Driving subject identification method and system in auxiliary driving environment

Abstract

The invention discloses a driving subject identification method and system in an auxiliary driving environment. The method comprises the following steps: acquiring driving data of a vehicle; processing the acquired driving data, and screening lane changing time, steering wheel angle and acceleration fragment data; inputting the screened fragment data into a pre-constructed Stacking integrated driving subject recognition model, and recognizing the probabilities of driving subjects corresponding to different time periods in the fragment data; the Stacking integrated driving subject identification model comprises a primary learner and a secondary classifier, the primary learner comprises a lane changing working condition driving subject identification model and a steering working condition driving subject identification model, and the secondary classifier is a variable speed working condition driving subject identification model. According to the invention, the identification accuracy of the driving subject can be improved.

Description

Technical Field

[0001] The present invention relates to the field of driving safety of assisted driving vehicles, and in particular to a method and system for identifying a driving subject in an assisted driving environment. Background Art

[0002] With the support of relevant national policies, the smart car industry has developed rapidly, and the number of cars with assisted driving functions has increased rapidly. In 2022, the penetration rate of L2 passenger cars with assisted driving functions reached more than one-third, and assisted driving technology and products are accelerating into practical application. As a result, assisted driving car accidents have become more frequent, and safety issues have become prominent.

[0003] Compared with traditional cars, assisted driving cars have added active safety perception, decision-making and execution systems. In the investigation of assisted driving car accidents, the traditional driver-centered driving method is gradually shifting to a new driving mode centered on the driving system, which has caused a huge impact on the traditional motor vehicle traffic accident liability system.

[0004] In the process of investigating traffic accidents involving assisted driving vehicles, there are problems such as the difference between investigation data and real driving data, and the difficulty in obtaining real driving data under the assisted driving function in a timely and accurate manner. From the perspective of the law enforcement process of relevant units, there is no clear method and technology to solve the problem of difficulty in identifying the subject of accident responsibility due to the intervention of assisted driving functions. However, with the support of relevant policies and standards, after the L2-level passenger car terminal with assisted driving function is successfully activated, it should report the real-time information of the assisted driving vehicle within 3 minutes of operation, charging, accident alarm or power failure to the remote service and management platform at a certain time period, which makes it possible to identify the driving subject through natural driving data.

[0005] Scientifically and efficiently determining the responsible party for assisted driving car accidents has become a key issue that needs to be solved urgently. Therefore, it is necessary to conduct research on driving subjects based on readily available natural driving data through the basic data in the assisted driving system, and establish a method for determining the driving subject of assisted driving cars based on natural driving data to determine the responsible party for the accident, so as to provide support for the realization of intelligent responsibility determination based on accident determination rules. Summary of the invention

[0006] In view of the problems of difficulty in obtaining real driving data and confusion in identifying the responsible party in assisted driving car accident investigations, the present invention proposes a method and system for identifying the driving subject of an assisted driving car based on natural driving data features, which can efficiently identify the driving subject based on basic driving data.

[0007] The technical solution adopted by the present invention is:

[0008] A method for identifying a driving subject in an assisted driving environment is provided, comprising the following steps:

[0009] Obtain the vehicle's driving data, process it, and filter out lane change time, steering wheel angle, and acceleration segment data;

[0010] The filtered fragment data are input into a pre-built fusion Stacking integrated driving subject identification model to identify the probability of the driving subject corresponding to different time periods in the fragment data; the fusion Stacking integrated driving subject identification model includes a primary learner and a secondary classifier, wherein the primary learner includes a lane changing condition driving subject identification model and a steering condition driving subject identification model, and the secondary classifier is a speed changing condition driving subject identification model.

[0011] According to the above technical solution, the lane-changing driving subject identification model and the steering driving subject identification model are both constructed based on artificial neural networks.

[0012] Following the above technical solution, the driving subject identification model for speed change conditions is constructed based on a one-dimensional convolutional neural network.

[0013] Following the above technical solution, when training the steering condition driving subject identification model, the data segment driving behavior types used include four types of conditions: left turn, right turn, large radius left turn, and large radius right turn.

[0014] Following the above technical solution, the number of hidden layers of the artificial neural network is set to 8, the number of nodes of the layer neurons is set to 16, and all other parameters are default, and the lane changing condition driving subject identification model and the steering condition driving subject identification model are constructed respectively.

[0015] According to the above technical solution, the one-dimensional convolutional neural network includes an 11-layer network structure, 1 sequence input layer, 2 convolutional layers, 1 fully connected layer, 1 output layer and several activation and normalization modules.

