Method for generating background traffic flow considering driver dynamic reaction time and application thereof

Abstract

The application belongs to the technical field of automatic driving simulation test, and relates to a background traffic flow generation method considering dynamic reaction time of a driver and application. The dynamic reaction time of the driver is extracted based on a natural driving trajectory, a historical driving state sequence of the vehicle is taken as model input, a real reaction time of the driver is taken as model output, a dynamic reaction time prediction model is constructed based on a time convolution network, the dynamic reaction time prediction model and a physical following model are integrated to construct a reaction perception following model, the dynamic reaction time prediction model predicts dynamic reaction time t ( t ),the physical following model calculates the acceleration at the current moment according to the state at the moment of t ‑ t ( t ), and a differential evolution optimizer is used to calibrate the parameters of the reaction perception following model. The TCN is responsible for simulating the complex and difficult-to-describe driver perception and cognition process with explicit rules, and the IDM is responsible for describing the clear physical execution process, so that the final model has strong fitting real data capability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic driving simulation test, and particularly relates to a background traffic flow generation method considering dynamic reaction time of drivers and application. BACKGROUND

[0002] With the rapid development of automatic driving technology, constructing simulation test scenes has become a key link to verify and ensure its safety. Since automatic driving vehicles and manually driving vehicles will coexist as a new normal of road traffic, in simulation tests, it is crucial to generate micro background traffic flow with individual heterogeneity and randomness, and the background traffic flow that truly reflects human driving behavior is directly related to the effectiveness and challenge of the test environment. At present, the mainstream method of generating background traffic flow is to describe the longitudinal driving behavior of vehicles based on micro traffic flow models, such as car following models. The existing technology has considered the influence of reaction time to some extent when constructing the car following model, but still has certain limitations:

[0003] First, the reaction time model is relatively simple and lacks dynamic adaptability. Most existing car following models generally simplify the reaction time of drivers too much, such as setting it as a fixed constant or a value determined by statistical analysis. However, the reaction time of human drivers will dynamically change due to various factors such as current driving situation, mental state, and traffic density.

[0004] Second, the extracted dynamic reaction time data cannot be directly used to generate behaviorally realistic traffic flow, and there is a lack of specific solutions to effectively integrate the extracted data into simulation models.

[0005] Third, most models are difficult to balance accuracy and interpretability. Existing data-driven micro traffic flow models have a large number of adjustable parameters, making the model have high degrees of freedom and the ability to depict any complex driving behavior. However, the prediction effect of such models is highly related to the quality and quantity of the original trajectory data, and the interpretability of the model is generally limited.

[0006] In summary, the existing technology fails to provide a reaction mechanism that allows each driver in the simulation world to have a human-like, contextually dynamic reaction. This leads to the generated background traffic flow being relatively monotonous in behavior, and unable to reproduce complex traffic phenomena (such as the formation and dissipation of congestion waves) caused by the reaction delay changes of the human driver group, thereby reducing the severity and credibility of automatic driving simulation tests. SUMMARY

[0007] To solve the above technical problems, the application provides a background traffic flow generation method considering driver reaction time, which core idea is to create a reaction perception car following model with dynamic reaction ability through deep fusion of a data-driven dynamic reaction time prediction model and a physical car following model based on physical rules, and then generate a high-fidelity background traffic flow.

[0008] The application discloses a background traffic flow generation method considering driver dynamic reaction time, including the following processes:

[0009] Step S1, based on natural driving trajectory, extracting driver's dynamic reaction time, and constructing dynamic label data set containing driving situation sequence-reaction time pair;

[0010] Step S2, taking the historical driving state sequence of a vehicle in a period of time as the model input, and taking the real reaction time of the driver as the model output, a dynamic reaction time prediction model is constructed based on a time convolution network;

[0011] Step S3, integrating the physical car following model and the dynamic reaction time prediction model to construct a reaction perception car following model, the dynamic reaction time prediction model predicts the dynamic reaction time τ(t), and the physical car following model calculates the acceleration at the current time t according to the state at the time t-τ(t); a differential evolution optimizer is used to calibrate the parameters of the reaction perception car following model;

[0012] Step S4, in a micro-traffic simulation platform, instantiating the reaction perception car following model as a plurality of independent driver agents, deploying and running the driver agents in a virtual road network, and making the driver agents interact and evolve to generate a background traffic flow.

