A scene complexity evaluation method and system

Abstract

The application discloses a scene complexity evaluation method and system, comprising the following steps: inputting a parameter set for describing a scene to be evaluated into a scene complexity evaluation model to output a scene complexity evaluation result, wherein the construction of the scene complexity evaluation model comprises the following steps: selecting several typical scenes; acquiring a parameter set for describing the typical scenes; acquiring driving data corresponding to the typical scenes, and obtaining difficulty labels corresponding to the typical scenes based on the driving data; and taking the parameter set of the typical scenes as input and taking the difficulty labels of the typical scenes as output, constructing a scene complexity evaluation model through an SVM or a decision tree algorithm. The application analyzes post-driving characteristics, takes driving difficulty as a form of scene complexity, takes difficulty labels as output, takes scene elements as previous input, constructs a scene complexity evaluation model, and thus measures complex traffic scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of transportation, and particularly relates to a scene complexity evaluation method and system. BACKGROUND

[0002] Quantification of scene complexity is conducive to transforming the road traffic environment from a complex state to a simple state, and is of great significance to the study of driving environment.

[0003] Existing research on traffic scene complexity mainly analyzes the influence weight of environmental factors or various elements in the scene. The complexity of environmental factors is calculated or obtained by expert voting method, and the weights of various complex factors in the environment are summed according to subjective weights to obtain the complexity evaluation; the influence weight of each part of the scene element is quantified, the corresponding complexity index of the composition element affecting the traffic complexity is analyzed, and the complexity classification of the overall scene is obtained, so as to realize the comprehensive quantitative evaluation of the scene complexity.

[0004] However, this method is mainly based on expert scoring, and is mixed with many subjective factors, lacking a certain objectivity. SUMMARY

[0005] In view of the above problems, the application provides a scene complexity evaluation method and system, which analyzes the post-driving characteristics, uses driving difficulty as a form of scene complexity to measure complex traffic scenes, and constructs a model based thereon for scene complexity evaluation.

[0006] In one aspect, the application provides a scene complexity evaluation method, comprising the following steps: inputting a parameter set for describing a scene to be evaluated into a scene complexity evaluation model to output a scene complexity evaluation result, wherein the construction of the scene complexity evaluation model comprises the following steps: selecting a plurality of typical scenes; obtaining a parameter set for describing the typical scenes; obtaining driving data corresponding to the typical scenes, and obtaining a difficulty label of the typical scenes based on the driving data; and constructing a scene complexity evaluation model by an SVM or a decision tree algorithm, taking the parameter set of the typical scenes as input and taking the difficulty label of the typical scenes as output.

[0007] Preferably, the parameter set comprises a lane width, a lane length and a lane curvature radius.

[0008] Preferably, the following factors are considered when selecting the typical scenes: straight roads, single-turn roads and continuous-turn roads.

[0009] Preferably, the driving data corresponding to the typical scenes is driving data in the driving process in the typical scenes.

[0010] Preferably, the driving data is driving data under a specified task.

[0011] Preferably, based on the driving data, the difficulty label of the typical scene is obtained, and the specific method is: based on the driving data, a representation feature for representing driving difficulty is obtained; based on the representation feature, a difficulty representation value of the typical scene is calculated; a K-means method is used to cluster the difficulty representation values of the obtained several typical scenes, and a difficulty label is assigned to the typical scene according to the clustering result.

[0012] Preferably, the representation feature is a feature related to vehicle collision frequency, vehicle collision severity, and vehicle stability.

