Method for accelerating iteration based on automatic driving edge scene parameters

Abstract

The invention relates to the field of scene optimization of edge calculation, in particular to a method for accelerating iteration based on automatic driving edge scene parameters. The method comprises the following steps: firstly, extracting and analyzing speed and position data of a vehicle driving into a ramp according to an NGSIM vehicle track data set; secondly, sensitivity analysis under Monte Carlo simulation is carried out on the extracted speed and position parameters of the vehicle; finally, the sensitivity parameters are added into the iteration process of the particle swarm algorithm, and a final detection result is obtained. The feature information can be extracted more efficiently through clustering analysis of the vehicle position data, the sensitive parameters are added in the particle swarm optimization process, the optimization speed of the particle swarm optimization is increased, and therefore the iteration process of scene parameters is accelerated. According to the method, the reliability and applicability of scene modeling can be effectively improved, the accuracy of feature extraction is enhanced, the calculation efficiency of an algorithm is optimized, and reliable support is provided for rapid optimization of scene parameters.

Description

Technical Field

[0001] The present invention relates to the field of accelerating the iteration of autonomous driving scenarios, and particularly to a method for accelerating the iteration based on autonomous driving edge scenario parameters. Background Art

[0002] With the rapid development of computer technology, the Autonomous Driving System (ADS), as a typical representative of the cyber-physical fusion system, aims to reduce traffic accidents caused by human errors. By eliminating the dependence on drivers, ADS significantly reduces the accident rate. However, to ensure the safety of autonomous vehicles, a large number of edge scenario tests must be carried out. Edge scenarios are a set of scenarios in the logical scenario parameter space near the collision risk and safety boundary regions, characterized by low probability and high risk, and can accelerate the testing of the ability boundary of the autonomous driving system. However, edge scenarios are scarce and difficult to reproduce in real traffic scenarios. Therefore, optimization algorithms are used to accelerate the iteration of edge scenario parameters and generate corresponding edge scenarios in virtual simulation software to meet the safety verification requirements before the large-scale commercialization of autonomous vehicles.

[0003] The document with the application number "202410207982.5" discloses a "method for generating an accelerated test scenario library for an autonomous driving system". Based on the selection of initial state elements of time series scenarios, a scenario parameterization model is established for the initial state characterized by multi-element coupling. The problem is that the requirement for extracting elements of the scenario initial state is relatively high, resulting in difficulty in the state of the initial scenario. Another document with the application number "202211711315.8" discloses a "method for online generating key edge test scenarios for autonomous driving acceleration". By constructing a scenario complexity model and based on the generated natural driving test scenarios and the constructed scenario complexity model, an autonomous driving vehicle is evaluated and tested, and scenario adaptive adjustment is performed to output key boundary test scenarios. The problem is that it depends on the generalization of natural driving data, restricting the diversity and coverage of scenario generation.

[0004] Generally speaking, when evaluating and testing autonomous vehicles, the above documents do not consider the characteristics of too many scenario elements and too few edge scenarios, have limitations on the applicable scope of scenarios, are difficult to model relatively complex scenarios, and have a slow scenario iteration speed. Summary of the Invention

[0005] The present invention proposes a method for accelerating the iteration based on autonomous driving edge scenario parameters to solve the problems in the prior art, such as the small number of edge scenarios, limitations on the applicable scope of scenarios, difficulty in modeling relatively complex scenarios, and slow scenario iteration speed.

[0006] To achieve the above object, the technical solution of the present invention is as follows: A method for accelerating the iteration of edge scene parameters based on autonomous driving, comprising the following steps:

[0007] Step 1: First, vehicle parameters are extracted based on the NGSIM vehicle trajectory dataset. The vehicle parameters include vehicle speed, vehicle position, lane position, distance between vehicles, and data on whether lane changes occur. Then, the extracted vehicle position data is subjected to cluster analysis to divide the vehicles into background vehicles and the autonomous driving vehicle under test.

[0008] Step 2: Perform sensitivity analysis of the extracted parameters by Monte Carlo simulation, and calculate the sensitivity values of the background vehicle parameters and the autonomous driving vehicle under test.

[0009] Step 3: Initialize the population of the particle swarm algorithm, and simultaneously construct a multi-factor coupling objective function of the particle swarm algorithm. Determine the level of the scene through the fitness function. Finally, add the sensitivity values obtained in Step 2 to the iterative process of the particle swarm algorithm to perform accelerated iteration of the edge scene parameters.