[0016] Based on the above technical solution, the Stacking integrated driving subject recognition model is integrated based on S tac The king ensemble model is constructed to integrate the lane changing condition driving subject identification model, the steering condition driving subject identification model and the speed change condition driving subject identification model.

[0017] Following the above technical solution, when training the Stacking integrated driving subject recognition model, a 5-fold cross-validation method is used for model training, and a Bayesian optimizer is used to optimize the network hyperparameters to obtain the optimal hyperparameters.

[0018] The present invention also provides a driving subject identification system in an assisted driving environment, comprising:

[0019] A data acquisition module, used to acquire vehicle driving data;

[0020] The data processing module is used to process the acquired driving data and filter out the lane change time, steering wheel angle and acceleration fragment data;

[0021] The recognition module is used to input the filtered segment data into a pre-built fusion Stacking integrated driving subject recognition model to identify the probability of the driving subject corresponding to different time periods in the segment data; the fusion Stacking integrated driving subject recognition model includes a primary learner and a secondary classifier, wherein the primary learner includes a lane changing condition driving subject recognition model and a steering condition driving subject recognition model, and the secondary classifier is a speed changing condition driving subject recognition model.

[0022] The present invention also provides a computer storage medium, which stores a computer program that can be executed by a processor, and the computer program executes the driving subject identification method in the assisted driving environment described in the above technical solution.

[0023] The beneficial effects of the present invention are as follows: the method for distinguishing the driving subject of an assisted driving vehicle based on the characteristics of natural driving data of the present invention selects lane change time, steering wheel angle and acceleration fragment data from driving data, and after analysis, it is found that these three data have obvious differences for different driving subjects; a lane change condition driving subject identification model, a steering condition driving subject identification model and a speed change condition driving subject identification model are respectively constructed for these three data, and then the three models are optimized and assembled through the Stacking ensemble learning algorithm, and the assisted driving vehicle driving subject identification model under the optimal parameter combination is obtained through group decision-making, and the probability of the driving subject corresponding to different time periods in the fragment data is further obtained through the assembled model, thereby improving the recognition accuracy of the driving subject.

[0024] Furthermore, since the steering and lane-changing data are composed of multiple driving segments, there is no continuity between the segments, and it is impossible to perform machine learning with the time axis as a sequence. The basic structure of the artificial neural network algorithm, the artificial neuron, is composed of nonlinear change units. The information it receives is the dot product of the neuron weight corresponding to the previous driving segment. It has a strong nonlinear mapping ability and conforms to the characteristics of steering and lane-changing data. Therefore, the lane-changing driving subject identification model and the steering driving subject identification model are both constructed based on artificial neural networks. The one-dimensional convolutional neural network is a variant of the traditional convolutional neural network, which can effectively realize the classification and learning of time series data and conform to the characteristics of speed change data. Therefore, the speed change driving subject identification model is constructed based on the one-dimensional convolutional neural network.

[0025] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 is a flow chart of a method for identifying a driving subject in an assisted driving environment according to Embodiment 1 of the present invention;

[0028] Figure 2 is a schematic diagram of the framework structure of Embodiment 2 of the present invention;

[0029] Figure 3 This is a schematic diagram of the process of building a driving subject recognition model that integrates Stacking;

[0030] Figure 4 It is a schematic diagram of the integration strategy of the Stacking integrated driving subject recognition model;

[0031] Figure 5 This is a diagram of the confusion matrix of the Stacking integrated driving subject recognition model;

[0032] Figure 6 It is a schematic diagram of the integrated accuracy of the Stacking integrated driving subject recognition model. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0034] It should be noted that the illustrations provided in the embodiments of the present invention are only used to illustrate the basic concept of the present invention in a schematic manner. Therefore, the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout type may also be more complicated.

[0035] In the present invention, it is also necessary to explain that, if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, if the terms "first" and "second" appear, they are only used for description and distinction purposes, and cannot be understood as indicating or implying relative importance.

[0036] Example 1

[0037] like Figure 1 As shown, the driving subject identification method in the assisted driving environment of the embodiment of the present invention includes the following steps:

[0038] S1. Obtaining vehicle driving data;

[0039] S2, processing the acquired driving data, and filtering out lane change time, steering wheel angle, and acceleration segment data;

[0040] S3. Input the filtered fragment data into a pre-built fusion Stacking integrated driving subject identification model to identify the probability of the driving subject corresponding to different time periods in the fragment data; the fusion Stacking integrated driving subject identification model includes a primary learner and a secondary classifier, wherein the primary learner includes a lane changing condition driving subject identification model and a steering condition driving subject identification model, and the secondary classifier is a speed changing condition driving subject identification model.