[0013] Preferably, the representation feature includes: vehicle collision frequency, I-class collision times, II-class collision times, III-class collision times, IV-class collision times, I-class stability times, II-class stability times, III-class stability times, and IV-class stability times; wherein the vehicle collision frequency refers to the frequency of vehicle collision, the I-class collision times refer to the number of times of vehicle deceleration after vehicle collision, the II-class collision times refer to the number of times of vehicle stop after vehicle collision, the III-class collision times refer to the number of times of vehicle turning around less than one circle after vehicle collision, the IV-class collision times refer to the number of times of vehicle turning around more than one circle after vehicle collision, the I-class stability times refer to the number of times of vehicle stability without collision, the II-class stability times refer to the number of times of vehicle smooth driving after shaking without collision, the III-class stability times refer to the number of times of vehicle turning around less than one circle without collision, and the IV-class stability times refer to the number of times of vehicle turning around more than one circle without collision.

[0014] Preferably, based on the driving data, the difficulty label of the typical scene is obtained, and the specific method is: based on the driving data, the average speed corresponding to the typical scene is obtained; based on the representation feature and the average speed, an input set is obtained, the difficulty label assigned according to the clustering result is taken as the output, a BP neural network is used for supervised learning to construct a difficulty label output model; the difficulty label output model is trained until the model converges, that is, the training is completed; and based on the trained difficulty label output model, the difficulty label true value of the typical scene is obtained.

[0015] Preferably, when the scene complexity evaluation model is constructed, the difficulty label true value of the typical scene is taken as the output.

[0016] In another aspect, the present application provides a scene complexity evaluation system, comprising: an evaluation unit configured to input a parameter set for describing a scene to be evaluated into a scene complexity evaluation model to output a scene complexity evaluation result; and a model construction unit configured to select a typical scene, obtain a parameter set for describing the typical scene, obtain driving data corresponding to the typical scene, obtain a difficulty label based on the driving data, and construct a scene complexity evaluation model by using an SVM or a decision tree algorithm with the parameter set of the typical scene as input and the difficulty label as output.

[0017] Compared with the prior art, the present application has the following beneficial effects:

[0018] (1) The present application analyzes the post-driving characteristics, uses driving difficulty as a form of scene complexity, uses a difficulty label as output, uses scene components as input, and constructs a scene complexity evaluation model, thereby realizing the measurement of complex traffic scenes.

[0019] (2) The evaluation method of the present application is objective and can reduce the influence of subjective factors. The scene parameter set (i.e., the parameter set obtained after quantizing the scene components) is input into the model, and the difficulty label of the current driving task can be directly output, the complexity of the driving situation can be judged, and the safety of the traffic environment can be further judged, thereby providing a new method for driving environment design, traffic safety warning, traffic control, etc.

[0020] (3) In different scenes, the specified tasks during driving (i.e., the purposes to be achieved during driving) are different, which will have a certain influence on the evaluation of scene complexity. The driving data of the present application is preferably driving data under a specified task, and the influence of the task requirements can be considered when evaluating the scene complexity.

[0021] (4) The selected characteristic features for calculating the difficulty representation value belong to result type data, and the difficulty representation value calculated based thereon can more intuitively represent the driving difficulty.

[0022] (5) The speed during driving in the scene is an important influencing factor of driving difficulty, which has a certain influence on driving difficulty. In the present application, in addition to considering the characteristic features representing driving difficulty, the average speed corresponding to the scene is also considered when using the BP network for supervised learning, so that a difficulty label true value that can more accurately represent the scene difficulty can be obtained, and subsequently, the difficulty label true value can be directly obtained based on the characteristic features and the average speed, without the need for redundant calculation, which is convenient and fast. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 Typical scene selected in Embodiment 1 of the present application;

[0024] Figure 2 This is a comparison chart of velocity data before and after preprocessing in Embodiment 1 of the present invention;

[0025] Figure 3 This is a comparison chart of the error between the predicted and expected values ​​of the BP neural network test set in Embodiment 1 of the present invention;

[0026] Figure 4 This is a performance graph of the model constructed using the SVM algorithm in Embodiment 1 of the present invention;

[0027] Figure 5 This is a performance graph of the model constructed using the decision tree algorithm in Embodiment 1 of the present invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1

[0030] This embodiment provides a method for evaluating scene complexity, including the following steps:

[0031] The set of parameters used to describe the scenario to be evaluated is input into the scenario complexity evaluation model to output the scenario complexity evaluation result.