[0013] Preferably, the step S1 includes the following processes:

[0014] Step S101, regarding the front vehicle acceleration sequence as a stimulus signal, and regarding the rear vehicle acceleration sequence as a response signal, the dynamic reaction time τ is the time delay between the occurrence of the stimulus and the generation of the response;

[0015] Step S102, defining a fixed-size analysis window on a car following trajectory, the analysis window moves forward along the entire time sequence at a fixed step, and for each position of the analysis window, the front vehicle acceleration sequence and the rear vehicle acceleration sequence in the analysis window are extracted;

[0016] Step S103, using a cross-correlation analysis method to determine the cross-correlation coefficient of the front vehicle acceleration sequence and the rear vehicle acceleration sequence;

[0017] Step S104, within a preset reasonable time range [α, β], the real reaction time τ* of the current analysis window driving situation is calculated according to the following formula:

[0018] ,

[0019] the independent variable when the objective function takes the maximum value ; the cross-correlation coefficient of the front vehicle and the rear vehicle acceleration sequence;

[0020] Step S105, traverse the analysis window through all the car-following trajectories to obtain the real reaction time with timestamp and its corresponding cross-correlation coefficient, and construct a dynamic label data set containing driving situation sequence-reaction time pair.

[0021] Preferably, in the step S103, the cross-correlation function of the front vehicle acceleration sequence and the rear vehicle acceleration sequence is calculated , the formula is as follows:

[0022] ,

[0023] the expected function; the acceleration of the front vehicle at time t; the acceleration of the rear vehicle at time t+τ;

[0024] The cross-correlation function is normalized to the cross-correlation coefficient:

[0025] ,

[0026] , the mean of the front vehicle and the rear vehicle in the corresponding time sequence, respectively; , the standard deviation of the front vehicle and the rear vehicle in the corresponding time sequence, respectively.

[0027] Preferably, in the step S2, the core of the dynamic reaction time prediction model is a stacked residual block, and each residual block contains two layers of dilated causal convolution, which operates on one-dimensional sequence as follows:

[0028] ,

[0029] the convolution result of the elements in the sequence ; the input sequence; the filter; the dilation factor; the filter size.

[0030] Preferably, in the step S2, the driving state features of the vehicle include headway, relative speed, rear vehicle speed and front vehicle speed.

[0031] Preferably, in the step S3, the acceleration of the physical car-following model is calculated as follows:

[0032]

[0033] is the maximum acceleration of the rear vehicle; is the relative speed, is the speed of the front vehicle, is the speed of the rear vehicle; is the distance between the rear vehicle and the front vehicle; is the desired dynamic headway; is the desired speed;

[0034]

[0035] is the minimum safe stopping distance; is the comfortable deceleration; is the desired headway.

[0036] Preferably, in the step S3, the objective function of the differential evolution optimizer is calculated as follows:

[0037]

[0038] is the independent variable when the objective function takes the minimum value ; is the physical parameter set of the physical car-following model; is the optimal physical parameter set of the physical car-following model; is the simulated acceleration at the moment when the output of the reactive perception car-following model; is the real trajectory acceleration at the moment ; is the total number of time series steps.

[0039] Correspondingly, the present application also discloses a computer device / system, which comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to realize the steps of the above method.

[0040] Correspondingly, the present application also discloses a computer readable storage medium, which stores a computer program / instruction, wherein the computer program / instruction is executed by a processor to realize the steps of the above method.

[0041] ​​​​​Correspondingly, the application also discloses a computer program product, comprising computer programs / instructions, which realize the steps of the above method when executed by a processor.

[0042] The application has the following beneficial effects:

[0043] 1. The application can predict reaction time in real time according to a vehicle historical driving state sequence by constructing a dynamic reaction time prediction model TCN based on deep learning, realizes dynamic changes of reaction time of a simulated driver in a driving situation, and makes the behavior of a simulated vehicle adaptively like a human driver, greatly improving the realism and credibility of background traffic flow at a micro-behavior level.