[0041] In step S2, the three types of fragment data, namely lane change time, steering wheel angle and acceleration, are selected because in the entire data fragment, indicators such as lane change time and steering wheel angle are all less than 0.01, and there are obvious differences. Other data, such as the following distance indicator, show obvious volatility in the interval [0.006-0.35]. Combined with the actual situation, the difference in following distance between the two types of driving subjects (drivers and auxiliary driving systems) shows a certain periodic change with the size of traffic flow: when the traffic flow is small, there is a more obvious difference in following distance; when the traffic flow is large, the difference in following distance is lower. Considering that this patent focuses on the investigation of assisted driving accidents, the occurrence of accidents is random, so driving parameters with full periodicity characteristics should be selected, and data that does not have full periodicity and only has volatility in a certain interval, such as following distance data, should be discarded.

[0042] Furthermore, in the acquired data segments, the steering and lane changing data are composed of multiple driving segments, and there is no continuity between the segments, so it is impossible to perform machine learning with the time axis as a sequence. The basic structure of the artificial neural network algorithm, the artificial neuron, is composed of nonlinear change units. The information it receives is the dot product of the neuron weight corresponding to the previous driving segment. It has a strong nonlinear mapping ability and conforms to the characteristics of steering and lane changing data. Therefore, the lane changing condition driving subject identification model and the steering condition driving subject identification model can both be constructed based on artificial neural networks. And the number of hidden layers of the artificial neural network can be set to 8, the number of nodes of the layer neuron can be set to 16, and all other parameters are defaulted to construct the lane changing condition driving subject identification model and the steering condition driving subject identification model respectively. When training the steering condition driving subject identification model, the data segment driving behavior types used include left turn, right turn, large radius left turn, and large radius right turn, a total of four conditions.

[0043] The one-dimensional convolutional neural network is a variant of the traditional convolutional neural network, which can effectively realize the classification and learning of time series data and conform to the characteristics of speed change data. Therefore, the speed change driving subject identification model can be constructed based on the one-dimensional convolutional neural network. The one-dimensional convolutional neural network may include an 11-layer network structure, a sequence input layer, two convolutional layers, a fully connected layer, an output layer, and several activation and normalization modules.

[0044] Specifically, the Stacking integrated driving subject identification model is constructed based on the Stacking set model, and the lane changing condition driving subject identification model, the steering condition driving subject identification model and the speed change condition driving subject identification model are integrated.

[0045] When training the fusion stacking integrated driving subject recognition model, a 5-fold cross-validation method can be used for model training, and a Bayesian optimizer can be used to optimize the network hyperparameters to obtain the optimal hyperparameters.

[0046] The method for distinguishing the driving subject of an assisted driving vehicle based on the characteristics of natural driving data of the present invention selects lane change time, steering wheel angle and acceleration fragment data from driving data. After analysis, it is found that these three data have obvious differences for different driving subjects. A lane change condition driving subject identification model, a steering condition driving subject identification model and a speed change condition driving subject identification model are respectively constructed for these three data, and then the three models are optimized and assembled through the Stacking ensemble learning algorithm. The assisted driving vehicle driving subject identification model under the optimal parameter combination is obtained through group decision-making, and the probability of the driving subject corresponding to different time periods in the fragment data is further obtained through the assembled model, thereby improving the recognition accuracy of the driving subject.

[0047] In the Stacking ensemble model, the primary learners should have certain predictive capabilities when used alone, that is, they should be able to predict the target variable more accurately to a certain extent, thereby improving the data generalization ability of the ensemble model; the secondary learners should select those learners that perform stably on different data sets, while avoiding overfitting, so as to learn from the prediction results of the primary learners and make better prediction results. Considering that the posterior probabilities output by the lane-changing driving subject identification model and the steering driving subject identification model are distributed around 0 and 1, the accuracy is relatively high, while the speed-changing driving subject identification model has a low recognition accuracy, and the output posterior probability distribution is relatively uniform. This condition is presented in a time series and has good data stability. At the same time, the amount of data in this segment is much larger than that of the lane-changing and steering conditions, which can effectively prevent overfitting. Therefore, the lane-changing and steering artificial neural networks are used as models as primary learners, and the one-dimensional convolutional neural network is used as a secondary classifier to construct a fusion Stacking ensemble driving subject identification model.