[0032] The construction of the scenario complexity evaluation model includes the following steps:

[0033] (1) Select several typical scenarios; obtain a set of parameters to describe the typical scenarios.

[0034] When selecting the typical scenarios, the following factors are considered: straight roads, roads with a single turn, and roads with continuous turns, to simulate various road sections that may occur in actual driving. In this embodiment, when selecting typical scenarios, the Project CARS racing game platform is used to select 3 typical game tracks and divide them into 13 typical scenarios, such as... Figure 1 As shown.

[0035] The parameter set is also a parameter set obtained after quantizing the elements constituting the scene. The elements constituting the driving scene in the actual driving task are various and interrelated, and the complexity of the driving scene is a comprehensive response under the action of various scene factors. The length of the lane in the scene affects the driving time; the narrow lane width increases the psychological pressure of the driver during driving, and when there are more participants in the scene, the risk of collision between vehicles will also increase; the small lane curvature radius increases the driving difficulty when the vehicle turns, which causes great hidden dangers to safe driving. Relevant statistical data shows that traffic accidents occurring on curves account for more than 60% of all traffic accidents. As can be seen, the lane length, lane width and lane curvature radius of the scene are basic and important parameters for representing the characteristics of the scene.

[0036] Based on this, in a preferred scheme of the embodiment, the parameter set includes: lane width, lane length, lane curvature radius. Wherein, the lane length is obtained by using the edge detection algorithm in the image recognition method, combined with the actual length of the given track in the game scene, combined with the edge detection to obtain the scene pixel points. The lane length adopts the average width of the lane. The lane curvature radius is obtained by using image denoising, image binarization processing, Hough transform circle detection, morphological operation road thinning and other steps to calculate the curvature radius; if there are multiple curvature radii obtained by Hough transform circle detection, the average value is taken. As shown in Table 1, the parameter set of 13 typical scenes of the embodiment is listed:

[0037] Table 1 Parameter set of typical scenes

[0038]

[0039]

[0040] (2) Obtain the driving data corresponding to the typical scene, and obtain the difficulty label of the typical scene based on the driving data.

[0041] The driving data corresponding to the typical scene is the driving data in the driving process in the typical scene. In this embodiment, the Project CARS racing game platform is used to simulate the driving task by using the game rudder, and the driving data of the test personnel in the driving process is collected, and the output form is a video recording in the driving process.

[0042] As a preferred solution, the embodiment considers that different task requirements in the same scene have certain influence on the determination of driving difficulty, and therefore the driving data is driving data under a specified task. In the embodiment, the specified task is selected from the following two tasks: one is to complete as quickly as possible under the premise of ensuring safety and reducing collision (that is, safety is preferred, denoted as a safety driving task); and the other is to ensure safety as much as possible under the premise of striving for a top ranking (that is, speed is preferred, denoted as a racing driving task).

[0043] Based on the driving data, a difficulty label of the typical scene is obtained, and the specific steps are as follows:

[0044] S1, based on the driving data, a representation feature for representing driving difficulty is obtained.

[0045] S2, based on the representation feature, a difficulty representation value of the typical scene is calculated.

[0046] S3, the K-means method is used to cluster the difficulty representation values of the obtained several typical scenes, and the difficulty label is assigned to the typical scene according to the clustering result.

[0047] Further, in order to more conveniently and accurately represent the driving difficulty of the typical scene, the following steps are further included:

[0048] S4, based on the driving data, an average speed corresponding to the typical scene is obtained.

[0049] S5, based on the representation feature and the average speed, an input set is obtained, with the difficulty label assigned according to the clustering result as the output, a BP neural network is used for supervised learning, and a difficulty label output model is constructed.

[0050] S6, the difficulty label output model is trained until the model converges, that is, the training is completed.

