reduction processing to determine dimensionality reduction feature values ​​and labeled data.

3. The laser radar anti-environmental interference identification method based on machine learning as claimed in claim 1, characterized in that: Before classifying and marking the interference data and the target data and performing feature value extraction, normalization, and dimension reduction processing, the method further includes: Collect environmental echoes in a complex environment to determine the interference data, wherein the environmental echoes include echoes of clouds, fog, rain, snow, and smoke; Target echoes in different scenarios are collected to determine the target data.

4. The laser radar anti-environmental interference identification method based on machine learning as claimed in claim 1, characterized in that: The step of training and determining the optimal classification hyperplane in the machine learning model based on the dimension reduction feature value and the labeled data comprises: Determine the purpose of training the support vector machine model in the machine learning, that is, the objective function of the training; Adding relaxation term parameters to the constraint conditions and adding penalty terms to the objective function; Mapping the reduced-dimensional feature values ​​and the labeled data into a high-dimensional space through a kernel function; Training is performed in the high-dimensional spatial data environment to determine the optimal classification hyperplane.

5. The laser radar anti-environmental interference identification method based on machine learning as claimed in claim 1, characterized in that: The method of performing classification prediction on the received laser radar echo data based on the principal component matrix and the optimal classification hyperplane includes: Perform eigenvalue processing on the laser radar echo data collected in real time to determine the laser eigenvalue; Performing dimensionality reduction processing on the laser eigenvalues ​​through the principal component matrix to obtain laser dimensionality reduction data; Normalizing the laser dimension reduction data to determine a data matrix; The support vector and kernel function are used to determine the distance from the data matrix to the optimal classification hyperplane, and classification prediction is performed based on the distance. If the distance is positive, it is identified as a target, and if the distance is negative, it is identified as interference.

6. A laser radar anti-environmental interference recognition system based on machine learning, characterized in that: The system comprises: a pre-training processing module, a training module, and a prediction module; The pre-training processing module is used to classify and mark the interference data and the target data and perform feature value extraction, normalization, and dimension reduction processing to determine the dimension reduction feature value and mark the data; The training module is used to train and determine the optimal classification hyperplane in the machine learning model based on the reduced dimension feature value and the labeled data; the classification hyperplane is a function that classifies the reduced dimension feature value according to a rule, and the rule is that the minimum distance interval from the point in the data set to the hyperplane is the largest, and the optimal classification hyperplane is the optimal solution among multiple classification hyperplanes; The prediction module is used to classify and predict the received lidar echo data based on the principal component matrix and the optimal classification hyperplane, and identify the detection results of the lidar, and the detection results include: identification as a target and identification as interference; the principal component matrix is ​​the matrix of the dimensionality reduction eigenvalue vectorization, which is used to perform dimensionality reduction processing on the lidar echo data.

7. The laser radar anti-environmental interference recognition system based on machine learning as claimed in claim 6, characterized in that: The pre-training processing module is specifically used for: Classify and mark the collected interference data and target data and extract characteristic values ​​to determine the characteristic value data of the one-dimensional range image; Normalizing the characteristic value data of the one-dimensional range image to determine normalized data; The normalized data is subjected to dimensionality reduction processing to determine dimensionality reduction feature values ​​and labeled data.

8. The laser radar anti-environmental interference recognition system based on machine learning as claimed in claim 6, characterized in that: The system further comprises: an acquisition module; The acquisition module is used to collect environmental echoes in a complex environment to determine the interference data, wherein the environmental echoes include echoes of clouds, fog, rain, snow, and smoke; and collect target echoes in different scenes to determine the target data.

9. The laser radar anti-environmental interference recognition system based on machine learning as claimed in claim 6, characterized in that: The training module is specifically used for: Determine the purpose of training the support vector machine model in the machine learning, that is, the objective function of the training; Adding relaxation term parameters to the constraint conditions and adding penalty terms to the objective function; Mapping the reduced-dimensional feature values ​​and the labeled data into a high-dimensional space through a kernel function; Training is performed in the high-dimensional spatial data environment to determine the optimal classification hyperplane.

10. The laser radar anti-environmental interference recognition system based on machine learning as claimed in claim 6, characterized in that: The prediction module is specifically used for: Perform eigenvalue processing on the laser radar echo data collected in real time to determine the laser eigenvalue; Performing dimensionality reduction processing on the laser eigenvalues ​​through the principal component matrix to obtain laser dimensionality reduction data; Normalizing the laser dimension reduction data to determine a data matrix; The support vector and kernel function are used to determine the distance from the data matrix to the optimal classification hyperplane, and classification prediction is performed based on the distance. If the distance is positive, it is identified as a target, and if the distance is negative, it is identified as interference.

Images