[0048] Example 2

[0049] This embodiment is based on Embodiment 1, but differs in that it provides a specific process for screening out three types of data, namely, lane change time, steering wheel angle and acceleration, and provides a specific process for training each model.

[0050] like Figure 2 As shown, the vehicle's driving parameter data can be connected to the CAN bus through the vehicle-mounted OBD interface. The series of data collected in this embodiment are: vehicle speed V {v1, v2, v3, ..., v n}, vehicle speed acceleration A{a1, a2, a3, …, a n}, steering wheel angle θ{θ1, θ2, θ3, …, θ n}、Following distance T d {t d1 , t d2 , t d3 ,…,t dn}, lane change time T{t1, t2, t3, …, t n} and relative distance D{d1, d2, d3, …, d n}.

[0051] Among them, the vehicle OBD interface is a physical interface on the vehicle that can obtain CAN bus data in real time. CAN bus refers to a network bus on the vehicle that connects various sensors and components of the vehicle. Through this bus, real-time data from various sensors of the vehicle can be obtained;

[0052] Furthermore, the data is preprocessed:

[0053] First, various types of data are interpolated, and the piecewise linear interpolation method is used to fill in the missing values:

[0054] W n ={w 1n , w 2n ,···,w mn}

[0055] Where W n is a set of data indicators; n is the number of data indicators; m is the number of elements of feature indicators;

[0056] Among them, the missing value W in =N i-1n w i-1n +N i+1n w i+1n ,i=(0,1,2,···,m)

[0057] In the formula,

[0058] According to the sampling rate Fs, various types of data are interpolated to obtain a series of data (in this embodiment, the sampling rate F s =15Hz):

[0059] Vehicle speed V1{v Δt1 , v Δt2 , v Δt3 , …, v Δtn}、

[0060] Vehicle speed acceleration A1{a Δt1 , a Δt2 , a Δt3 , …, a Δtn}、

[0061] Steering wheel angle θ1{θ Δt1 ,θ Δt2 ,θ Δt3 ,…,θ Δtn}、

[0062] Following distance T d1 {T dΔt1 , T dΔt2 , T dΔt3 ,…,T dΔtn}、

[0063] Lane changing time T{t Δt1 , t Δt2 , t Δt3 ,…,t Δtn}

[0064] Relative distance D1{d Δt1 , d Δt2 , dΔt3 , …, d Δtn}.

[0065] in

[0066] Furthermore, the new data set obtained by interpolation is cleaned, and the MAD algorithm is used to detect outliers on the cleaned driving data, and the median of the data is calculated:

[0067]

[0068] Calculate MAD:

[0069]

[0070] In the formula, MAD is the set of characteristic indicators W n Subtract the characteristic index W from each element in n The median of the absolute value of the median of the set, b is a constant, usually b = 1.48;

[0071] In order to detect outliers in the observed data, it is necessary to calculate each observation value w mn The coefficient of determination is:

[0072]

[0073] When the determination coefficient T exceeds the given threshold d, the observation value w is considered mn For abnormal data, according to a large number of scientific experiments and engineering practices, T=2.5 is taken to eliminate abnormal data.

[0074] Clean the new data set obtained by interpolation. If T>2.5, delete the data w mn If the condition is not met, the data is retained. After this step, the data set is obtained, and after normalization, the final data set is obtained: vehicle speed V2{v Δt1 , v Δt2 , v Δt3 , …, v Δtn}、Vehicle speed acceleration A2{a Δt1 , a Δt2 , a Δt3 , …, a Δtn}、Steering wheel angle θ2{θ Δt1 ,θ Δt2 ,θ Δt3 ,…,θ Δtn}、Following distance T d2 {T dΔt1 , T dΔt2 , T dΔt3 ,…,T dΔtn}、Lane changing time T2{tΔt1 , t Δt2 , t Δt3 ,…,t Δtn}, relative distance D2{d Δt1 , d Δt2 , d Δt3 , …, d Δtn}.

[0075] For the above six data sets, a K-means clustering model is established through learning the differences in driving subject behaviors for classification.

[0076] The value of the number of categories K is determined according to the SSE (Sum of Squared Error) indicator, and K cluster centers are randomly initialized. The specific formula is as follows:

[0077]

[0078] Where: C k is a cluster; m k is the cluster center; w p Cluster C k The sample points in ; k = 1,…,K.

