Data-driven vehicle following behavior prediction method based on driver characteristics

Abstract

The present invention discloses a data-driven vehicle following behavior prediction method based on driver characteristics, comprising the following steps: respectively obtaining mutation points of the speed change of two vehicles over time, estimating the instantaneous reaction delay of the current vehicle according to the speed mutation points; dividing the driving behavior according to the intrinsic characteristics of the current vehicle's multidimensional time series, clustering the divided driving behaviors, and calculating the proportion of each driving mode respectively; using the gated recurrent unit GRU as a prediction model, the input of the prediction model includes a multidimensional time series related to the current vehicle's following behavior, and the behavioral characteristics of the current driver, and the prediction target is the forward distance of the current vehicle in the next period; the length of the multidimensional time series is determined by the instantaneous reaction delay. The present invention also incorporates the driver's reaction delay and driving style into the following model, mines the reaction delay and driving mode based on actual data, and performs differentiated modeling for different drivers to more accurately estimate the driver's following behavior.

Description

Technical Field

[0001] The present invention relates to the field of microscopic traffic flow, and in particular to a data-driven vehicle following behavior prediction method based on driver characteristics. Background Art

[0002] As my country's urbanization process accelerates, the number of motor vehicles increases, and transportation demand further increases. Roads and related infrastructure are also constantly developing. At present, emerging technologies represented by big data have become a new driving force for the development of my country's smart transportation industry. Under this strategic background, vehicles are developing in the direction of intelligence and automation, and data-driven following models should provide technical support for the development of autonomous driving related industries.

[0003] The data-driven car-following model is an important part of microscopic traffic flow theory. It does not require assumptions about microscopic traffic behavior, is more flexible to apply, and can take into account a wider range of information. It has a wide range of application value in the fields of capacity analysis, traffic safety evaluation, active safety control, automatic driving, and assisted driving systems. With the rich resources of traffic big data and the further development of traffic data analysis and mining technology, it is of great significance to reshape the modeling ideas of the car-following model in the big data environment and realize fine modeling.

[0004] The existing following models have achieved good results in the field of autonomous driving, but they do not take the driver's reaction delay and driving style into consideration. Unlike human drivers, autonomous vehicles collect and analyze data based on multiple sensors to make driving behavior choices, which often involve different data sources and data types. In mixed traffic flows, future autonomous vehicles will need to specifically judge the following behavior of human drivers; for assisted driving systems, accurately judging the driving behavior of specific drivers is conducive to accurate prompts and warnings, and there are still few existing following models based on differentiated modeling. Summary of the invention

[0005] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a data-driven vehicle following behavior prediction method based on driver characteristics, while incorporating the driver's reaction delay and driving style into the following model, performing differentiated modeling for different drivers, and more accurately predicting the driver's following behavior.

[0006] Technical solution: A data-driven vehicle following behavior prediction method based on driver characteristics of the present invention comprises the following steps:

[0007] Step 1: According to the actual trajectory data of the current vehicle and the preceding vehicle in the microscopic traffic flow, the mutation points of the speed changes of the two vehicles over time are obtained respectively, and the instantaneous reaction delay of the current vehicle is estimated according to the speed mutation points;

[0008] Step 2: divide the driving behavior according to the intrinsic characteristics of the current vehicle multidimensional time series, cluster the divided driving behaviors, obtain three driving modes, calculate the proportion of each driving mode, and use the proportion of each driving mode to represent the behavioral characteristics of the current driver;

[0009] Step 3, using the gated recurrent unit GRU as the prediction model. The input of the prediction model includes the multidimensional time series related to the current vehicle's following behavior and the current driver's behavior characteristics. The prediction target is the forward distance of the current vehicle in the next period. The length of the multidimensional time series is determined by the instantaneous reaction delay.

[0010] Furthermore, the mutation points of speed changes over time include acceleration mutation points and deceleration mutation points. After obtaining the mutation points of the speed changes over time of the two vehicles, the relationship between the speed mutation points of the two vehicles is matched, and the time difference between the two when the speed mutation of the leading vehicle causes the speed mutation of the following vehicle after matching is calculated, which is the instantaneous reaction delay.