[0051] S7, based on the trained difficulty label output model, a difficulty label true value of the typical scene is obtained.

[0052] In S1, the characteristic features are related to vehicle collision frequency, vehicle collision severity, and vehicle stability. In this embodiment, the characteristic features include: vehicle collision frequency, level I collision frequency, level II collision frequency, level III collision frequency, level IV collision frequency, level I stability frequency, level II stability frequency, level III stability frequency, and level IV stability frequency. The vehicle collision frequency refers to the frequency of vehicle collisions, i.e., the ratio of total collision frequency to driving time. The level I collision frequency refers to the number of times the vehicle decelerates after a collision. The level II collision frequency refers to the number of times the vehicle stops after a collision. The level III collision frequency refers to the number of times the vehicle turns around less than one full circle after a collision. The level IV collision frequency refers to the number of times the vehicle turns around more than one full circle after a collision. The level I stability frequency refers to the number of times the vehicle is stable without a collision. The level II stability frequency refers to the number of times the vehicle stabilizes after shaking without a collision. The level III stability frequency refers to the number of times the vehicle turns around less than one full circle without a collision. The level IV stability frequency refers to the number of times the vehicle turns around more than one full circle without a collision.

[0053] In the above characteristic features, the number of events related to collision severity and vehicle stability is obtained based on the derived video, and the event is labeled according to the corresponding condition, and the number is counted. The data related to speed is obtained by image recognition, frame extraction, and screenshot of the speed indicator position, and recognized by OCR recognition algorithm. In order to avoid affecting the calculation results and model results in subsequent data application, the manually labeled event records need to be preprocessed, and the loop data is removed and integrated. The video recognition data is preprocessed to remove non-numeric type and error data such as random code, and the incorrectly recognized numeric type data is smoothed, such as Figure 2

[0054] In S2, the characteristic features belong to the result type data, i.e., the post-driving features. The difficulty representation value calculated based on this can more intuitively represent the driving difficulty. In this process, the entropy weight method is used to calculate the weight. It should be noted that in order to avoid the influence of individual differences of test personnel on the objectivity of the scene difficulty, multiple test personnel are selected to drive in each scene. Based on the driving data of each test personnel in the scene, a set of characteristic features can be obtained, and based on each set of characteristic features, a difficulty representation value can also be obtained. The difficulty representation value of the typical scene obtained in S2 is the average of the difficulty representation values of all test personnel in the scene.

[0055] ​In S3, the difficulty label of the embodiment has 4 levels, that is, the driving difficulty level is divided into 4 levels (recorded as 1, 2, 3, and 4 from easy to difficult), and too many difficulty levels will lead to indistinct division between difficulty levels, and too few difficulty levels will lead to unclear scene difficulty. In order to verify the effectiveness of the selected representation features in the embodiment, the difficulty labels assigned after clustering are compared with the subjective cognitive difficulty, as shown in Tables 2 and 3.

[0056] Table 2 Comparison of safety driving task difficulty labels

[0057]

[0058]

[0059] Table 3 Comparison of racing driving task difficulty labels

[0060] Scene number Subjective difficulty rating Rating result 1 3 3 2 1 1 3 4 3 4 2 1 5 1 1 6 3 3 7 4 3 8 4 2 9 2 2 10 2 1 11 2 3 12 4 4 13 3 2

[0061] As can be seen from the above table, the difficulty labels obtained in the embodiment are mostly consistent with the subjective cognitive results of people.

