Abnormal driving investigation method based on high-precision track depth expression adversarial model

Abstract

The invention discloses an abnormal driving investigation method based on a high-precision track depth expression confrontation model. The method comprises the following steps: 1, obtaining high-precision track data and investigation record data of a vehicle needing abnormal driving investigation; 2, constructing and training a vehicle high-precision track depth expression adversarial model, including two self-encoding deep neural networks, and forming a dual adversarial structure; 3, verifying and optimizing a vehicle high-precision track depth expression confrontation model; and 4, deploying and applying the vehicle high-precision track depth expression confrontation model meeting the requirements, specifically deploying at a background center system and a roadside edge end, and carrying out information intercommunication with police individual soldier equipment checked on site. According to the invention, the efficiency and accuracy of abnormal driving behavior discrimination in road traffic law enforcement can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic management, and in particular to an abnormal driving detection method based on a high-precision trajectory deep expression adversarial model. Background Art

[0002] Abnormal driving conditions such as driving under the influence of alcohol and driving under the influence of drugs seriously endanger the normal operation of road traffic. Cracking down on such abnormal driving behaviors is an important part of ensuring traffic safety and order. Therefore, it is particularly important to accurately identify whether a moving vehicle has abnormal driving behavior and effectively carry out inspections and disposal.

[0003] Currently, there are three main methods for abnormal driving identification, namely biological signal analysis, surveillance image analysis, and driving data analysis, but the above methods all have certain problems in practical applications.

[0004] Among them, the first two categories rely on individual drivers to implant or wear sensors, as well as the installation of monitoring sensors on specific vehicles. They are generally targeted at taxis, online ride-hailing vehicles, freight vehicles, etc., and are highly targeted, making it difficult to carry out universal public road traffic management and illegal behavior detection;

[0005] The third type of driving data analysis method can use vehicle sensor data, or consider using observation methods other than vehicles and drivers to collect data, such as vehicle speed, trajectory location information, etc. However, it is still limited by traditional sensors and related technical methods, and there are problems such as insufficient data accuracy and low accuracy in abnormal driving recognition, as follows:

[0006] (1) The data provided by traditional sensors have large inherent errors, which can easily be amplified during the identification process, resulting in misjudgment, which can easily cause the final identification to be inaccurate for practical application.

[0007] (2) Driving data are mostly continuous, high-dimensional sequence data. It is necessary to mine and identify the microscopic vehicle operation characteristics of a very small number of abnormal driving behaviors and accurately identify the differences between abnormal driving behaviors and normal driving behaviors in data such as driving trajectories, which places high demands on the recognition model.

[0008] In recent years, with the development of technologies such as LiDAR, millimeter-wave radar, Beidou positioning, and high-precision maps, it has become easier to obtain high-precision trajectory data of vehicles on the road, providing new tools for data-driven road traffic management. The high-precision trajectory single-point positioning accuracy reaches the centimeter level, which can describe the vehicle driving conditions in a more granular manner, including trajectory fluctuations caused by abnormal driving behavior of the driver.

[0009] In terms of recognition models, the extensive application of deep learning in images, language, etc. has proved its advantages in mining large-scale complex data features. Through effective and appropriate model construction and training, it can more accurately perform deep expression and anomaly identification of complex high-dimensional data.

[0010] Therefore, how to change the traditional inspection method from manual "random sampling" like looking for a needle in a haystack to the mode of "first machine intelligent full inspection, then targeted manual verification" to improve the efficiency and accuracy of identifying abnormal driving behaviors in road traffic law enforcement has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0011] In view of the above-mentioned defects of the prior art, the present invention provides an abnormal driving detection method based on a high-precision trajectory deep expression adversarial model. The purpose of the present invention is to apply the idea of ​​deep learning adversarial training, take advantage of the high-precision trajectory detection, and mine the microscopic features of the driving trajectory of the vehicle driven by the driver in the abnormal driving state, so as to change the traditional inspection method from the manual "random sampling" like finding a needle in a haystack to the mode of "first machine intelligent full inspection, then targeted manual verification", so as to improve the efficiency and accuracy of identifying abnormal driving behavior in road traffic law enforcement.