[0010] Further, the sensitivity analysis in Step 2 above includes four steps:

[0011] First, perform sensitivity analysis between the background vehicle parameters and the autonomous driving vehicle under test. The given model function is the minimum time to collision of the vehicle:

[0012]

[0013] where d b-y represents the y-direction position of the background vehicle, and d c-y represents the y-direction position of the vehicle under test. d b-x represents the x-direction position of the background vehicle, and d c-x represents the x-direction position of the vehicle under test. v b represents the speed of the background vehicle, and v c represents the speed of the vehicle under test;

[0014] Then, set the parameter variable range and define the range of the input variables. Generate N = 1000 random samples within the set range, and then use the model function to calculate the output value of each sample;

[0015] Next, screen the valid samples to ensure that the output values are within the set range;

[0016] Finally, perform statistical analysis on the screened samples to obtain the sensitivity values of the background vehicle parameters and the autonomous driving vehicle under test.

[0017] Further, in Step 3 above, first perform initialization of the particle population by Latin square oversampling and construct a multi-factor coupling objective function:

[0018]

[0019] Among them, x and v respectively represent the position and velocity of the particle; the parameters in A include the velocity of the test vehicle, the distance coordinates in the x and y directions, the velocity of the lane-changing vehicle, and the coordinates in the x and y directions; ω1, ω2, and ω3 respectively represent the weights of the three factors.

[0020] Further, in the above step three, the weights of the objective function of the genetic algorithm are determined, and the objective function after determining the weights is used as the fitness function, and the value of the fitness function is used to determine the scenario level.

[0021] Further, in the above step three, during the particle iteration process, the adaptation iteration relationship function is expressed as follows:

[0022]

[0023] Among them, w max and w min respectively represent the maximum and minimum values of the inertia weight, iteration represents the current iteration number, and num iterations represents the total number of iterations.

[0024] Compared with the existing technology, the beneficial effects of the present invention are:

[0025] (1) Abstract logical modeling based on the NGSIM vehicle trajectory dataset and clustering analysis of vehicle data. The present invention first effectively extracts the speed and position data of the vehicles in the scenario and obtains the value range, and uses K-means clustering to divide the vehicle position data into background vehicles and the tested autonomous vehicles. This method can more accurately simulate the actual driving environment, ensure the true effectiveness of the parameters, and can focus more on the research of the vehicle interaction process. Thereby improving the reliability and applicability of the scenario modeling.

[0026] (2) The combination of sensitivity analysis and particle swarm algorithm. The present invention obtains the sensitivity analysis results between the parameters through Monte Carlo simulation of the extracted parameters, and constructs the objective function and fitness function of the multi-factor coupling of the particle swarm algorithm on this basis. By adding sensitive analysis parameters during the iteration process of the particle swarm algorithm, the particle swarm algorithm can achieve rapid convergence and precise optimization, and can quickly make the scenario approach the edge scenario.

[0027] (3) The present invention innovatively adopts a clustering analysis method for vehicle position data, which improves the efficiency of feature information extraction. On this basis, a sensitivity parameter is introduced to optimize the particle swarm algorithm, effectively improving the optimization performance of the algorithm, and then accelerating the iterative convergence speed of the scenario parameters. This improved method not only enhances the accuracy of feature extraction but also optimizes the computational efficiency of the algorithm, providing reliable support for the rapid optimization of scenario parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is the overall flowchart of the present invention;

[0029] Figure 2 is the comparison diagram of the convergence result after adding sensitivity analysis and the convergence result without adding sensitivity analysis. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the appended embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, which are only used to illustrate the present invention but not to limit the scope of the present invention.

[0031] As Figure 1 shown in the overall flowchart of the present invention, the basic idea of the present invention is to first perform an abstract logical modeling of the scenario based on the natural driving dataset, extract and analyze the speed and position data of relevant vehicles therein, determine the model function, and perform a sensitivity analysis of the parameters. At the same time, the sensitivity analysis results are added to the update iteration of the particle swarm algorithm, and finally the edge scenario parameters are obtained.

[0032] Based on the above basic idea, the present invention provides a method for accelerating the iteration of edge scenario parameters for autonomous driving, and the specific implementation steps are as follows.

[0033] Step 1: First, construct an abstract logical scenario, extract the vehicle parameters in the dataset from the NGSIM vehicle trajectory dataset. The vehicle parameters include the value ranges of vehicle speed and vehicle position, lane position, distance between vehicles, and trajectory data such as whether the vehicle is a lane-changing vehicle, etc. Then, perform K-means clustering analysis on the vehicle parameters of all vehicles entering the ramp, classify the vehicles into background vehicles and the measured autonomous driving vehicles, analyze and select the distribution relationship between the parameters, and provide support for scenario iteration generation. The parameter analysis is expressed as:

[0034]

[0035] where J is the sum of the squared errors within the clusters. K is the number of clusters, n is the total number of data points, x j is the j-th data point in the data, and μ j is the clustering center of the i-th cluster.