[0044] 2. The application successfully integrates a dynamic reaction time prediction model driven by data and a physical following model based on physical rules by deep fusion, and unifies calibration by using a global optimization algorithm, seamlessly integrates the dynamic reaction mechanism learned from data into a micro driving behavior model (a physical following model), and makes traffic flow generated by the method endogenously emerge more complex traffic phenomena (such as congestion waves that walk and stop), thereby providing a more rigorous and more realistic test environment for an automatic driving system, enhancing the challenge of a simulation environment, and helping to find potential safety risks.

[0045] 3. The application adopts a hybrid architecture of a data-driven model (TCN) + a physical model (IDM), realizes the complementary advantages of a "black box" model and a "white box" model, TCN is responsible for simulating a complex driver perception and cognition process that is difficult to describe by explicit rules, and IDM is responsible for describing a clear physical execution process, so that the final model has strong ability to fit real data, and retains the actual significance of physical parameters (such as expected speed and expected vehicle headway), realizes the coexistence of model interpretability and performance. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of a background traffic flow generation method of the application is shown in the figure;

[0047] Figure 2 A structure diagram of a reaction perception following model of the application is shown in the figure;

[0048] Figure 3 A structure diagram of a dynamic reaction time prediction model based on a time convolution network of the application is shown in the figure;

[0049] Figure 4 A schematic diagram of driver reaction time label extraction in embodiment 1 of the application is shown in the figure;

[0050] Figure 5A comparison chart of simulation results of the reaction-aware car-following model and the physical car-following model in Embodiment 1 of the present application is shown in the figure; DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. If not specifically indicated, the technical means used in the embodiments are conventional means familiar to those skilled in the art.

[0052] As shown in the figure, Figure 1 The present application discloses a background traffic flow generation method considering dynamic reaction time of a driver, including the following processes:

[0053] Step S1, based on a natural driving trajectory (real driving trajectory), the dynamic reaction time of the driver is extracted, and a dynamic label data set containing a driving situation sequence-reaction time pair is constructed as a training label for a subsequent dynamic reaction time prediction model.

[0054] Step S2, taking the historical driving state sequence of a vehicle in a period of time as the model input and the real reaction time of the driver as the model output, a dynamic reaction time prediction model TCN is constructed and trained based on a time convolution network. The trained dynamic reaction time prediction model can be used as a function module to predict the corresponding dynamic reaction time according to the real-time driving situation sequence.

[0055] Step S3, a classic physical car-following model (intelligent driver model IDM) is integrated with the dynamic reaction time prediction model TCN to construct a reaction-aware car-following model RA-IDM. In the reaction-aware car-following model RA-IDM, the dynamic reaction time prediction model predicts the dynamic reaction time τ(t) at each time step t, and the physical car-following model uses the vehicle state delayed by τ(t), i.e. the state at time t-τ(t), to calculate the acceleration at the current time t, which simulates the delay process from perception to action of a human driver;

[0056] A differential evolution DE optimizer is used to calibrate the parameters of the reaction-aware car-following model, and a differential evolution algorithm, a global optimization algorithm, is used to uniformly calibrate the physical parameters of the physical car-following model. The DE algorithm is a heuristic search algorithm based on groups, which finds the optimal solution by simulating mutation, crossover and selection operations in the biological evolution process. The calibration process is a closed-loop optimization, and the goal is to find a set of physical car-following model physical parameters that can minimize the root mean square difference between the simulation acceleration of the reaction-aware car-following model RA-IDM and the real trajectory acceleration.

[0057] Step S4: In the microscopic traffic simulation platform, the calibrated reaction perception car-following model (i.e., the complete model containing the dynamic reaction time prediction module and the calibrated physical car-following model parameters) is instantiated into multiple independent driver agents. The driver agents are deployed and run in the virtual road network, allowing them to interact and evolve. Since each agent has a built-in reaction mechanism that can dynamically adapt to the situation, they work together to generate macroscopic background traffic flow with highly realistic behavior.

[0058] Specifically, such as Figure 2 As shown, the background traffic flow generation method of the present invention is carried out in parallel by two major modules. Figure 2 On the left is the dynamic reaction time prediction module. This module is entirely driven by real data. Its main task is to train a perceptual brain that can predict the driver's dynamic reaction time based on real-time driving scenarios from large-scale natural driving trajectory data through a temporal convolutional network. Figure 2 On the right is the Reaction-Aware Carry-Driven Model (RA-IDM) and its calibration module, which integrates the Dynamic Reaction Time and Physical Carry-Driven Model (IDM). It combines the trained sensory brain (left side) with a classic, physically meaningful executive limb to form a hybrid RA-IDM model. Then, a global optimization algorithm is used to uniformly calibrate the RA-IDM model, ensuring that both work in harmony.