[0079] Each data point in the sample data is divided into the cluster with the closest cluster center.

[0080] After all the data are distributed, the cluster center of each cluster is recalculated based on all the data in the cluster to obtain its average value.

[0081] Finally, iterate again to redistribute the data in the sample data set and continuously update until the cluster center no longer changes.

[0082] Furthermore, according to the clustering results, the clustering parameters are statistically analyzed. The parameters are the mean, standard deviation, maximum and minimum values ​​of lane change time, speed, acceleration, following distance, and relative distance to the front vehicle. The variance test is performed on the parameters with obvious differences. The calculation formulas for the mean, standard deviation, maximum and minimum values ​​of the statistical analysis are as follows:

[0083]

[0084] Among them, i is a single sample; x i are the lane-changing sample parameters, and n is the total number of samples.

[0085] Furthermore, the clustering parameters are statistically analyzed for differences:

[0086] Calculate the driving time periods t1, t2, ···, t n Mean relative distance between vehicles Standard deviation σ1, σ2, ···, σ n , median M1, M2, ..., M n Etc. features to construct feature sets;

[0087] Will Following distance samples D1{d Δt1 , d Δt2 , d Δt3 ,···,d Δtn}Arrange in ascending order;

[0088] Calculate the average In the formula: i is the sample individual, n is the total number of samples;

[0089] Calculate standard deviation The maximum value X max =max(x i ), minimum value X min =min(x i );

[0090] Furthermore, the non-parametric Kruskal-Wallis H test (KW test method) was used to test for insignificant differences.

[0091] The basic steps of the KW test are to convert the data into the rank statistic R i , with R i (i=1, ..., n) represents the nth sample in the data i Observations, such as: (d Δt1 , …, d Δtn ) can be expressed as (R t1 , …, R tn ), thereby calculating the average rank of each group of samples to determine the test statistic H value and correct H c , and then look up the table to obtain the P value for testing. The H value calculation formula is:

[0092]

[0093] Mean values ​​of lane change time, speed, acceleration, following distance, and relative distance to the vehicle ahead (within 95% confidence interval).

[0094] According to the significance level, the statistical results were subjected to KW test, and the driving data of the driver and the assisted driving car were compared. It was found that the medians of v and x0 were significantly different, 0.023 and 0.532 respectively, which were greater than 0.01, and there was no significant difference;

[0095] The indicators such as lane change time and steering wheel angle are all less than 0.01, which shows differences;

[0096] Furthermore, the wavelet time-frequency difference analysis is performed on the acceleration data of the two. The wavelet transform can perform fine processing on the time-frequency of the signal, such as signal f (t) , r0 is the starting point of the wavelet to be measured, then its series expansion is as follows:

[0097]

[0098] Among them, r is the scale; t is the time; s is the displacement, is the approximate coefficient, d r (s) is the detail coefficient. is the scaling function, is a wavelet function.

[0099] The captured driving period data are numbered in chronological order and compiled into the original signal sequence x(l), which is sampled L times in one driving cycle. The specific formula is:

[0100]

[0101] Where: x(k) is the data after discrete Fourier transform of L points; x(l) is the original signal sequence; L is the number of periodic sampling points; l is the sampling point; k is the frequency parameter;

[0102] The frequency of acceleration during the driving period can be obtained by using the acceleration spectrum diagram after discrete Fourier transformation;

[0103] By capturing the driving data of the driving period and the similarity between each sample in the sample set, the driving period corresponding to the acceleration can be extracted.

[0104] The results show that there is a significant difference in acceleration between human drivers and assisted driving systems;

[0105] Through the above analysis, it can be seen that the parameters such as lane change time, steering wheel angle, acceleration, etc. in the driving clips of the driver and the assisted driving system are obviously different and consistent with objective facts, which can be used to establish the next step of driving subject identification model.

[0106] Furthermore, a sample set was constructed based on natural driving data, and an artificial neural network (ANN) algorithm and a one-dimensional convolutional neural network model (1D-Convolutional Neural Network) were used to establish driving subject identification models based on different driving parameters. The artificial neural network models were constructed with the steering and lane changing data of the vehicle as input, and the one-dimensional convolutional neural network model was constructed with the speed change data as input.

[0107] In order to put the artificial neural network model of the above method into specific application, an embodiment of driving subject identification based on natural driving data is now given.