[0011] Furthermore, the multidimensional time series can be expressed as Indicates that the time series data contains information of T moments, each moment has three dimensions, including the speed of the vehicle at each moment ν i (t), headway distance s i (t), and the speed difference Δν with the preceding vehicle i (t).

[0012] Further, the step 2 specifically includes:

[0013] Step 201, assuming that the characterization vector of each moment of the multidimensional time series obeys a multidimensional Gaussian distribution, a greedy Gaussian partitioning algorithm is used to obtain the optimal partitioning point sequence of the multidimensional time series, and the parameter mean of each subsequence after the multidimensional time series is partitioned;

[0014] Step 202, using the parameter mean to characterize each time interval, clustering the time intervals using the K-means method, and obtaining three categories by clustering, including three driving modes: normal, cautious, and aggressive, which are represented by numbers 0, 1, and 2, respectively;

[0015] Step 203, calculate the time proportion of each driving mode respectively, expressed as

[0016] Further, the step 3 specifically includes: starting from time t, vehicle i inputs the multidimensional time series of the GRU model Indicates that v i (t) represents the speed of vehicle i at time t; s i (t) represents the headway distance of vehicle i at time t; Δv i (t) = vi (t)-v i-1 (t) represents the speed difference between vehicle i and the preceding vehicle at time t; Δx i-1 (t-1) is the change in the position of the preceding vehicle, indicating the distance the preceding vehicle has traveled from time t-1 to time t; dv i-1 (t) = v i-1 (t)-v i-1 (t-1) is the speed change of the preceding vehicle, indicating the speed difference between the preceding vehicle at time t and time t-1; τ is the instantaneous reaction delay of the driver;

[0017] Another input of the GRU model is the time proportion of each driving mode d i , d i It is fused with the last hidden layer of the GRU model to output the forward distance of the current vehicle in the next period.

[0018] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the present invention simultaneously incorporates the driver's reaction delay and driving style into the following model, mines the reaction delay and driving mode based on actual data, performs differentiated modeling on different drivers, and more accurately estimates the driver's following behavior; the method of the present invention has strong theoretical and operationality, relies on traffic flow big data, integrates the driver's reaction delay and driving style into the data-driven following model, and has certain application prospects in the warning of assisted driving systems, and the prediction and judgment of autonomous driving vehicles under mixed traffic flow conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the vehicle following behavior prediction method of the present invention;

[0020] Figure 2 It is the average reaction delay distribution map of all vehicles in the NGSIM dataset;

[0021] Figure 3 This is the multi-dimensional time series partition result diagram of vehicle number 2924;

[0022] Figure 4 The driving mode for vehicle number 2924;

[0023] Figure 5 This is a diagram of the multi-step prediction effect of the model for changing vehicle attributes. DETAILED DESCRIPTION

[0024] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0025] like Figure 1As shown in the flow chart, the data-driven vehicle following behavior prediction method based on driver characteristics described in this embodiment includes the following steps:

[0026] Step 1: According to the actual trajectory data of the current vehicle and the preceding vehicle in the microscopic traffic flow, the mutation points of the speed changes of the two vehicles over time are obtained respectively, and the instantaneous reaction delay of the current vehicle is estimated according to the speed mutation points.

[0027] Estimate the instantaneous reaction delay of the vehicle based on the data, use the speed time series of the leading vehicle and the following vehicle as the basis, find the mutation point of speed change over time, that is, the position of the vehicle's brake signal, and study the brake signal delay of the following vehicle compared to the leading vehicle as the instantaneous reaction delay. The mutation points of speed change over time include acceleration mutation points and deceleration mutation points, which correspond to the speed mutation situations that occur during acceleration and deceleration, respectively. Since the change in vehicle speed is not only related to the driving state of the leading vehicle, but also has a certain degree of randomness, when looking for mutation points, they need to be screened to a certain extent. In this embodiment, taking 3s as an example, if there are multiple speed mutation points within the range of 3s, only the mutation point with the largest mutation is retained.