[0012] To achieve the above object, the present invention discloses an abnormal driving detection method based on a high-precision trajectory deep expression adversarial model, comprising the following steps:

[0013] Step 1: Obtain high-precision trajectory data and investigation record data of vehicles that need to be investigated for abnormal driving;

[0014] Step 2: Construction and training of a high-precision vehicle trajectory deep expression adversarial model, including two autoencoder deep neural networks, and forming a dual adversarial structure;

[0015] Step 3: Verification and optimization of the vehicle high-precision trajectory deep expression adversarial model;

[0016] Step 4: Deploy and apply the high-precision trajectory deep expression adversarial model of the vehicle that meets the requirements, specifically deploy it in the supporting background center system and the roadside edge end, and exchange information with the individual equipment of the police force conducting on-site inspections.

[0017] Preferably, step 1 is as follows:

[0018] Step 1.1, select a relatively gentle road section with a linear longitudinal slope of no more than 3% as a data collection section;

[0019] Constructing a digital base map of the data collection section;

[0020] Extract and fit the center line of the lane where the vehicle passes and record it as curve y 0 = l(x0 );

[0021] Among them, y 0 、x 0 The coordinates of each point on the center line on the digitized base map;

[0022] Step 1.2, acquiring point cloud data of the vehicle through high-precision trajectory acquisition devices continuously deployed on the roadside of the road section, and deducing the high-precision trajectory data of the vehicle through the point cloud data;

[0023] The high-precision trajectory data includes a high-precision vehicle driving trajectory l with a trajectory positioning error accuracy of at least centimeter level;

[0024] Calculate the vehicle's high-precision driving trajectory l and the center line y 0 = l(x 0 ) and form a sequence ε;

[0025] The longitudinal variation sequence of the front and rear track points of the high-precision driving track l of the vehicle is δ, and the speed of the vehicle in the road section is v. When the sampling time interval is small, such as less than or equal to 0.1 second, it is considered to be approximately the instantaneous speed;

[0026] Then, ε=[ε 1 ,ε 2 ,...,ε L ] T ;

[0027] δ=[δ 1 ,δ 2 ,...,δ L ] T ;

[0028] v=[v 1 ,v 2 ,...,v L ] T ;

[0029] Where L is the coordinate of the trajectory point on l at time t (x t ,y t ), i.e., t = 1, 2, ..., L;

[0030] The coordinates of the ath trajectory point (x a ,y a ) to the center line y 0 = l(x 0 ) has a lateral displacement of ε a =y a -l(x a ), longitudinal variation speed

[0031] Where a=1, 2, 3, ..., L, when a≥2, x a 、x a-1 respectively refer to the a-th and the previous track point's horizontal coordinate values;

[0032] Step 1.3: At the end of the data collection section, a temporary manual inspection point is set up in combination with daily traffic management and inspection work to evaluate and record the status of drivers of some passing vehicles. The evaluation status label is s = [s 1 ,s 2 ] T ;

[0033] in, j = 1 or 2;

[0034] j=1 corresponds to normal driving status; j=2 corresponds to abnormal status including driving under the influence of alcohol, driving under the influence of drugs, driving while fatigued and driving aggressively;

[0035] Other situations where the driver's status is not collected and assessed on-site are distinguished by larger numerical marks;

[0036] Step 1.4: Continue to collect data until a dataset C = {(ε q ,δ q ,v q ,s q )q=0,1,2,…,N}

[0037] Among them, for the qth vehicle, the sample data collected and retained for its passage includes the lateral offset sequence ε q , the longitudinal variation series δ q , speed sequence v q and evaluate record status q , that is, the corresponding data pair is constructed as (ε q ,δ q ,v q ,s q ); N is the size of the data set C;

[0038] Step 1.5: Divide the samples in the data set C into three sets, specifically:

[0039] The samples without on-site record status labels are divided into training set C train ,

[0040] The samples with the field record status labels are randomly divided into the test set C test ,

[0041] And the validation set C validate .

