CN111797000A - A Scenario Complexity Evaluation Method Based on Gradient Boosting Decision Tree Model - Google Patents

Abstract

The present invention provides a method for evaluating scene complexity based on a gradient boosting decision tree model, comprising the following steps: S1, collecting parameters, and generating a simulated driving scene; S2, scoring the simulated driving scene samples for complexity; S3, summarizing The feature elements of the complexity are input into the decision tree model for calculation; S4, the decision tree is upgraded; S5, the feature parameter data set that affects the model is obtained, 80% of the data set is used as the training set and 20% is used as the test set, using 50% off Cross-validation and debugging to obtain the complexity evaluation model; S6, calculate the scene complexity of the data to be evaluated; S7, split the input driving scene data into dynamic features and static features, and then comprehensively score according to each impact feature Get the scene complexity. The method of the present invention can provide a clear and concise estimation of the complexity of the automatic driving test scene, and meet the requirement that the tester can select the driving scene according to the complexity of the scene.

Description

A Scenario Complexity Evaluation Method Based on Gradient Boosting Decision Tree Model

technical field

The invention belongs to the field, and in particular relates to a scene complexity evaluation method based on a gradient boosting decision tree model.

Background technique

The classification of the existing simulation test driving scenes is mostly based on the tested functions. The driving scene categories obtained by using this type of classification method are relatively general, and the testers will not be able to know the relevant test scenes concisely and clearly. For the test difficulty generated by the automatic driving control algorithm test, it is impossible to achieve a progressive test. In the actual testing process, in addition to the different types of scenarios, the testing of scenarios with different difficulties is also an important factor that the testers need to consider. Therefore, this paper designs a scene complexity evaluation model based on the gradient boosting decision tree model for the research on the complexity calculation of the intelligent networked simulation test scene, which is used for the scene complexity evaluation of the simulation test of the intelligent networked vehicle.

SUMMARY OF THE INVENTION

In view of this, the present invention aims to propose a scene complexity evaluation method based on a gradient boosting decision tree model, so as to solve the test difficulty that the tester will not be able to know the relevant test scene concisely and clearly for the test of the automatic driving control algorithm, It is also impossible to do the problem of progressive testing.

In order to achieve the above object, the technical scheme of the present invention is achieved in this way:

A method for evaluating scene complexity based on a gradient boosting decision tree model, comprising the following steps:

S1. Generate a simulation scene with the help of simulation software based on the driving scene data based on multi-sensor fusion;

S2. With the help of the experience of a number of simulation test and scene development experts, the complexity of the simulated driving scene samples is scored;

S3. Input the results of multiple expert evaluations and the dynamic and static features that affect the complexity into the decision tree model for model training;

S4, upgrade the decision tree model;

S5. Obtain the characteristic parameter data set that affects the model, take part of the data set as the training set and the other part as the test set, and use 5-fold cross-validation to debug to obtain the complexity evaluation model;

S6. Directly input the driving scene data to be evaluated into the model to calculate the scene complexity;

S7. The input driving scene data is firstly divided into dynamic features and static features, and then the scene complexity is obtained after comprehensive scoring according to the influencing features.

Further, the parameters in the step S1 include dynamic traffic flow information and road information.

Further, in the step S4, the decision tree model is upgraded to a LightGBM gradient-based decision tree model.

Further, the characteristic parameters in the step S5 include TTC, light intensity, the longitudinal distance between the target vehicle and the main vehicle, the average vehicle speed of the main vehicle, the acceleration dimension of the target vehicle, the initial vehicle speed of the main vehicle, the maximum speed of the target vehicle, the average speed of the target vehicle, Interfering with the driving behavior of the vehicle and the driving behavior of the host vehicle.

Further, the static features in the step S3 or S7 include: environmental information class features, road information class features;

Further, the environmental information class features include weather, light intensity, and visibility.

Further, the road information class features include road curvature, lane line features, and the number of lanes tunnels.

Further, the dynamic features in the step S7 include: main vehicle information type characteristics and traffic participant information type characteristics.

Further, the vehicle owner information class features include vehicle owner behavior, average speed, and acceleration dimensions.

Further, the traffic participant information class features include the driving behavior of the target vehicle, the number of traffic participants, the distance from the host vehicle, and the relative speed.

Compared with the prior art, the method for evaluating the scene complexity based on the gradient boosting decision tree model according to the present invention has the following advantages:

The method of the present invention provides a classification method of intelligent networked driving simulation scenarios based on scene complexity, and provides a simulation scenario evaluation model based on scene complexity, and the evaluation scheme can provide clear and concise automatic driving The complexity estimation of the test scene meets the tester's requirement that the driving scene can be selected according to the complexity of the scene.