[0108] This embodiment uses lane change time and steering wheel angle as the first layer of feature data processed by two ANN algorithms. The number of artificial neurons in the input layer is consistent with the feature dimension of the input data. The input layer contains k nodes, which transmit feature X^ and weight W^ to the output node. The model output is the driving segment number corresponding to the driving period (the driving segment number after labeling). The input layer itself does not perform any calculations.

[0109] Different network hyperparameters have a great impact on the classification accuracy and performance of various indicators of the network model. It is necessary to optimize the model's hyperparameters according to different input data and classification objects in order to obtain the best model effect.

[0110] Firstly, the steering wheel angle fragment data in the lateral driving behavior is used as input, and the artificial neural network model is combined to construct a steering condition driving subject identification model.

[0111] The preprocessed data is constructed by setting the time window length and using a sliding window to construct samples and a sample set, and the original data set is set to D = {(x i ,y i ):x i ∈R m , ∈R}, which contains n driving trips and m driving behavior features.

[0112] Furthermore, the data segment driving behavior types used in this example include four types of conditions: left turn, right turn, large radius left turn, and large radius right turn. The distribution of the data used is shown in Table 1.

[0113] Table 1 Sample numbers of some steering driving conditions

[0114]

[0115] In this step, the driving behavior sample data is divided into a test set and a training set in a ratio of 3:7. 30% of the entire trip is randomly selected as the test set, and the remaining 70% of the trip is used as the training set and validation set. The neural network model is trained using a 5-fold cross-validation method.

[0116] This embodiment constructs an ANN model by setting the number of hidden layers to 8, the number of nodes of the neurons in this layer to 16, and all other parameters to default. The output formula of neuron i in the hidden layer is:

[0117]

[0118] Where: xi (i=1, 2, 3, ..., n) is the input value of the i-th neuron; w ij is the connection weight from the i-th neuron to the j-th neuron; b is the threshold of this neuron; F is the activation function; and Y is the output of the artificial neuron.

[0119] In this example, the driver identification result learned the best hyperparameters at the 25th iteration. The optimized network hyperparameters are:

[0120] The number of hidden layers in the artificial neural network is 2, and the activation function of each layer is Rectified Linear Unit (ReLU); the regularization strength is 3.5×10 -5 ; The number of columns in the weight matrix of the first layer of artificial neurons is 71; The number of columns in the weight matrix of the second layer of artificial neurons is 46.

[0121] The prediction speed of the model is 12000 obs / s, and the prediction accuracy is 89.0%. The model effect is tested using test data that is not involved in the model training process. The test accuracy is 89.6%. The test accuracy is close to the training accuracy, and there is no overfitting phenomenon.

[0122] Furthermore, the lane-changing segment data in the lateral driving behavior is used as input and combined with the artificial neural network model to construct a lane-changing driving subject identification model. The process is the same as described above.

[0123] The data segment condition type used is lane change, and the data distribution used is shown in Table 2.

[0124] Table 2 Number of lane-changing driving clip samples

[0125]

[0126] This model has undergone an automatic optimization process during training. The optimized network hyperparameters are: number of fully connected layers: 3; activation function: Tanh; regularization strength: 1×10 -5 ; First layer size: 12; Second layer size: 142; Third layer size: 297.

[0127] This model has undergone an automatic optimization process during training, and the optimized network hyperparameters are:

[0128] The number of hidden layers in the artificial neural network is 3. The activation functions of the first and second layers are ReLU, and the activation function of the third layer is hyperbolic tangent function (Tanh); the regularization strength is 3.5×10 -5; The number of columns in the weight matrix of the first layer of artificial neurons is 71; The number of columns in the weight matrix of the second layer of artificial neurons is 46.

[0129] The prediction speed of the model is 5700 obs / s, and the prediction accuracy is 85.8%. The model effect is tested using test data that is not involved in the model training process. The test accuracy is 86%, which is close to the training accuracy.

[0130] Furthermore, the speed change segment data in the longitudinal driving behavior is used as input to construct a driving subject identification model for the speed change condition.

[0131] The data segment condition type used is speed change, and the data distribution used is shown in Table 3:

[0132] Table 3 Number of longitudinal driving condition samples

[0133]

[0134] Furthermore, considering that acceleration and deceleration behaviors are more frequent and relatively simple, and the speed change characteristics are significantly lower than the above-mentioned lateral driving behavior parameters, this model uses a one-dimensional convolutional neural network algorithm to construct a driving subject identification model for longitudinal driving conditions.

[0135] In this speed-changing driving subject identification model, an 11-layer network structure is constructed, including a sequence input layer, two convolutional layers, a fully connected layer, an output layer, and several activation and normalization modules. The specific structure and number of parameters are shown in Table 4.