[0042] More preferably, step 2 is as follows:

[0043] Step 2.1: Construct two autoencoding deep neural networks M 1 、M 2 , respectively including the first encoder M E1 and the first decoder M D1 , and the second encoder M E2 and the second decoder M D2 ;

[0044] Step 2.2: M 1 、M 2 The first stage of training is carried out, from the training set C train Select the matrix X consisting of the model input data (ε, δ, v), then M 1 The output is M 1 (X) = M D1 (M E1 (X)), M 2 The output is M 2 (X) = M D2 (M E2 (X));

[0045] Define the first stage of training M 1 The loss functions are: LOSS 1 1 =||XM 1 (X)|| 2 ,

[0046] M 2 The loss functions are: LOSS 2 1 =||XM 2 (X)|| 2 ;

[0047] Among them, ||·|| 2 represents the second-order norm calculation, that is, the first-stage training objectives are minLOSS 1 1 and minLOSS 2 1 ;

[0048] Step 2.3: M 1 、M 2 In the second stage of training, M 1 The output is M 2 The input training M 2 , M 2 The output is M 1 The input training M 1;

[0049] Define the second stage of training M 1 The loss functions are: LOSS 1 2 =||XM 2 (M 1 (X))|| 2 , M 2 The loss functions are: LOSS 2 2 =-||XM 1 (M 2 (X))|| 2 ;

[0050] Define the second stage of training M 1 and M 2 The training objectives are:

[0051]

[0052] Step 2.4: Establish the abnormal discrimination score index θ(X), as follows:

[0053] θ(X)=β||XM 2 (M 1 (X))|| 2 +(1-β)||XM 1 (M 2 (X))|| 2 ;

[0054] Among them, β is a hyperparameter, which takes 0≤β≤1;

[0055] Step 2.5, calibrate the abnormal discrimination score threshold; set the abnormal discrimination score threshold Θ;

[0056] When θ(X) ≥ θ, it is determined that the current train The matrix X composed of the model input data (ε, δ, v) is selected as an abnormal sample; otherwise, it is judged that the current training set C train The matrix X consisting of the model input data (ε, δ, v) is selected as the normal sample.

[0057] More preferably, in step 2.3 and step 2.4, the Adam optimizer is used, the initial learning rate is λ, and during the training process, the learning rate adopts a dynamic strategy and gradually decreases as the training progresses; wherein, the number of training and learning iterations of the first stage training and the second stage training are both K times, and the training is terminated after completing K learning iterations.

[0058] More preferably, in step 2.5, usingtest The abnormal discrimination scores are calculated for the samples, the obtained abnormal discrimination scores are arranged from large to small, and the upper 5% quantile is taken as the abnormal discrimination score threshold Θ.

[0059] More preferably, step 3 is as follows:

[0060] Step 3.1: Use the validation set C validate The sample is input into the vehicle high-precision trajectory deep expression adversarial model, the corresponding abnormal discrimination score index θ(X) is calculated and whether it is an abnormal sample is determined, and then the result is compared with the sample original state label s to calculate the discrimination accuracy;

[0061] Step 3.2: If the accuracy of the vehicle high-precision trajectory deep expression adversarial model is lower than 80%, perform model tuning;

[0062] After the accuracy of the vehicle high-precision trajectory deep expression adversarial model reaches above 80%, proceed to the next step.

[0063] More preferably, the model tuning includes adjusting the initial training parameters and hyperparameters of the vehicle high-precision trajectory deep expression adversarial model.

[0064] Beneficial effects of the present invention:

[0065] The present invention is conducive to the implementation of efficient traffic management work. By judging in advance whether there are any abnormal signs in the vehicle's driving trajectory, it is conducive to the reasonable deployment and efficient interception of the police force in the downstream section, thereby improving the efficiency of on-site inspection.

[0066] The present invention has a wide range of applications. Compared with traditional monitoring and data collection applications targeting specific drivers or vehicles, it is universally applicable to all types of vehicles traveling on the road.

[0067] The vehicle high-precision trajectory deep expression adversarial model proposed in the present invention has high accuracy and fully applies the idea of ​​deep learning adversarial training, which is conducive to further improving the efficiency of model training and enhancing the accuracy of abnormal driving judgment.