[0059] In a further implementation, step S1, constructing a dynamic label dataset, is based on the stimulus-response relationship of car-following behavior. A sliding window approach is used to continuously analyze the complete car-following process, thereby capturing the time-varying characteristics of dynamic reaction time. Specifically, the process includes the following:

[0060] Step S101: Establish stimulus-response relationship: regard the acceleration sequence of the vehicle in front as the stimulus signal, the acceleration sequence of the vehicle behind as the response signal, and the dynamic reaction time τ as the time delay between the occurrence of stimulus and the generation of response;

[0061] Step S102: Apply a sliding window for continuous analysis: Define a fixed-size analysis window on a complete car-following trajectory. The analysis window moves forward along the entire time series with a fixed step size. For each position of the analysis window, extract the acceleration sequences of the preceding vehicle and the following vehicle within the analysis window.

[0062] Step S103: Calculate the time delay within the analysis window: Within each analysis window, in order to automatically calculate the dynamic reaction time τ, a cross-correlation analysis method is used to calculate the cross-correlation function of the acceleration sequences of the preceding and following vehicles. The formula is as follows:

[0063] ,

[0064] is the desired function; is the acceleration of the front vehicle at time t; is the acceleration of the rear vehicle at time t+τ;

[0065] To eliminate the dimension effect, the cross-correlation function is normalized to the cross-correlation coefficient:

[0066] ,

[0067] , is the mean of the front vehicle, respectively, the rear vehicle, in the corresponding time series; , is the standard deviation of the front vehicle, respectively, the rear vehicle, in the corresponding time series;

[0068] Step S104, determine and mark the reaction time: within the preset reasonable time range [α, β], find the time delay when the cross-correlation coefficient reaches the maximum value, at this time, the time delay is regarded as the real reaction time label under the driving situation represented by the current analysis window, the time stamp of the label can be recorded as the center time of the analysis window; the real reaction time τ* is calculated as follows:

[0069] ,

[0070] is the independent variable when the objective function takes the maximum value ; is the cross-correlation coefficient of the acceleration series of the front vehicle and the rear vehicle;

[0071] Step S105, construct a dynamic label data set: traverse the analysis window to all the car-following trajectories meeting the conditions to obtain a series of real reaction times with time stamps and their corresponding cross-correlation coefficients, filter out high-confidence results, and construct a dynamic label data set containing driving situation sequence-reaction time pairs for subsequent model training.

[0072] Further embodiments, the structure of the dynamic reaction time prediction model based on the time convolution network is as Figure 3As shown, the reaction time can be predicted according to the real-time driving situation. The dynamic reaction time prediction model takes the historical driving state sequence of the vehicle in a period of time as input, and the state features include the inter-vehicle distance (i.e. the distance between the front of the rear vehicle and the rear of the front vehicle), the relative speed, the rear vehicle speed and the front vehicle speed. The core of the dynamic reaction time prediction model is a stacked residual block, each residual block contains two layers of dilated causal convolution. The causal convolution ensures that only the historical input at time t and before t is used when outputting at time t, and no future information is seen. The dilated convolution increases the receptive field exponentially through the dilated rate, so that the network can efficiently capture long-range dependencies in time series without deep layers. Its operation on one-dimensional sequence is as follows:

[0073] ,

[0074] is the convolution result of the elements in the sequence ; is the input sequence; is the filter; is the dilation factor; is the filter size.

[0075] In a further embodiment, the acceleration of the physical car following model is calculated as follows:

[0076] ,

[0077] is the maximum acceleration of the rear vehicle; is the relative speed, , is the front vehicle speed, is the rear vehicle speed; is the distance between the rear vehicle and the front vehicle; is the desired dynamic headway of the rear vehicle; is the desired speed of the rear vehicle;

[0078] ,

[0079] is the minimum safe stopping distance of the rear vehicle; is the comfortable deceleration of the rear vehicle; is the desired headway of the rear vehicle.