[0136] Table 4 One-dimensional convolutional neural network structure parameters

[0137]

[0138] Furthermore, the driving parameters of the driving subject can be extracted by capturing the similarity between the driving subject's speed change data and each sample in the sample set.

[0139] In this embodiment, a sliding time window method is used to sample the preprocessed and filtered data, the time window length is W_L, and the time window overlap is O_V. The time window moves along the time axis for sequential sampling.

[0140] The data samples are input into the convolutional neural network. The input of the convolutional neural network is a two-dimensional data matrix, and the network output is the matching driving segment number.

[0141] The accuracy of the speed change recognition of the trained driver is higher than 80%. The test accuracy is similar to the training accuracy, and there is no overfitting phenomenon.

[0142] Furthermore, based on the output results and model characteristics of the steering condition driving subject identification model, lane change condition driving subject identification model and speed change condition driving subject identification model constructed above, combined with the actual situation that the above two or three driving conditions will appear in natural driving experiments in all driving periods, the Stacking ensemble model is selected as the integrated strategy for constructing the driving subject identification model. Figure 3 , 4 As shown, the specific steps are as follows:

[0143] The output data of the driving subject identification model constructed above based on specific driving conditions is W n×1 ={w1, w2, w3, ..., w n}, where n is the total number of IDs for identifying driving entities, w i is the posterior probability that the test data belongs to the i-th driver. The predictor variable data is in the form of a 1×84 numerical matrix, and the response variable is the driver ID.

[0144] Furthermore, different operating condition segments belonging to the same driver are randomly arranged to form training samples to simulate the arrangement and combination of operating condition segments that may occur in actual scenarios.

[0145] Furthermore, a neural network model is built to integrate the output of the driving subject identification model, the neural network is trained using a 5-fold cross-validation method, and the network hyperparameters are optimized using a Bayesian optimizer, so that the model learning obtains the optimal hyperparameters.

[0146] According to the above steps, considering that the speed change driving subject identification model has a low identification accuracy, the output posterior probability distribution is relatively uniform. The posterior probabilities output by the lane change driving subject identification model and the steering driving subject identification model are distributed around 0 and 1. Therefore, the lane change and steering artificial neural networks are used as models as primary learners, and the one-dimensional convolutional neural network is used as a secondary classifier to construct a fusion Stacking integrated driving subject identification model. The algorithm structure is as follows:

[0147]

[0148]

[0149] Where: W is the output data set result of the artificial neural network model, x m is the characteristic attribute of the data set, y m is the target value attribute of the data set, m is the number of fusion indicators of the data set, and x in this model m Lane changing, turning, and speed changing characteristics at different time periods, y m It is human driving state and assisted driving state; Tis the primary learning classifier, T is the final number of primary learning classifiers, ξ1 in this model is the lane changing artificial neural network model, ξ2 is the steering artificial neural network model, T in this model is 2; ζ is the secondary learning classifier, and ζ is LR; h t is the trained primary learner, t is the number variable of primary classifiers; D` is the prediction value generated by cross-validation of the primary learner for the training set of the secondary classifier; Z it is the predicted value obtained by the primary learner for the training set in the original data set; x i is the training set of the original data set; y i is the predicted value of the test set in the original data set; h` is the secondary learner trained by the secondary learner; H(x) is the final output result.

[0150] Furthermore, according to the above Stacking ensemble learning model, the confusion matrix of the output classification result is obtained (such as Figure 5 As shown in the figure, the recognition accuracy of the driver in each driving period is obtained. It can be seen that under the three working conditions, the recognition accuracy after integration is improved compared with that before integration. The specific results are as follows Figure 6 shown.

[0151] In summary, this embodiment uses natural driving data with more real and comprehensive driving information to extract driving characteristics of natural driving data in different driving periods, and has good universality. Three characteristic parameters that can characterize driving characteristics, namely lane change time, steering wheel angle, and acceleration, are determined from the two aspects of vehicle status and vehicle operation, and used as input variables for two types of neural network driving subject identification models, fully considering the interaction characteristics between people, vehicles, and roads, with theoretical basis and in line with objective facts. The Stacking integrated recognition model used has excellent recognition effect, high recognition accuracy, and can achieve higher recognition accuracy by using shorter time series. Compared with other driver identity recognition models, it has better stability and also has higher recognition accuracy for driving condition data with longer time intervals.

