1. Introduction
The roads on which vehicles drive are generally divided into structured roads and unstructured roads. Structured roads refer to roads with clear road markings, a single environment, and distinct geometric features, such as highways and urban arteries. In a broad sense, structured roads can be understood as unobstructed road environments with flat ground and good visual effects for white line navigation. Unstructured roads refer to roads with unclear road boundaries, complex environments, and diverse road shapes that have a low degree of structural such as urban nonmain roads and rural streets. Compared with structured roads, unstructured roads also have the characteristics of exhibiting diverse types and states of barriers and surface coverage, as well as complex and variable road conditions, all of which increase the difficulty for intelligent vehicles to perceive the environment.
Vehicles have been developing in intelligent and directions. The testing and evaluation of intelligent vehicles are the basic guarantee to achieving the safe use of transport. According to the degree of interaction between the test environment and external traffic elements and the degree of test risk, the test and evaluation can be divided into four stages, including a closed-environment normative test, a closed-environment validation test, a limited open-environment validation test, and an open-environment demonstrative validation [
1]. Closed-environment normative testing has the lowest degree of interaction and test risk compared to the other phases. The purpose of this test is to assess the performance of the vehicle by its performance in various aspects. This phase of testing does not allow access to other vehicles and related elements and is conducted in a dedicated closed environment. The test conditions are defined, the test methods, instruments and results are reproducible, and the test results are precise values of physical quantities. Closed-environment validation tests are also conducted in dedicated closed environments and require the performance of multiple tests to obtain probabilistic indicators. Part of the limited open-environment validation test allows other vehicles and related elements to enter and exit the area in order to evaluate the safety, reliability and other performance of the vehicle. Open-environment demonstration validation is conducted in a completely open real-traffic environment to improve the environmental synergy between the vehicle, the road, traffic facility equipment and other traffic participants. Although the scenarios, test methods, test results, and relative difficulty of the four phases are different, they are all essential and have means of providing technical guarantees of the safety, reliability, and use of intelligent vehicles.
At present, the research on the testing and evaluation system for intelligent vehicles in special roads is not perfect, and the intelligent barrier avoidance function of vehicles cannot be assessed. As one of the core functions of autonomous driving, the intelligent function of barrier avoidance plays an important role in improving traffic efficiency and ensuring driving safety. Therefore, establishing a scientific and reasonable comprehensive evaluation method for application to the vehicle intelligent barrier avoidance function under special roads is of great significance. Testing and evaluation before the actual use of the vehicle is of great significance in the process of verifying function and ensuring driving safety.
The premise of a comprehensive evaluation is to build an evaluation index system, and the selection of indexes should follow the principles of feasibility and scientific. Integrated evaluations are centered on obtaining the evaluation results of the subject through the use of the appropriate methodology. The process mainly includes calculating the index weight and establishing a comprehensive evaluation model. According to the source of data, the methods for calculating weights can be divided into three categories, including the subjective method, objective method, and combination weighting method [
2]. According to the knowledge and experience of experts in the corresponding fields, the subjective method compares, assigns, and calculates the importance of the evaluation indicators to determine the weight. Examples include the analytical hierarchy process (AHP) [
3], the expert survey method (Delphi) [
4], and the order relationship analysis (G1) method [
5]. Subjective weight depends too much on expert experience. The objective method uses mathematical methods to calculate weights according to the quantitative relationship between the initial data of indicators. Examples include principal component analysis [
6], factor analysis [
7], the entropy method [
8], and the CRITIC (criteria importance though intercriteria correlation) method [
9]. If the measured data are not typical and universal, there may be an unreasonable weight distribution. The combined weighting method combines the weights obtained using the subjective and objective methods according to different preference coefficients. It not only retains the information expression of expert experience and knowledge and the subjective intention of decision-makers in the subjective method, but it also retains the information of the internal relationship between indicators and evaluation objects in the objective method. It has the effect of offering complementary advantages, and the evaluation results are relatively more scientific and reasonable. The multiple evaluation indicators are synthesized into a whole comprehensive evaluation by establishing a comprehensive evaluation model. The main methods include the grey correlation method [
10], TOPSIS (technique for order preference by similarity to an ideal solution) method [
11], BP (back propagation) network [
12], and fuzzy comprehensive evaluation method [
13].