[0068] The present invention is intensive in the construction and application of urban infrastructure and is conducive to the full utilization of infrastructure and its data.

[0069] The present invention realizes the mining and discovery of the microscopic representation of abnormal driving behavior in the spatiotemporal trajectory big data through high-precision trajectory data collection of vehicles, which is conducive to the full mining of the data value of roadside equipment and multi-scenario application.

[0070] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Shown is an execution flow chart of an embodiment of the present invention.

[0072] Figure 2 A schematic diagram of the construction and training architecture of a high-precision vehicle trajectory deep expression adversarial model in one embodiment of the present invention is shown. DETAILED DESCRIPTION

[0073] Example

[0074] like Figure 1 As shown, the abnormal driving detection method based on the high-precision trajectory deep expression adversarial model includes the following steps:

[0075] Step 1: Obtain high-precision trajectory data and investigation record data of vehicles that need to be investigated for abnormal driving;

[0076] Step 2: Construction and training of a high-precision vehicle trajectory deep expression adversarial model, including two autoencoder deep neural networks, and forming a dual adversarial structure;

[0077] Step 3: Verification and optimization of the vehicle high-precision trajectory deep expression adversarial model;

[0078] Step 4: Deploy and apply the high-precision trajectory deep expression adversarial model of the vehicle that meets the requirements, specifically deploy it in the supporting background center system and the roadside edge end, and exchange information with the individual equipment of the police force conducting on-site inspections.

[0079] The high-precision vehicle trajectory deep expression adversarial model implemented by the present invention can support abnormal driving identification and management at the front and back ends. By collecting high-precision trajectory data of passing vehicles in real time, remote monitoring of the background center can be realized, or police forces can be deployed on the downstream road sections of data collection to conduct on-site inspections. Accurate vehicle interception and abnormal driving behavior verification can be carried out based on the real-time judgment results sent by the edge end to the individual equipment.

[0080] In some embodiments, step 1 is as follows:

[0081] Step 1.1, select a relatively gentle road section with a linear longitudinal slope of no more than 3% as a data collection section;

[0082] Constructing a digital base map of the data collection section;

[0083] Extract and fit the center line of the lane where the vehicle passes and record it as curve y 0 = l(x 0 );

[0084] Among them, y 0 、x 0The coordinates of each point on the center line on the digitized base map;

[0085] Step 1.2, acquiring point cloud data of the vehicle through high-precision trajectory acquisition devices continuously deployed on the roadside of the road section, and deducing the high-precision trajectory data of the vehicle through the point cloud data;

[0086] The high-precision trajectory data includes a high-precision vehicle driving trajectory l with a trajectory positioning error accuracy of at least centimeter level;

[0087] Calculate the vehicle's high-precision driving trajectory l and the center line y 0 = l(x 0 ) and form a sequence ε;

[0088] The longitudinal variation sequence of the front and rear track points of the high-precision driving track l of the vehicle is δ, and the speed of the vehicle in the road section is v. When the sampling time interval is small, such as less than or equal to 0.1 second, it is considered to be approximately the instantaneous speed;

[0089] Then, ε=[ε 1 ,ε 2 ,...,ε L ] T ;

[0090] δ=[δ 1 ,δ 2 ,...,δ L ] T ;

[0091] v=[v 1 ,v 2 ,...,v L ] T ;

[0092] Where L is the coordinate of the trajectory point on l at time t (x t ,y t ), i.e., t = 1, 2, ..., L;

[0093] The coordinates of the ath trajectory point (x a ,y a ) to the center line y 0 = l(x 0 ) has a lateral displacement of ε a =y a -l(x a ), longitudinal variation speed

[0094] Where a=1, 2, 3, ..., L, when a≥2, x a、x a-1 respectively refer to the a-th and the previous track point's horizontal coordinate values;

[0095] Step 1.3: At the end of the data collection section, a temporary manual inspection point is set up in combination with daily traffic management and inspection work to evaluate and record the status of drivers of some passing vehicles. The evaluation status label is s = [s 1 ,s 2 ] T ;

[0096] in, j = 1 or 2;

[0097] j=1 corresponds to normal driving status; j=2 corresponds to abnormal status including driving under the influence of alcohol, driving under the influence of drugs, driving while fatigued and driving aggressively;

[0098] Other situations where the driver's status is not collected and assessed on-site are distinguished by larger numerical marks;

[0099] The above-mentioned larger numerical mark distinction means that the value should significantly exceed the state value 0 or 1, for example, s = [999,999] T .