[0152] Example 3

[0153] This embodiment is based on embodiment 1. The driving subject recognition system in the assisted driving environment of this embodiment is mainly for realizing the above method embodiment. The system includes:

[0154] A data acquisition module, used to acquire vehicle driving data;

[0155] The data processing module is used to process the acquired driving data and filter out the lane change time, steering wheel angle and acceleration fragment data;

[0156] The recognition module is used to input the filtered segment data into a pre-built fusion Stacking integrated driving subject recognition model to identify the probability of the driving subject corresponding to different time periods in the segment data; the fusion Stacking integrated driving subject recognition model includes a primary learner and a secondary classifier, wherein the primary learner includes a lane changing condition driving subject recognition model and a steering condition driving subject recognition model, and the secondary classifier is a speed changing condition driving subject recognition model.

[0157] Each module is mainly used to implement each step of the method, which will not be described in detail here.

[0158] Example 4

[0159] The present application also provides a computer-readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a disk, an optical disk, a server, an App application store, etc., on which a computer program is stored, and when the program is executed by a processor, a corresponding function is realized. When the computer-readable storage medium of this embodiment is executed by a processor, the driving subject recognition method in the assisted driving environment of the method embodiment is realized.

[0160] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0161] The order of execution of each step in the above embodiment does not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0162] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A method for identifying a driving subject in an assisted driving environment, characterized in that: The following steps are involved: Obtain vehicle driving data; Process the acquired driving data and filter out lane change time, steering wheel angle and acceleration segment data; The filtered fragment data are input into a pre-built fusion Stacking integrated driving subject identification model to identify the probability of the driving subject corresponding to different time periods in the fragment data; the fusion Stacking integrated driving subject identification model includes a primary learner and a secondary classifier, wherein the primary learner includes a lane changing condition driving subject identification model and a steering condition driving subject identification model, and the secondary classifier is a speed changing condition driving subject identification model.

2. The method for identifying a driving subject in an assisted driving environment according to claim 1, characterized in that: The lane-changing driving subject identification model and the steering driving subject identification model are both constructed based on artificial neural networks.

3. The method for identifying a driving subject in an assisted driving environment according to claim 1, characterized in that: The driving subject identification model for speed change conditions is constructed based on a one-dimensional convolutional neural network.

4. The method for identifying a driving subject in an assisted driving environment according to claim 1, characterized in that: When training the steering condition driving subject identification model, the data segments used for driving behavior types include left turn, right turn, large radius left turn, and large radius right turn, a total of four conditions.

5. The method for identifying a driving subject in an assisted driving environment according to claim 2, characterized in that: The number of hidden layers of the artificial neural network was set to 8, the number of nodes of the layer neurons was set to 16, and all other parameters were kept as default. The lane-changing driving subject identification model and the steering driving subject identification model were constructed respectively.

6. The method for identifying a driving subject in an assisted driving environment according to claim 3, characterized in that: The one-dimensional convolutional neural network consists of an 11-layer network structure, 1 sequence input layer, 2 convolutional layers, 1 fully connected layer, 1 output layer, and several activation and normalization modules.

7. The method for identifying a driving subject in an assisted driving environment according to claim 1, characterized in that: The Stacking integrated driving subject identification model is built based on the Stacking set model, and integrates the lane changing condition driving subject identification model, the steering condition driving subject identification model and the speed change condition driving subject identification model.

8. The method for identifying a driving subject in an assisted driving environment according to any one of claims 1 to 7, characterized in that: When training the fusion stacking integrated driving subject recognition model, a 5-fold cross-validation method is used for model training, and the Bayesian optimizer is used to optimize the network hyperparameters to obtain the optimal hyperparameters.

9. A driving subject recognition system in an assisted driving environment, characterized in that: include: A data acquisition module, used to acquire vehicle driving data; The data processing module is used to process the acquired driving data and filter out the lane change time, steering wheel angle and acceleration fragment data; The recognition module is used to input the filtered segment data into a pre-built fusion Stacking integrated driving subject recognition model to identify the probability of the driving subject corresponding to different time periods in the segment data; the fusion Stacking integrated driving subject recognition model includes a primary learner and a secondary classifier, wherein the primary learner includes a lane changing condition driving subject recognition model and a steering condition driving subject recognition model, and the secondary classifier is a speed changing condition driving subject recognition model.

10. A computer storage medium, characterized in that: A computer program executable by a processor is stored therein, and the computer program executes the driving subject identification method in an assisted driving environment as described in any one of claims 1 to 8.

Images