[0028] There is a certain correspondence between the speed mutation points of the leading vehicle and the following vehicle. After obtaining the mutation points of the speeds of the two vehicles over time, the speed mutation points of the two vehicles are matched. The time difference between the two when the speed mutation of the leading vehicle causes the speed mutation of the following vehicle is calculated, which is the instantaneous reaction delay. Their correspondence is not strictly one-to-one, so they need to be identified and matched.

[0029] This embodiment uses high-precision actual vehicle trajectory data Next Generation Simulation (NGSIM) data for algorithm construction and case analysis. The NGSIM data set is initially collected by a camera at a fixed position, and then information is extracted through video processing. The sampling frequency is 0.1 seconds, and each sample includes the vehicle's instantaneous speed, acceleration, longitudinal and lateral position coordinates, vehicle length, vehicle type and other information. The trajectory of a vehicle in a following state is selected, and the mutation point where the speed changes over time is found. The mutation point relationship is matched to obtain the instantaneous reaction delay. After calculation, the average reaction delay of all vehicles in the data set is 1.67s, the minimum value is 0.9s, and the standard deviation is 2.43. Its distribution is as follows: Figure 2 The mean value of the instantaneous reaction delay during the operation of the following vehicle is taken as the length of the multidimensional time series in the subsequent prediction model.

[0030] Step 2: Divide the driving behaviors according to the intrinsic characteristics of the current vehicle multidimensional time series, cluster the divided driving behaviors, obtain three driving modes, and calculate the proportion of each driving mode to represent the behavioral characteristics of the current driver.

[0031] Step 201, assuming that the characterization vector of each moment of the multidimensional time series obeys a multidimensional Gaussian distribution, a greedy Gaussian partitioning algorithm is used to obtain an optimal partitioning point sequence of the multidimensional time series, the optimal partitioning point sequence is a set of multidimensional Gaussian distribution parameter change points, and the parameter mean of each subsequence after the multidimensional time series is partitioned is obtained;

[0032] The multidimensional time series is defined as:

[0033] Indicates that the time series data has information at T moments, and each moment has three dimensions, including the speed ν of the vehicle at each moment i (t), headway distance s i (t), and the speed difference Δν with the preceding vehicle i (t). After the multidimensional time series is divided, each set of driving data is divided into several shorter components, each of which is a driving mode of the current driver. Based on the greedy Gaussian partitioning algorithm, the multidimensional time series that reflects the driver's car-following behavior decision is divided. 1 ,b 2 ,…,b K Divide into K+1 subsequences, the division points are a set of increasing sequences that meet the following conditions:

[0034] 1 = b 0 <b 1 <...<b K <b K+1 =T

[0035] Assuming that the representation vector of each moment of the multidimensional time series obeys a multidimensional Gaussian distribution, we have And the distribution function (μ t ,Σ t ) The change point is at the dividing point b 1 ,b 2 ,...,b K Place.

[0036] The goal of multidimensional time series partitioning is to find the best partition point sequence, establish the likelihood function of the partition point sequence, and obtain the by-product (μ (k) ,Σ (k) ) sequence, find the best partition point sequence to maximize the likelihood function value. The expression of the likelihood function is as follows:

[0037]

[0038] Ignore the vehicle number i, and have:

[0039]

[0040] In the formula, det means to find the determinant of the matrix.

[0041] In practical applications, if the data dimension is too large and the number of data points is small, the maximum likelihood estimation result of the covariance will produce a large error. Therefore, a regularization term about the covariance is added to the likelihood function, and the regularization parameter λ is introduced to obtain a new objective function Φ, which is expressed as:

[0042]

[0043] Wherein, the regularization parameter λ≥0, and Tr represents the sum of the eigenvalues ​​of the matrix.