[0100] Step 1.4: Continue to collect data until a dataset C = {(ε q ,δ q ,v q ,s q )q=0,1,2,…,N}

[0101] Among them, for the qth vehicle, the sample data collected and retained for its passage includes the lateral offset sequence ε q , the longitudinal variation series δ q , speed sequence v q and evaluate record status q , that is, the corresponding data pair is constructed as (ε q ,δ q ,v q ,s q ); N is the size of the data set C;

[0102] Step 1.5: Divide the samples in the data set C into three sets, specifically:

[0103] The samples without on-site record status labels are divided into training set C train ,

[0104] The samples with the field record status labels are randomly divided into the test set C test ,

[0105] And the validation set C validate .

[0106] like Figure 2 As shown, in some embodiments, step 2 is specifically as follows:

[0107] Step 2.1: Construct two autoencoding deep neural networks M 1 、M 2 , respectively including the first encoder M E1 and the first decoder M D1 , and the second encoder M E2 and the second decoder M D2 ;

[0108] Step 2.2: M 1 、M 2 The first stage of training is carried out, from the training set C train Select the matrix X consisting of the model input data (ε, δ, v), then M 1 The output is M 1 (X) = M D1 (M E1 (X)), M 2 The output is M 2 (X) = M D2 (M E2 (X));

[0109] The purpose of the first stage of training is to make the input real data consistent with M 1 、M 2 The error of each output is minimized;

[0110] Define the first stage of training M 1 The loss functions are: LOSS 1 1 =||XM 1 (X)|| 2 ,

[0111] M 2 The loss functions are: LOSS 2 1 =||XM 2 (X)|| 2 ;

[0112] Among them, ||·|| 2 represents the second-order norm calculation, that is, the first-stage training objectives are minLOSS 1 1 and minLOSS 2 1 ;

[0113] Step 2.3: M 1 、M2 In the second stage of training, M 1 The output is M 2 The input training M 2 , M 2 The output is M 1 The input training M 1 ;

[0114] The purpose of the second stage training is to make M 1 The first decoder M D1 The output is as close as possible to the real data X and M 2 The output of M is the same 2 The second decoder M D2 The output is as close to the real data X as possible but different from M 1 The output is different, thus forming M 1 、M 2 Confront each other and promote each other's improvement.

[0115] Define the second stage of training M 1 The loss functions are: LOSS 1 2 =||XM 2 (M 1 (X))|| 2 , M 2 The loss functions are: LOSS 2 2 =-||XM 1 (M 2 (X))|| 2 ;

[0116] Define the second stage of training M 1 and M 2 The training objectives are:

[0117]

[0118] Step 2.4: Establish the abnormal discrimination score index θ(X), as follows:

[0119] θ(X)=β||XM 2 (M 1 (X))|| 2 +(1-β)||XM 1 (M 2 (X))|| 2 ;

[0120] Among them, β is a hyperparameter, which takes 0≤β≤1;

[0121] Step 2.5, calibrate the abnormal discrimination score threshold; set the abnormal discrimination score threshold Θ;

[0122] When θ(X) ≥ θ, it is determined that the current train The matrix X composed of the model input data (ε, δ, v) is selected as an abnormal sample; otherwise, it is judged that the current training set C train The matrix X consisting of the model input data (ε, δ, v) is selected as the normal sample.

[0123] In some embodiments, the Adam optimizer is used in step 2.3 and step 2.4, the initial learning rate is λ, and during the training process, the learning rate adopts a dynamic strategy and gradually decreases as the training progresses; wherein, the number of training and learning iterations of the first stage training and the second stage training are both K times, and the training is terminated after completing K learning iterations.