[0080] In a further embodiment, the objective function of the differential evolution optimizer is calculated as follows:

[0081] ,

[0082] ​for the independent variable when the objective function takes the minimum value ; for the physical parameter set of the physical car-following model for the optimal physical parameter set of the physical car-following model for the output of the reaction-aware car-following model the simulation acceleration at the moment for the real trajectory acceleration at the moment for the total number of time series steps

[0083] The specific calibration process of the differential evolution optimizer is as follows:

[0084] Input: population size M; dimension D of solution; maximum number of iterations P.

[0085] Output: optimal solution vector Δ.

[0086] (1) Initialization: randomly generate an initial population containing M individuals (candidate solutions) within the preset parameter boundary; each individual is a D=5-dimensional vector corresponding to a set of IDM parameters .

[0087] (2) Define fitness function: the fitness function is used to evaluate the pros and cons of each individual (parameter set), and in the present application, the fitness function is the objective function , and the calculation process is as follows:

[0088] a. input an IDM parameter set;

[0089] b. In the simulation verification module, use the parameter set and the trained dynamic reaction time prediction model to perform trajectory simulation; for a real car-following data, at each simulation time step :

[0090] c. Extract the historical driving state sequence before the current moment, input the dynamic reaction time prediction model TCN, and obtain the dynamic reaction time ;

[0091] d. The physical car-following model IDM uses the vehicle state after delay, i.e. the speed , relative speed , and headway at the moment , to calculate the simulation acceleration at the current moment;

[0092] e. After the simulation is completed, calculate the simulation acceleration at all time steps and the real trajectory acceleration The root mean square error (RMSE) between them;

[0093] f. Return the RMSE as the fitness value for this parameter group. The smaller the RMSE, the better the fitness.

[0094] (3) Iterative optimization: Enter the main loop and perform at most P iterations. In each iteration, perform the following operations on each individual in the population:

[0095] a. Mutation: A mutation vector is generated by differentially combining other randomly selected individuals in the population;

[0096] b. Crossover: Cross the mutation vector with the current individual to generate a test vector;

[0097] c. Selection: Calculate the fitness value of the test vector, i.e. call the fitness function in step (2); if the fitness of the test vector is better than the fitness of the current individual (i.e., the RMSE is smaller), then replace the current individual with the test vector in the next generation population; otherwise, retain the current individual.

[0098] (4) Termination and Output: The algorithm terminates when the maximum number of iterations is reached or the error meets the convergence condition. The best-fit individual found so far is returned, i.e., the optimal solution vector Δ, as the final calibrated IDM physical parameters.

[0099] The method of this invention utilizes the powerful temporal feature extraction capability of a dynamic reaction time prediction model to learn and predict the driver's reaction time, which dynamically changes with the driving situation, from the vehicle's historical state sequence, thus replacing the static or simple random reaction time in the prior art; the dynamic reaction time predicted by the dynamic reaction time prediction model... As input, it is applied in real time to the decision-making calculation process of the physical car-following model, that is, the physical car-following model is based on Traffic conditions at any given time determine The acceleration at any moment forms a reaction-perception car-following model that tightly couples perception and execution. By combining a dynamic reaction time prediction model for predicting the driver's dynamic reaction time with a traditional microscopic physical car-following model for describing the vehicle's physical motion, separate modeling and coupling of the perception and cognition and physical execution stages in driving behavior are achieved. A global optimization algorithm is used to uniformly calibrate the parameters of the physical car-following model with the goal of minimizing the error between the simulated trajectory and the real trajectory. This ensures that the data-driven module and the physical model module can work together optimally, making the output of the entire reaction-perception car-following model highly consistent with the real world.

[0100] Example 1

[0101] Taking the processing of the disclosed NGSIM vehicle trajectory dataset as an example, the implementation process of a background traffic flow generation method considering the dynamic reaction time of drivers is elaborated in detail.

[0102] 1. Data preparation and feature engineering

[0103] Load the original vehicle trajectory data file of NGSIM I-80 highway. Preprocess the data, including unit conversion, data cleaning, and select "front-rear" vehicle pairs with a following time of more than 30 seconds. For each vehicle pair at each time point (data frame rate 10 Hz, i.e. 0.1 s / frame), calculate key car-following features, including: headway, relative speed, own vehicle speed, front vehicle speed, and own vehicle acceleration. Divide the dataset into a training set (used for dynamic reaction time prediction model TCN training and DE parameter calibration) and a test set (used for final model performance verification).