[0062] In S4-S7, considering that the speed of driving in the scene has a great influence on the driving difficulty, the average speed is also used for model construction. In S5, when constructing the difficulty label output model, the output of the model is the difficulty label assigned according to the clustering result, and the input set of the model is obtained based on the representation features and the average speed. Specifically, since there is not only one test personnel in each scene, the input of the model based on the representation features can be the sum of the representation features of all test personnel in the scene, or the average of the representation features of all test personnel, for example, there are n test personnel in a scene, and the representation features of the “Ⅰ-level collision times” are x1, x2, …, xn respectively. Then the input of the model obtained based on the representation features of the “Ⅰ-level collision times” in S5 can be (x1+x2+…+xn), or (x1+x2+…+xn) / n. It should be noted that if the sum of the representation features of all test personnel is used as the input of the model, the number of test personnel in each scene needs to be the same. Similarly, the input of the model based on the average speed can be the average of the average speeds of all test personnel in the scene. n n n

[0063] Based on the input set obtained based on the representation features and the average speed, the model is constructed with the difficulty label assigned according to the clustering result as the output, to obtain the difficulty label true value. Subsequently, the model can be directly used to obtain the difficulty label true value of the scene based on the representation features and the average speed, without unnecessary calculation. The error comparison between the predicted value and the expected value of the test set of the BP neural network in the embodiment is as follows​​​Figure 3 The model error results are shown in Table 4.

[0064] Table 4 Model error results

[0065] Mean absolute error MAE 7.801e-06 Mean squared error MSE 6.6971e-06 Root mean squared error RMSE 0.0025879 Coefficient of determination R^2 1

[0066] (3) Taking the parameter set of the typical scene as input and the difficulty label as output (here, the difficulty label can be a difficulty label assigned after clustering by the K-means method, or a difficulty label true value obtained through a difficulty label output model, and in this embodiment, the difficulty label true value is preferred), a scene complexity evaluation model is constructed through an SVM or decision tree algorithm.

[0067] The model performance constructed through the SVM algorithm is shown in Table 5. Figure 4 As shown in Table 5, the prediction accuracy of the SVM model is 82.93%; the model performance constructed through the decision tree algorithm is shown in Table 6. Figure 5 As shown in Table 6, the prediction accuracy of the decision tree model is 85%. It can be seen that the SVM classifier and the decision tree model constructed can both well complete the classification purpose of the driving task, and achieve a relatively optimal performance.

[0068] This embodiment also provides a scene complexity evaluation system, comprising:

[0069] An evaluation unit configured to input a parameter set for describing a scene to be evaluated into a scene complexity evaluation model to output a scene complexity evaluation result.

[0070] A model construction unit configured to:

[0071] select a plurality of typical scenes;

[0072] obtain a parameter set for describing the typical scenes;

[0073] obtain driving data corresponding to the typical scenes, and obtain difficulty labels of the typical scenes based on the driving data;

[0074] take the parameter set of the typical scene as input and the difficulty label of the typical scene as output, and construct a scene complexity evaluation model through an SVM or decision tree algorithm.

[0075] The above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, the ordinary skilled in the art can still modify the technical solutions recorded in the above embodiments, or make equivalent replacements for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the specification of the present application.

Claims

1. A method of scene complexity assessment, characterized by, The method comprises the following steps: Inputting a parameter set for describing a scene to be evaluated into a scene complexity evaluation model to output a scene complexity evaluation result, wherein the construction of the scene complexity evaluation model comprises the following steps: Selecting a plurality of typical scenes; Obtaining a parameter set for describing the typical scenes; Obtaining driving data corresponding to the typical scenes, and obtaining a difficulty label of the typical scenes based on the driving data; the specific method is: Based on the driving data, obtain a representation feature for representing driving difficulty; Based on the representation feature, calculate a difficulty representation value of the typical scene; Using a K-means method to cluster the difficulty representation values of the plurality of typical scenes, and assigning a difficulty label to the typical scenes according to the clustering result; The representation feature is a feature related to vehicle collision frequency, vehicle collision severity and vehicle stability; The representation feature includes vehicle collision frequency, I-level collision frequency, II-level collision frequency, III-level collision frequency, IV-level collision frequency, I-level stability frequency, II-level stability frequency, III-level stability frequency and IV-level stability frequency; The vehicle collision frequency refers to the frequency of vehicle collision, The I-level collision frequency refers to the frequency of vehicle deceleration after vehicle collision, The II-level collision frequency refers to the frequency of vehicle stop after vehicle collision, The III-level collision frequency refers to the frequency of vehicle turning less than one round after vehicle collision, The IV-level collision frequency refers to the frequency of vehicle turning more than one round after vehicle collision, The I-level stability frequency refers to the frequency of vehicle stability without collision, The II-level stability frequency refers to the frequency of vehicle smooth driving after shaking without collision, The III-level stability frequency refers to the frequency of vehicle turning less than one round without collision, The IV-level stability frequency refers to the frequency of vehicle turning more than one round without collision; The parameter set includes lane width, lane length and lane curvature radius.