In terms of the evaluation of intelligent vehicles, Sun [
14] built an evaluation index system from three aspects, namely, safety, intelligence, and ride comfort. Then, the author calculated the weight based on the EAHP (extended analytical hierarchy process) method, and comprehensively evaluated the intelligent behavior of driverless vehicles through gray correlation analysis. Zhao [
15] established an evaluation index system based on typical working conditions, such as intersections and car following, and evaluated them through the entropy–cost function method. Although subjective factors were excluded, the system relied too much on the measured data and did not have universality. Huang [
16] evaluated the overall performance of the driverless vehicle through the AHP–entropy method and fuzzy comprehensive evaluation. To avoid the complexity of the AHP calculation and the possibility of failing the consistency test, Li [
17] analyzed the test content of the China Smart Car Future Challenge, built an evaluation index system from the four aspects of safety, systematicity, stability, and speed, and calculated the weight based on the G1 method and entropy method. The G1 method is an improved weighting method based on AHP that avoids the complex process of constructing judgment matrices and consistency checks. However, the entropy method ignores the internal relationship between indicators. CRITIC method has better applications in other comprehensive evaluation fields, such as power grid [
18] and mining [
19], compared with the entropy method, which considers the conflict and contrast degree of evaluation indicators. The driverless vehicle is in the development stage, and so the subjective assessment of experts cannot be excessively relied on in the function evaluation process. Additionally, the measured data of indicators cannot be completely relied on.
Firstly, this paper focuses on the closed-environment normative test, the evaluation object is the intelligent barrier avoidance function of the vehicle under special roads, and the performance of the function is evaluated by measuring the specific physical quantities of each index. For indicators that can obtain data, the CRITIC method of objective weighting is used to determine the weight. For qualitative indicators that are difficult to quantify, the G1 method is used to determine the weight in order to avoid the single influence of subjective and objective factors in the evaluation method. A complex evaluation index system can be regarded as a grey system. To reduce the influence of subjective factors in the final evaluation result, the grey correlation analysis method is chosen as the evaluation method, which provides a basis for the application of vehicles.
4. Discussion
The comprehensive evaluation score clearly reflects the overall situation of the intelligent barrier avoidance function of the test vehicle and the differences in each specific function. The overall score for the intelligent barrier avoidance function of the test vehicle under special roads is 87.03, and some of the functions of this vehicle can still be improved. The intelligent barrier avoidance function of the test vehicle facing negative barriers has a higher score than that facing positive barriers. This indicates that in the same scenario, the test vehicle’s ability to avoid negative barriers is better than its ability to avoid positive barriers. The difference between the two scores may be influenced by the test data, with some indicators scoring too high and subjective weights only partially corrected. Among the intelligent barrier avoidance functions facing positive barriers, the score of the intelligent barrier crossing function is the best, being significantly superior to that of detour. In the intelligent barrier avoidance function for negative barriers, the ability to recognize and perceive barriers and the intelligent barrier avoidance function through detour are superior to those of braking and crossing. The scores of each function can reflect the functional differences in the same scene, providing a targeted basis for improving the intelligent barrier avoidance function of the test vehicle.
The test evaluation in this paper only verifies the correctness and validity of the proposed method. Due to the limitation of the number of test samples, the test results (mean values) are selected as the reference sequence in the comprehensive evaluation study, which cannot further evaluate the advantages and disadvantages of vehicle functions. In subsequent research, the law of the test results can be studied through multiple-vehicle tests to obtain the optimal limit value of each index, and a more reasonable reference sequence can be set accordingly. Along these lines, further research on the evaluation score of the advantages and disadvantages of the interval can be used as the basis for the comprehensive evaluation. The evaluation method can then be applied to compare the functional advantages and disadvantages of different vehicles.