[0036] Step 2: Calculate the sensitivity values of the background vehicle parameters and the autonomous driving vehicle under test by performing Monte Carlo sensitivity analysis on the parameters extracted from the modeling scenario. The specific steps include:

[0037] Step 2.1 Perform sensitivity analysis between parameters: Use Monte Carlo simulation to estimate the output distribution of the model through repeated random sampling to evaluate the response of the model to changes in the input variables.

[0038] Specifically, a sensitivity analysis is conducted between the background vehicle parameters and the parameters of the autonomous driving vehicle under test, and the minimum collision time of the vehicle given the model function is:

[0039]

[0040] Among them, d b-y Indicates the y-direction position of the background vehicle, d c-y Indicates the y-direction position of the vehicle being measured. b-x represents the x-direction position of the background vehicle, d c-x Indicates the x-direction position of the vehicle being measured. b represents the background vehicle speed, v c Indicates the speed of the vehicle being tested.

[0041] Step 2.2 sets the parameter variable range and defines the range of the input variable, generates N = 1000 random samples within the set range, and then uses the model function to calculate the output value of each sample. It is generally recognized that TTC is at the dangerous boundary when it is equal to 1.5s. In the present invention, N = 1000 random samples are generated within the range of [1,2] when TTC belongs to;

[0042] Step 2.3: Screen valid samples to ensure that the output value is within the set range;

[0043] Step 2.4 evaluates the impact of small changes in input variables on the model function through sensitivity analysis, and uses the finite difference method to calculate the sensitivity of each input variable. Introduce small changes to each input variable, calculate the changes in the model function, and perform statistical analysis on the screened samples based on the changes in the model function, and calculate the sensitivity values of the background vehicle parameters and the tested autonomous vehicle respectively.

[0044] The specific operations for calculating the sensitivity value are as follows:

[0045] Y base =model(v b ,d b-x ,d b-y ,v c ,d c-x ,d c-y )(3)

[0046]

[0047] Among them, X1, X2, ..., X i represent the speed and position of the vehicle.

[0048] Step 3: First, initialize the population of the particle swarm algorithm, and at the same time construct the objective function of the multi-factor coupling of the particle swarm algorithm. Determine the scenario level through the fitness function, and finally perform the accelerated iteration of the edge scenario parameters. The specific description is as follows:

[0049] The way to initialize the population in Step 3.1 is to use the Latin square oversampling to initialize the population particles. Suppose there are k parameters, and the range of each parameter is divided into n intervals. The range of the i-th parameter is [a i , b i . The range of each parameter [a i , b i is divided into n intervals. The specific operation is shown in the following equation:

[0050]

[0051] Among them, the j-th interval is [a i +(j - 1)Δx, a i +jΔx], where j = 1, 2, 3…, n. Randomly select a sample point in each interval [a i +(j - 1)Δx, a i +jΔx], as shown in the following formula:

[0052] x ij = [a i +(j - 1)Δx, a i +jΔx](6)

[0053] Among them, u ij is a random number uniformly distributed between [0, 1]. Then the sample points of each parameter are randomly arranged to generate a combination of sample points. For each parameter i, the generated combination of sample points is shown in the following formula:

[0054] x i = (x iπ(1) , x iπ(2) , ..., x iπ(n) )(7)

[0055] Among them, π represents a random sequence, and x represents a set of vehicle parameters extracted. The Latin square oversampling initialization method is used in the present invention.

[0056] Step 3.2: Construct a multi-factor coupling objective function based on the particle swarm algorithm. The design of the objective function is divided into three parts, as shown in the following formula:

[0057]

[0058] Among them, x and v represent the position and velocity of the particle respectively. In the present invention, the objective function comprehensively considers three parts. The first part is the minimum collision time of the vehicle; the second part is the lane-changing time; the third part is the variance of the velocity change, so as to comprehensively evaluate the performance of each group of particles. Among them, the parameters of the first part A are mainly the speed of the measured autonomous driving vehicle and the distance coordinates in the x and y directions. The speed of the background vehicle, as well as the coordinates in the x and y directions. ω1, ω2, and ω3 represent the weights of the three factors. The weights of the objective function are determined by the genetic algorithm. Through the powerful search ability of the genetic algorithm, the weights of the three parts in the particle swarm algorithm objective function are determined. Through its selection, crossover, and mutation processes, the optimal combination of the weights of the three parameters is continuously selected, and finally a set of optimal weights is obtained, and the objective function under the optimal weights is used as the fitness function.