2. The scene complexity assessment method of claim 1, wherein, The following factors are considered when selecting the typical scenes: straight road, single-turn road and continuous-turn road.

3. The scene complexity assessment method of claim 1, wherein, The driving data is driving data under a specified task.

4. The scene complexity assessment method of claim 1, wherein, Based on the driving data, obtaining the average speed corresponding to the typical scenes; 5. The scene complexity assessment method of claim 1, wherein, Based on the representation feature and the average speed, obtaining an input set, and using a BP neural network to perform supervised learning to construct a difficulty label output model, wherein the difficulty label output model takes the difficulty label assigned according to the clustering result as the output; Training the difficulty label output model until the model converges, i.e., completing the training; Obtaining the difficulty label true value of the typical scenes based on the trained difficulty label output model. When constructing the scene complexity evaluation model, taking the difficulty label true value of the typical scenes as the output. The method comprises the following steps:

6. The scene complexity assessment method of claim 5, wherein, An evaluation unit configured to input a parameter set for describing a scene to be evaluated into a scene complexity evaluation model to output a scene complexity evaluation result; 7. A scene complexity assessment system, characterized by ​ ​ The model construction unit is configured to: select a plurality of typical scenes; obtain a parameter set for describing the typical scenes; obtain driving data corresponding to the typical scenes, and obtain a difficulty label of the typical scenes based on the driving data; construct a scene complexity evaluation model by using an SVM or a decision tree algorithm, with the parameter set of the typical scenes as input and the difficulty label of the typical scenes as output. The scene complexity evaluation model is constructed by the following steps: select a plurality of typical scenes; obtain a parameter set for describing the typical scenes; obtain driving data corresponding to the typical scenes, and obtain a difficulty label of the typical scenes based on the driving data; the specific method is: obtain a representation feature for representing driving difficulty based on the driving data; calculate a difficulty representation value of the typical scenes based on the representation feature; cluster the difficulty representation values of the plurality of typical scenes by using a K-means method, and assign difficulty labels to the typical scenes according to the clustering result; the representation feature is a feature related to vehicle collision frequency, vehicle collision severity, and vehicle stability; the representation feature includes vehicle collision frequency, I-level collision times, II-level collision times, III-level collision times, IV-level collision times, I-level stability times, II-level stability times, III-level stability times, and IV-level stability times; the vehicle collision frequency refers to the frequency of vehicle collision, the I-level collision times refer to the number of times of vehicle deceleration after vehicle collision, the II-level collision times refer to the number of times of vehicle stop after vehicle collision, the III-level collision times refer to the number of times of vehicle rotation less than one round after vehicle collision, the IV-level collision times refer to the number of times of vehicle rotation more than one round after vehicle collision, the I-level stability times refer to the number of times of vehicle stability without collision. the II-level stability times refer to the number of times of vehicle smooth driving after shaking without collision, the III-level stability times refer to the number of times of vehicle rotation less than one round without collision, the IV-level stability times refer to the number of times of vehicle rotation more than one round without collision; construct a scene complexity evaluation model by using an SVM or a decision tree algorithm, with the parameter set of the typical scenes as input and the difficulty label of the typical scenes as output.

Images