[0044] At the known dividing point b 1 ,b 2 ,…,b K In the case of , the optimal value of the mean of subsequence k is the empirical mean of this sequence, and the optimal value of the covariance is the sum of the empirical covariance of this sequence and the regularization term.

[0045] Among them, S (k) is the empirical covariance of subsequence k, expressed as:

[0046]

[0047] Substituting the mean and variance into the objective function, we get the following objective function:

[0048]

[0049] In the formula, C = -(TK / 2)(2π+1), is a constant, and Ψ(b k-1 ,b k ) is defined as:

[0050]

[0051] Therefore, the objective function is simplified to:

[0052]

[0053] Find the best partition point sequence to maximize the objective function.

[0054] The greedy Gaussian partitioning algorithm is used for solving the problem. The core idea is to add a partitioning point to the sequence in each iteration, and then adjust all the previously added partitioning points to maximize the objective function. The greedy Gaussian partitioning algorithm requires a hyperparameter, which is the maximum number of partitioning points k of the time series data. The hyperparameter k is determined by the following steps: (1) Divide the total driving time of the vehicle by a certain time interval ΔT to obtain the maximum number of partitioning points; (2) Calculate the partitioning scheme from top to bottom and output the likelihood function value under different numbers of partitioning points; (3) Take 80% of the maximum likelihood function value as the threshold to obtain the corresponding number of partitioning points.

[0055] In this embodiment, ΔT is set to 5s. The vehicle trajectory data numbered 2924 is selected as an example. The number of division points is initially selected to be 20. The division results are shown as follows: Figure 3 As shown in the figure. The two solid lines in the figure represent the actual data and the estimated data mean of the current time interval. In addition to the mean, the dotted line plots the range of the mean plus or minus one standard deviation at each time interval. The same partition result has different manifestations in three dimensions. Obviously, the partition result is consistent with the actual change trend of the data, and the range of the standard deviation is small in each time interval.

[0056] From the perspective of time, the driver may switch between different driving modes or maintain a certain driving mode. Each driving mode has certain similarities, so it is necessary to cluster a large number of driving modes after dividing the time series.

[0057] In step 202, each time interval is characterized by a parameter mean, and the time intervals are clustered using a K-means method to obtain three categories, including normal, cautious, and aggressive driving modes, which are represented by numbers 0, 1, and 2, respectively.

[0058] In this embodiment, the cluster centers are shown in Table 1. The driving mode numbered 0 selects a smaller driving speed when the headway is small, and the speed is slower than the vehicle in front, and the driving strategy is normal; the driving mode numbered 1 is when the headway is large, the driving speed is still slower than the vehicle in front, and the driving strategy is cautious; the driving mode numbered 2 is when the headway is small, and the driving strategy is aggressive.

[0059] Table 1 Driving mode clustering results

[0060]

[0061]

[0062] The trajectory data of vehicle No. 2924 was extracted. As time went by, its driving mode showed the following characteristics: Driving mode 0 (normal) was the main mode, accounting for about 46.7%, interspersed with multiple short-duration driving modes 2 (aggressive), and one long-duration driving mode 1 (cautious). This is consistent with the actual situation: for safety reasons, drivers will not continue to choose aggressive driving strategies, but will occasionally chase the vehicle in front when conditions permit. In the following state, this behavior usually lasts for a short time; when driving conditions are greatly restricted, drivers will choose a cautious strategy, and this behavior usually lasts for a long time. The driving mode prototype (V_ID-2924) is as follows: Figure 4 shown.

[0063] Step 203, calculate the time proportion of each driving mode respectively, expressed as and exists There is an obvious relationship between the three, so in the model input,

[0064] Each driver's strategy is not static, but varies with road traffic conditions and driver behavior. And the duration of the three driving modes of different vehicles varies, so the proportion of the three driving modes is used to represent the characteristics of different drivers.

[0065] Step 3, using the gated recurrent unit GRU as the model, the input is the multidimensional time series related to the current vehicle's following behavior and the current driver's behavior characteristics, and the prediction target is the forward distance of the current vehicle in the next period of time; the length of the multidimensional time series is determined by the instantaneous reaction delay. In order to make accurate predictions, the average instantaneous reaction delay during the vehicle's driving process can be selected.