[0104] 2. Training of dynamic reaction time prediction model

[0105] Using the sliding window and cross-correlation analysis method, extract the "driving situation-reaction time" label dataset from the trajectory data of the training set. An example of driver reaction time label extraction is shown in Figure 4 . Then, build and train the physical car-following model TCN. The model input is the driving situation sequence of the past 2.5 seconds (25 time steps), and the output is the predicted reaction time. After 20 cycles of training, a TCN model is obtained that can predict the driver's reaction time according to the real-time driving situation.

[0106] 3. Unified calibration of RA-IDM physical parameters

[0107] Calibrate the 5 physical parameters of the IDM model using the differential evolution (DE) algorithm. The population size of the DE algorithm is set to 15, and the maximum number of iterations is 30 generations. The objective function is to minimize the root mean square error (RMSE) between the predicted acceleration of RA-IDM and the true acceleration on the training set.

[0108] After optimization, a set of IDM physical parameters that best match the TCN dynamic reaction module is obtained, and the calibration results are:

[0109] Expected speed = 19.38 m / s

[0110] Expected headway T = 0.53 s

[0111] Maximum acceleration = 0.51 m / s²

[0112] Comfort deceleration b = 3.24 m / s²

[0113] Minimum safety gap = 1.00 m

[0114] So far, a complete reaction-aware car-following model (RA-IDM) is constructed, which is composed of the trained TCN model and the calibrated set of IDM parameters above.

[0115] 4. Model performance verification and result analysis

[0116] To verify the performance of the RA-IDM model proposed in the present application, trajectory simulation experiments are performed on the test set that has never been used for training and calibration, and compared with the traditional IDM model using fixed reaction time. The parameters of the comparative IDM model can use the published calibration values in the field. (Literature: Kesting A, Treiber M, Helbing D. Enhanced intelligent driver model to access the impact of driving strategies on traffic capacity [J]. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2010, 368(1928): 4585-605.)

[0117] Statistical evaluation is performed on the entire test set, and the average RMSE of the trajectory prediction of the RA-IDM model of the present application is 4.12 m, and the average RMSE of the traditional IDM model is 4.65 m, which is reduced by 11.4%, which proves the significant beneficial effects of the present application in improving the simulation reality and prediction accuracy.

[0118] A typical car-following event in the test set is selected for detailed analysis, and the time sequence comparison chart of each motion parameter is drawn, as shown in Figure 5 .

[0119] From Figure 5The comparison results can be seen: in the trajectory reproduction accuracy, the RA-IDM model of the application is significantly better than the traditional IDM model. The root mean square error (RMSE) of RA-IDM is marked in the figure as 2.43 meters, lower than 3.17 meters of the traditional IDM, indicating that the fitting degree of the simulated vehicle displacement trajectory and the true observed value is higher. In the speed dynamic following, RA-IDM can more accurately reproduce the observed speed dynamic change, especially in the deceleration process near 660 seconds and the acceleration process near 690 seconds, its following effect is obviously better than that of traditional IDM. In the maintenance of headway, the traditional IDM simulated spacing after 655 seconds deviates from the true value, while the RA-IDM simulated headway curve has smaller difference with the observed value. In the acceleration response, although both models smooth the true acceleration with large noise, the acceleration and deceleration trend of RA-IDM is more consistent with the observed data, indicating that its response to traffic flow change is more sensitive and accurate.

[0120] 5. Generation of high-fidelity background traffic flow

[0121] In the traffic simulation software SUMO, the above calibrated RA-IDM model is packaged as a driver agent class. Deploy the agent in a virtual road network, and generate a large number of vehicles according to the traffic demand. After starting the simulation, each agent will make driving decisions according to its built-in RA-IDM model, interact with each other, and form a macroscopic, dynamically evolving, and highly realistic behavior background traffic flow. This traffic flow can be used for various application scenarios such as algorithm testing and safety verification of autonomous vehicles.