Description of drawings

The accompanying drawings constituting a part of the present invention are used to provide further understanding of the present invention, and the exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the attached image:

1 is a flowchart of a method for evaluating scene complexity based on a gradient boosting decision tree model according to an embodiment of the present invention;

2 is a structural diagram of a method for evaluating scene complexity according to an embodiment of the present invention;

FIG. 3 is an analysis diagram of a feature parameter with a large extraction influence according to an embodiment of the present invention.

Detailed ways

It should be noted that the embodiments of the present invention and the features of the embodiments may be combined with each other under the condition of no conflict.

In the description of the present invention, it should be understood that the terms "center", "portrait", "horizontal", "top", "bottom", "front", "rear", "left", "right", " The orientation or positional relationship indicated by vertical, horizontal, top, bottom, inner, outer, etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and The description is simplified rather than indicating or implying that the device or element referred to must have a particular orientation, be constructed and operate in a particular orientation, and therefore should not be construed as limiting the invention. In addition, the terms "first", "second", etc. are used for descriptive purposes only, and should not be construed as indicating or implying relative importance or implying the number of indicated technical features. Thus, a feature defined as "first", "second", etc., may expressly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

In the description of the present invention, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense, unless otherwise expressly specified and limited, for example, it may be a fixed connection or a detachable connection Connection, or integral connection; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, can be internal communication between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood through specific situations.

The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

As shown in Figures 1 to 2, a method for evaluating scene complexity based on a gradient boosting decision tree model includes the following steps:

Step 1: Extract the parameters of the collected driving scene data, including dynamic traffic flow information and road information, and use the decision tree model and the simulation scene modeling software to generate a corresponding simulated driving scene;

Step 2: With the help of the experience of multiple simulation testing and scene development experts, the complexity of the simulated driving scene samples is scored;

Step 3: Input the results of evaluation by multiple experts and the summarized complexity feature elements into the decision tree model. With the help of the decision tree model, it is easy to understand, has strong interpretability, and has fast prediction speed. It can automatically combine multiple features without any The interactions between features are dealt with stressfully and are computed with the advantage of non-parameterization.

As shown in Figure 3, the characteristic parameters with greater influence are obtained through calculation and analysis;

Step 4: The simulated driving scene is upgraded on the basis of the decision tree model, and upgraded to the LightGBM model to simulate the driving scene. LightGBM is an evolutionary version of the gradient boosting decision tree algorithm. The complexity of the driving scene data.

As shown in Figure 2, step 5: Obtain the feature parameters that affect the model, and the feature parameters used when the decision tree node is split, reflecting the importance of the feature. It can be seen from Table 1 that the top ten most important features are TTC, Light intensity, the longitudinal distance between the target car and the host car, the average speed of the host car, the acceleration dimension of the target car, the initial speed of the host car, the maximum speed of the target car, the average speed of the target car, the driving behavior of the distracting car, and the driving behavior of the host car.

Step 6: Use 80% of the data set as the training set and 20% as the test set, using 5-fold cross-validation. After repeated debugging, the model parameters are finally selected, and the complexity evaluation model is obtained.

Step 7: After the complexity evaluation model is generated, the driving scene data to be evaluated can be directly input into the model to calculate the scene complexity.

Step 8, Step 9, Step 10: The input driving scene data is firstly divided into dynamic features and static features, and then the scene complexity is obtained after comprehensive scoring according to each influence feature.

Table 1 Feature map annotations

1. Scene static factors

In terms of scene static complexity, environmental information features are extracted: weather (sunny, cloudy, rain and snow), illumination (light intensity), visibility (low, medium, high); road information features: road curvature ([- 0.01,0.01]), number of lane features, lane line features (complete and clear, incomplete or unclear), ground indication line features (complete and clear, incomplete or unclear, none), tunnel (into the tunnel, in the tunnel Driving, out of tunnel, non-tunnel driving), toll station (passing, not passing).

2 scene dynamic factors

In terms of the dynamic complexity of the scene, the information features of the main vehicle are extracted: the behavior of the main vehicle (following the line, following the vehicle, changing lanes), the initial speed of the main vehicle, the average speed of the main vehicle, the maximum speed of the main vehicle, and the acceleration dimension of the main vehicle. (close to uniform driving, normal driving, intense driving); traffic participant characteristics: target car driving behavior (no target car, steady driving, cut-out, cut-out failure, cut-in, acceleration, deceleration, deceleration to stop, stop-and-go ), the number of traffic participants (0, 1, 2), the initial speed of the target car, the average speed of the target car, the maximum speed of the target car, the relative speed of the target car and the own car, the relative acceleration of the target car, the acceleration dimension of the target car (close to a constant speed driving, normal driving, intense driving), the longitudinal distance between the target car and the host car, the lateral distance between the target car and the host car; the driving behavior of the interfering car (no target car, stable driving, cut-out, cut-out failure, cut-in, acceleration, deceleration, Deceleration to stop, stop-and-go), the relative speed of the interfering car and the own vehicle, the relative acceleration of the interfering car, the dimension of the interfering car acceleration (close to constant speed, normal driving, intense driving), the initial speed of the interfering car, and the average speed of the interfering car , the maximum speed of the jamming car, the longitudinal distance between the jamming car and the host car, and the lateral distance between the jamming car and the host car; TTC.