[0124] In some embodiments, in step 2.5, using test The abnormal discrimination scores are calculated for the samples, the obtained abnormal discrimination scores are arranged from large to small, and the upper 5% quantile is taken as the abnormal discrimination score threshold Θ.

[0125] In some embodiments, step 3 is as follows:

[0126] Step 3.1: Use the validation set C validate The sample is input into the vehicle high-precision trajectory deep expression adversarial model, the corresponding abnormal discrimination score index θ(X) is calculated and whether it is an abnormal sample is determined, and then the result is compared with the sample original state label s to calculate the discrimination accuracy;

[0127] Step 3.2: If the accuracy of the vehicle high-precision trajectory deep expression adversarial model is lower than 80%, perform model tuning;

[0128] After the accuracy of the vehicle high-precision trajectory deep expression adversarial model reaches above 80%, proceed to the next step.

[0129] The abnormal driving detection method based on the high-precision trajectory deep expression adversarial model according to claim 6 is characterized in that the model tuning includes adjusting the training initial parameters and hyperparameters of the vehicle high-precision trajectory deep expression adversarial model.

[0130] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. An abnormal driving detection method based on a high-precision trajectory deep expression adversarial model; characterized in that: The steps include: Step 1: Obtain high-precision trajectory data and investigation record data of vehicles that need to be investigated for abnormal driving; Step 2: Construction and training of a high-precision vehicle trajectory deep expression adversarial model, including two autoencoder deep neural networks, and forming a dual adversarial structure; Step 3: Verification and optimization of the vehicle high-precision trajectory deep expression adversarial model; Step 4: Deploy and apply the high-precision trajectory deep expression adversarial model of the vehicle that meets the requirements, specifically deploy it in the supporting background center system and the roadside edge end, and exchange information with the individual equipment of the police force conducting on-site inspections.

2. The abnormal driving detection method based on the high-precision trajectory deep expression adversarial model according to claim 1 is characterized in that: Step 1 is as follows: Step 1.1, select a relatively gentle road section with a linear longitudinal slope of no more than 3% as a data collection section; Constructing a digital base map of the data collection section; Extract and fit the center line of the lane where the vehicle passes and record it as a curve y0=l(x0); Wherein, y0 and x0 are the coordinates of each point on the center line on the digitized base map; Step 1.2, acquiring point cloud data of the vehicle through high-precision trajectory acquisition devices continuously deployed on the roadside of the road section, and deducing the high-precision trajectory data of the vehicle through the point cloud data; The high-precision trajectory data includes a high-precision vehicle driving trajectory l with a trajectory positioning error accuracy of at least centimeter level; Calculate the lateral offset between the high-precision driving trajectory l of the vehicle and the center line y0=l(x0) and form a sequence ε; The longitudinal variation sequence of the front and rear track points of the high-precision driving track l of the vehicle is δ, and the speed of the vehicle in the road section is v. When the sampling time interval is small, such as less than or equal to 0.1 second, it is considered to be approximately the instantaneous speed; Then, ε = [ε1,ε2,...,ε L ] T ; δ=[δ1,δ2,...,δ L ] T ; v=[v1,v2,...,v L ] T ; Where L is the coordinate of the trajectory point on l at time t (x t ,y t ), i.e., t = 1, 2, ..., L; The coordinates of the ath trajectory point (x a ,y a ) to the center line y0 = l(x0) is ε a =y a -l(x a ), longitudinal variation speed Where a=1, 2, 3, ..., L, when a≥2, x a 、x a-1 respectively refer to the a-th and the previous track point's horizontal coordinate values; Step 1.3: At the end of the data collection section, a temporary manual inspection point is set up in combination with daily traffic management and inspection work to evaluate and record the status of drivers of some passing vehicles. The evaluation status label is s = [s1, s2] T ; in, j=1 corresponds to normal driving status; j=2 corresponds to abnormal status including driving under the influence of alcohol, driving under the influence of drugs, driving while fatigued and driving aggressively; Other situations where the driver's status is not collected and assessed on-site are distinguished by larger numerical marks; Step 1.4: Continue to collect data until a dataset C = {(ε q ,δ q ,v q ,s q )q=0,1,2,...,N} Among them, for the qth vehicle, the sample data collected and retained for its passage includes the lateral offset sequence ε q , the longitudinal variation series δ q , speed sequence v q and evaluate record status q , that is, the corresponding data pair is constructed as (ε q ,δ q ,v q ,s q ); N is the size of the data set C; Step 1.5: Divide the samples in the data set C into three sets, specifically: The samples without on-site record status labels are divided into training set C train , The samples with the field record status labels are randomly divided into the test set C test , And the validation set C validate .