[0122] The above-described embodiments are only to describe the preferred modes of the present application, and not to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications, variations, modifications, and replacements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for generating background traffic flow taking into account the dynamic reaction time of the driver, characterized in that: The method comprises the following steps: Step S1, extracting the dynamic reaction time of the driver based on the natural driving trajectory, and constructing a dynamic label data set containing driving situation sequence-reaction time pairs; Step S2, taking the historical driving state sequence of the vehicle in a period of time as the model input, and taking the real reaction time of the driver as the model output, a dynamic reaction time prediction model is constructed based on a time convolution network; In the step S2, the core of the dynamic reaction time prediction model is a stacked residual block, each residual block contains two layers of dilated causal convolution, and the operation on the one-dimensional sequence is as follows: ， is a convolution result of elements in a sequence ; is an input sequence; is a filter; is a dilation factor; is a filter size; Step S3, integrating the physical car following model and the dynamic reaction time prediction model to construct a reaction perception car following model, the dynamic reaction time prediction model predicts the dynamic reaction time τ(t), and the physical car following model calculates the acceleration at the current time t according to the state at the time t-τ(t); a differential evolution optimizer is used to calibrate the parameters of the reaction perception car following model; In the step S3, the objective function of the differential evolution optimizer is calculated as follows: ， independent variable for the target function to take the minimum value ; set of physical parameters for the physical car-following model set of optimal physical parameters for the physical car-following model output of the reactive perception car-following model simulated acceleration at time instant is the real trajectory acceleration at time instant is the real trajectory acceleration at time instant is the total number of time steps Step S4, in the micro-traffic simulation platform, the reaction perception car following model is instantiated as a plurality of independent driver agents, and the driver agents are deployed and run in the virtual road network to interact and evolve, thereby generating a background traffic flow.

2. The method for generating background traffic flow considering driver dynamic reaction time according to claim 1, characterized in that: The step S1 comprises the following steps: Step S101, taking the front vehicle acceleration sequence as the stimulus signal and the rear vehicle acceleration sequence as the response signal, and the dynamic reaction time τ is the time delay between the occurrence of the stimulus and the generation of the response; Step S102, defining a fixed-size analysis window on a car following trajectory, and moving the analysis window along the entire time sequence at a fixed step, for each position of the analysis window, extracting the front vehicle acceleration sequence and the rear vehicle acceleration sequence in the analysis window; Step S103, determining the cross-correlation coefficient of the front vehicle acceleration sequence and the rear vehicle acceleration sequence by using cross-correlation analysis method; Step S104, within a preset reasonable time range [α,β], the real reaction time τ* of the current analysis window under the driving situation is calculated as follows: ， the argument of the target function for which the maximum is sought the cross-correlation coefficient of the acceleration sequences of the preceding and following vehicles;​ Step S105, traversing the analysis window on all car following trajectories to obtain the real reaction time with timestamp and the corresponding cross-correlation coefficient, and constructing a dynamic label data set containing driving situation sequence-reaction time pairs.

3. The method for generating background traffic flow considering driver dynamic reaction time according to claim 2, characterized in that: In the step S103, the cross-correlation function of the front vehicle acceleration sequence and the rear vehicle acceleration sequence is calculated The formula is as follows: ， is the desired function; is the acceleration of the preceding vehicle at time t; is the acceleration of the following vehicle at time t + τ; The cross-correlation function is normalized to a cross-correlation coefficient: ， , are the mean values of the front and rear vehicles in the respective time series; , are the standard deviations of the front and rear vehicles in the respective time series.

4. The method for generating background traffic flow considering driver dynamic reaction time according to claim 1, characterized in that: In the step S2, the driving state features of the vehicle include inter-vehicle distance, relative speed, rear vehicle speed and front vehicle speed.

5. The method for generating background traffic flow considering driver dynamic reaction time according to claim 3, characterized in that: In the step S3, the acceleration of the physical follow-up model is calculated as follows: The calculation formula is as follows: ， is the maximum acceleration of the rear vehicle; is the relative speed, , is the front vehicle speed, is the rear vehicle speed; is the distance between the rear vehicle and the front vehicle; is the desired dynamic headway; is the desired speed; ， for minimum safe stopping distance; for comfortable deceleration; for desired headway.

6. A computer apparatus / device / system, characterized by: The computer program product comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that: The computer program product comprises a computer program / instruction, which is executed by a processor to implement the steps of the method according to any one of claims 1-5.

8. A computer program product, characterised in that: The computer program product comprises a computer program / instruction, which is executed by a processor to implement the steps of the method according to any one of claims 1-5.

Images