Finally, the gradient decision tree algorithm is used to train the model through massive samples.

The specific implementation method is as follows:

As shown in Figure 1, for the above training set that has been scored by experts, the complexity rating results are used as labels of different samples, denoted as f(x), and the different features of samples are denoted as vector x. The gradient boosting decision tree model is selected as the model for evaluating the complexity of the scene, there are:

where w is the weight and Φ is the set of weak classifiers (regressors).

Input: training data set T={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…,(x _N ,y _N )}; loss function L(y,f(x)); basis function set {b(x;γ)};

Output: Additive model f(x).

The specific steps of the algorithm are as follows:

(1) Initialize f ₀ (x)=0

(2) For m=1,2,...M

(a) Minimize the loss function

get parameters β _m , γ _m

(b) Update

f _m (x)=f _m-1 (x)+βb(x; γ _m )

(3) Get the additive model

In the model training process, 80% of the labeled data set is used as the training set and 20% as the test set, and 5-fold cross-validation is used. After repeated debugging, the parameters of the model are selected, and the obtained weights and regression data of the relevant parameters are used to evaluate the complexity of the new autonomous driving scenarios.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the present invention. within the scope of protection.

Claims (10)

1. A scene complexity evaluation method based on a gradient lifting decision tree model is characterized by comprising the following steps:

s1, generating a simulation scene based on the driving scene data of the multi-sensor fusion by means of simulation software;

s2, carrying out complexity grading on the simulated driving scene sample by means of experience of a plurality of simulation tests and scene development experts;

s3, inputting the results of multiple expert evaluations, dynamic characteristics and static characteristics influencing complexity into a decision tree model for model training;

s4, upgrading the decision tree model;

s5, obtaining a characteristic parameter data set influencing the model, taking one part of the data set as a training set and the other part of the data set as a test set, and obtaining a complexity evaluation model by adopting 5-fold cross validation debugging;

s6, directly inputting driving scene data to be evaluated into the model for calculating scene complexity;

and S7, splitting the input driving scene data into dynamic features and static features, and then comprehensively scoring according to the influence features to obtain scene complexity.

2. The method for evaluating scene complexity based on the gradient boosting decision tree model according to claim 1, wherein: the parameters in the step S1 include dynamic traffic flow information and road information.

3. The method for evaluating scene complexity based on the gradient boosting decision tree model according to claim 1, wherein: in step S4, the decision tree model is upgraded to a LightGBM-based gradient decision tree model.

4. The method for evaluating scene complexity based on the gradient boosting decision tree model according to claim 1, wherein: the characteristic parameters in the step S5 comprise TTC, illumination intensity, longitudinal distance between the target vehicle and the main vehicle, average speed of the main vehicle, acceleration dimension of the target vehicle, initial speed of the main vehicle, maximum speed of the target vehicle, average speed of the target vehicle, driving behavior of an interference vehicle and driving behavior of the main vehicle.

5. The method for evaluating scene complexity based on the gradient boosting decision tree model according to claim 1, wherein: the static features in the step S3 or S7 include: environmental information type characteristics, road information type characteristics.

6. The method according to claim 5, wherein the scene complexity evaluation method based on the gradient boosting decision tree model comprises: the environmental information type characteristics comprise weather, illumination intensity and visibility.

7. The method according to claim 5, wherein the scene complexity evaluation method based on the gradient boosting decision tree model comprises: the road information type characteristics comprise road curvature, lane line characteristics and lane number tunnels.

8. The method for evaluating scene complexity based on the gradient boosting decision tree model according to claim 1, wherein: the dynamic features in step S7 include: the main vehicle information type characteristics and the traffic participant information type characteristics.

9. The method according to claim 8, wherein the scene complexity evaluation method based on the gradient boosting decision tree model comprises: the owner information type characteristics comprise owner behaviors, average speed and acceleration dimensionality.

10. The method according to claim 8, wherein the scene complexity evaluation method based on the gradient boosting decision tree model comprises: the traffic participant information type characteristics comprise driving behaviors of the target vehicles, the number of traffic participants, distances between the traffic participants and the main vehicles and relative speeds.

Images