[0066] Starting from time t, the multidimensional time series of vehicle i input into the GRU model is:

[0067]

[0068] in v i (t) represents the speed of vehicle i at time t; s i (t) represents the headway distance of vehicle i at time t; Δv i (t) = v i (t)-v i-1 (t) represents the speed difference between vehicle i and the preceding vehicle at time t; Δx i-1 (t-1) is the change in the position of the preceding vehicle, indicating the distance the preceding vehicle has traveled from time t-1 to time t; dv i-1 (t) = v i-1 (t)-v i-1(t-1) is the change in the speed of the preceding vehicle, indicating the speed difference between the preceding vehicle at time t and time t-1; τ is the instantaneous reaction delay of the driver.

[0069] Another input of the GRU model is the time proportion of each driving mode d i , d i It is fused with the last hidden layer of the GRU model to output the forward distance of the current vehicle in the next period.

[0070] The specific structure of the GRU model is shown in Table 2. The model input consists of two parts. One of them is the multidimensional time series related to the current vehicle following behavior, where the sequence length is variable and depends on the vehicle reaction delay. The second part is the current driver's behavior characteristics, which are clustered with the driving mode proportion d i The final output prediction target is the distance that the current vehicle will travel in the next period of time.

[0071] Table 2 GRU prediction model structure

[0072]

[0073] In order to verify the effect of the vehicle following behavior prediction method in the present invention, further evaluation and analysis are carried out, mainly from the two aspects of single-step prediction and multi-step prediction, and the Intelligent Driver Model (IDM) with excellent performance is selected as the benchmark model. Among them, the single-step prediction is to predict the following behavior of the next time step given the current state of the vehicle under study and other necessary phenomena. The prediction of each time step is based on real data, and the single-step prediction uses the average absolute error of the current vehicle's speed and position as the core metric. Multi-step prediction is to predict the vehicle following behavior of multiple time steps in the future given the initial state of the vehicle under study and other necessary phenomena. The prediction of each time step is based on the prediction result of the previous time step, and then the running state of the following vehicle is updated iteratively, and the actual information of the leading vehicle and the position and speed information of other surrounding vehicles are used as new inputs to predict the following behavior of the research object in the future. The multi-step prediction uses the Euclidean distance (EUC), Fréchet distance (FRE), area error ratio (QE) and the farthest moment position prediction accuracy (FD) between the actual vehicle trajectory curve and the predicted trajectory curve as measurement indicators. The prediction result evaluation is shown in Table 3.

[0074] Table 3 Evaluation of prediction results

[0075]

[0076] The influence of driver characteristics on car-following behavior is analyzed by changing the reaction delay and the representation of driver characteristics. The attributes of vehicle No. 744 in the NGSIM dataset are assigned to vehicle No. 684, and its behavior characteristics are d 684 =[0.30,0.21] is changed to d 744 = [0.54, 0], the reaction delay changes from 1.7s to 1.4s. Assign the vehicle attribute numbered 684 to vehicle 2695, and its behavior characteristics d 2695 =[0.44,0.36] is changed to d 684 = [0.54, 0], the reaction delay changes from 1.8s to 1.7s. Assign the vehicle attribute numbered 2695 to vehicle 744, and its behavior characteristics d 744 =[0.54,0] is changed to d 2695 =[0.44,0.36], the reaction delay changes from 1.4s to 1.8s, other information remains unchanged, and multi-step prediction is performed again.

[0077] Evaluation of the car-following behavior prediction method considering driver characteristics of the present invention:

[0078] In the single-step prediction task, the prediction model is significantly better than the baseline model in all aspects. In terms of multi-step prediction, from the perspective of the upper quartile, whether it is the "absolute" accuracy metric EUC of the position, the reliability metric FRE, or the accuracy metric QE that is insensitive to the length of the spatiotemporal trajectory, the model is slightly better than the traditional baseline model. This shows that the model is better than the traditional car-following model in terms of model stability in multi-step prediction, and can better predict uniform motion.