3. The abnormal driving detection method based on the high-precision trajectory deep expression adversarial model according to claim 2 is characterized in that: Step 2 is as follows: Step 2.1: construct two self-encoding deep neural networks M1 and M2, each including a first encoder M E1 and the first decoder M D1 , and the second encoder M E2 and the second decoder M D2 ; Step 2.2: Perform the first stage training on M1 and M2, and select the training set C train Select the matrix X composed of the model input data (ε, δ, v), then the output of M1 is M1(X)=M D1 (M E1 (X)), the output of M2 is M2(X)=M D2 (M E2 (X)); The loss functions of M1 in the first stage of training are defined as: LOSS1 1 =||X-M1(X)||2, The loss functions of M2 are: LOSS2 1 =||X-M2(X)||2; Among them, ||·||2 represents the second-order norm calculation, that is, the training objectives of the first stage are minLOSS1 1 and minLOSS2 1 ; Step 2.3, perform the second stage training on M1 and M2, use the output of M1 as the input of M2 to train M2, and use the output of M2 as the input of M1 to train M1; The loss functions of M1 in the second stage of training are defined as: LOSS1 2 =||X-M2(M1(X))||2, The loss functions of M2 are: LOSS2 2 = -||X-M1(M2(X))||2; The training objectives of M1 and M2 in the second stage of training are defined as: Step 2.4: Establish the abnormal discrimination score index θ(X), as follows: θ(X)=β||X-M2(M1(X))||2+(1-β)||X-M1(M2(X))||2; Among them, β is a hyperparameter, which takes 0≤β≤1; Step 2.5, calibrate the abnormal discrimination score threshold; set the abnormal discrimination score threshold Θ; When θ(X) ≥ θ, it is determined that the current train The matrix X composed of the model input data (ε, δ, v) is selected as an abnormal sample; otherwise, it is judged that the current training set C train The matrix X consisting of the model input data (ε, δ, v) is selected as the normal sample.

4. The abnormal driving detection method based on the high-precision trajectory deep expression adversarial model according to claim 3 is characterized in that: In step 2.3 and step 2.4, the Adam optimizer is used, the initial learning rate is λ, and during the training process, the learning rate adopts a dynamic strategy and gradually decreases as the training progresses; wherein, the number of training and learning iterations of the first stage training and the second stage training are both K times, and the training is terminated after completing K learning iterations.

5. The abnormal driving detection method based on the high-precision trajectory deep expression adversarial model according to claim 3 is characterized in that: In step 2.5, use the test The abnormal discrimination scores are calculated for the samples, the obtained abnormal discrimination scores are arranged from large to small, and the upper 5% quantile is taken as the abnormal discrimination score threshold Θ.

6. The abnormal driving detection method based on the high-precision trajectory deep expression adversarial model according to claim 3 is characterized in that: Step 3 is as follows: Step 3.1: Use the validation set C validate The sample is input into the vehicle high-precision trajectory deep expression adversarial model, the corresponding abnormal discrimination score index θ(X) is calculated and whether it is an abnormal sample is determined, and then the result is compared with the sample original state label s to calculate the discrimination accuracy; Step 3.2: If the accuracy of the vehicle high-precision trajectory deep expression adversarial model is lower than 80%, perform model tuning; After the accuracy of the vehicle high-precision trajectory deep expression adversarial model reaches above 80%, proceed to the next step.

7. The abnormal driving detection method based on the high-precision trajectory deep expression adversarial model according to claim 6 is characterized in that: The model tuning includes adjusting the initial training parameters and hyperparameters of the vehicle high-precision trajectory deep expression adversarial model.

Images