[0059] Use the genetic algorithm to determine the weights of the objective function, use the objective function with the determined weights as the fitness function, and use the value of the fitness function to determine the scenario level.

[0060] Step 3.3: Add the sensitivity values of the background vehicle parameters and the measured autonomous driving vehicle parameters obtained in Step 2.2 to the iterative process of the particle swarm algorithm to obtain the detection result and perform accelerated iteration of the edge scenario parameters:

[0061] It is necessary to design an adaptive iteration relationship function. The following formula is the adaptive inertia weight:

[0062]

[0063] Among them, w max and w min represent the maximum and minimum values of the inertia weight respectively, iteration represents the current iteration number, and num iterations represents the total number of iterations.

[0064] As Figure 2 shown, after introducing sensitivity analysis in the present invention, the convergence speed of the particle swarm optimization algorithm in the initial stage is significantly faster than that of the algorithm without introducing sensitivity analysis, the number of iterations to reach stability is also less, and the final fitness value is lower, indicating better optimization effect.

[0065] As shown in Table 1, the number of edge scenarios obtained by the particle swarm optimization algorithm without sensitivity analysis is significantly less than that obtained by the particle swarm optimization algorithm with sensitivity analysis, indicating the advancement of the improvement of the present invention.

[0066] Table 1

[0067]

[0068] The above description is an illustration of the specific implementation of the present invention, rather than a limitation thereof. Those skilled in the relevant technical field can also make various equivalent technical solutions without departing from the scope of the present invention. Therefore, all equivalent technical solutions should be classified into the scope of the invention protection of the present invention.

Claims

1. A method for accelerating iteration based on autonomous driving edge scenario parameters, characterized in that: It includes the following steps: Step 1: First, vehicle parameters are extracted based on the NGSIM vehicle trajectory dataset. The vehicle parameters include vehicle speed, vehicle position, lane position, distance between vehicles, and data on whether lane changes occur. Then, the extracted vehicle position data is subjected to cluster analysis to divide the vehicles into background vehicles and the autonomous driving vehicle under test; Step 2: Conduct a sensitivity analysis of the extracted parameters through Monte Carlo simulation, and calculate the sensitivity values of the background vehicle parameters and the autonomous driving vehicle under test; Step 3: Initialize the population of the particle swarm algorithm, and simultaneously construct a multi-factor coupling objective function of the particle swarm algorithm. Determine the level of the scenario through the fitness function. Finally, add the sensitivity values obtained in Step 2 to the iterative process of the particle swarm algorithm to perform accelerated iteration of the edge scenario parameters.

2. The method for accelerating iteration based on autonomous driving edge scenario parameters according to claim 1, wherein: The sensitivity analysis in Step 2 includes four steps: First, conduct a sensitivity analysis between the background vehicle parameters and the autonomous driving vehicle under test. The given model function is the minimum time to collision of the vehicle: Among them, d b-y represents the y-direction position of the background vehicle, and d c-y represents the y-direction position of the vehicle under test. d b-x represents the x-direction position of the background vehicle, and d c-x represents the x-direction position of the vehicle under test. v b represents the speed of the background vehicle, and v c represents the speed of the vehicle under test; Then, set the parameter variable range and define the range of the input variables. Generate N = 1000 random samples within the set range, and then use the model function to calculate the output value of each sample; Next, screen the valid samples to ensure that the output values are within the set range; Finally, conduct a statistical analysis of the screened samples to obtain the sensitivity values of the background vehicle parameters and the autonomous driving vehicle under test.

3. A method for accelerating iterative optimization of parameters in an autonomous driving edge scenario according to claim 1, characterized in that: In Step 3, first initialize the particle population by Latin hypercube sampling and construct a multi-factor coupling objective function: Among them, x and v respectively represent the position and velocity of the particle; the parameters in A include the speed of the test vehicle, the distance coordinates in the x and y directions, the speed of the lane-changing vehicle, and the coordinates in the x and y directions; ω1, ω2, and ω3 respectively represent the weights of the three factors.

4. A method for accelerating iteration of parameters in an autonomous driving edge scenario according to claim 1, characterized in that: In Step 3, determine the weights of the objective function of the genetic algorithm, use the objective function with the determined weights as the fitness function, and determine the scenario level with the value of the fitness function.

5. A method for accelerating iteration based on parameters of an autonomous driving edge scenario according to claim 1, characterized in that: In Step 3, during the particle iteration process, the adaptive iteration relationship function is expressed as follows: where, w max and w min represent the maximum and minimum values of the inertia weight respectively, iteration represents the current iteration number, and num iterations represents the total number of iterations.

Images