[0079] Analyze the impact of driver characteristics on car-following behavior, and change the multi-step prediction effect after input. Figure 5 . The selected vehicles are all in the test set, so the model does not know the relationship between its vehicle attributes and following behavior during the training process. The MAE of position prediction for vehicles 684, 744, and 2695 changed from 9.42 to 10.87, 19.33 to 45.13, and 7.34 to 7.76, respectively. Changing the vehicle attribute input affects the accuracy of the model's multi-step prediction. Therefore, vehicle attribute information is of great significance in the model. Intuitively, changing vehicle attributes will lead to greater volatility in speed prediction values ​​and weaker predictive ability for speed change trends. The excellent performance of this model is mainly because the model differentiates different drivers and more accurately estimates the driver's following behavior.

Claims

1. A data-driven vehicle following behavior prediction method based on driver characteristics, characterized in that: The following steps are involved: Step 1: According to the actual trajectory data of the current vehicle and the preceding vehicle in the microscopic traffic flow, the mutation points of the speed changes of the two vehicles over time are obtained respectively, and the instantaneous reaction delay of the current vehicle is estimated according to the speed mutation points; Step 2: divide the driving behavior according to the intrinsic characteristics of the current vehicle multidimensional time series, cluster the divided driving behaviors, obtain three driving modes, calculate the proportion of each driving mode, and use the proportion of each driving mode to represent the behavioral characteristics of the current driver; Step 3, using the gated recurrent unit GRU as the prediction model. The input of the prediction model includes the multidimensional time series related to the current vehicle's following behavior and the current driver's behavior characteristics. The prediction target is the forward distance of the current vehicle in the next period. The length of the multidimensional time series is determined by the instantaneous reaction delay. Multidimensional time series are represented as Indicates that the time series data contains information of T moments, each moment has three dimensions, including the speed of the vehicle at each moment ν i (t), headway distance s i (t), and the speed difference Δν with the preceding vehicle i (t); The step 2 specifically includes: Step 201, assuming that the characterization vector of each moment of the multidimensional time series obeys a multidimensional Gaussian distribution, a greedy Gaussian partitioning algorithm is used to obtain the optimal partitioning point sequence of the multidimensional time series, and the parameter mean of each subsequence after the multidimensional time series is partitioned; Step 202, using the parameter mean to characterize each time interval, clustering the time intervals using the K-means method, and obtaining three categories by clustering, including three driving modes: normal, cautious, and aggressive, which are represented by numbers 0, 1, and 2, respectively; Step 203, calculate the time proportion of each driving mode respectively, expressed as The step 3 specifically includes: starting from time t, vehicle i inputs the multidimensional time series of the GRU model Indicates that v i (t) represents the speed of vehicle i at time t; s i (t) represents the headway distance of vehicle i at time t; Δv i (t) = v i (t)-v i-1 (t) represents the speed difference between vehicle i and the preceding vehicle at time t; Δx i-1 (t-1) is the change in the position of the preceding vehicle, indicating the distance the preceding vehicle has traveled from time t-1 to time t; dv i-1 (t) = v i-1 (t)-v i-1 (t-1) is the speed change of the preceding vehicle, indicating the speed difference between the preceding vehicle at time t and time t-1; τ is the instantaneous reaction delay of the driver; Another input of the GRU model is the time proportion of each driving mode d i , d i It is fused with the last hidden layer of the GRU model to output the forward distance of the current vehicle in the next period.

2. The prediction method according to claim 1, characterized in that: The mutation points of speed change over time include acceleration mutation points and deceleration mutation points. After obtaining the mutation points of the speed change over time of the two vehicles, the relationship between the speed mutation points of the two vehicles is matched, and the time difference between the two when the speed mutation of the leading vehicle causes the speed mutation of the following vehicle after matching is calculated, which is the instantaneous reaction delay.

Images