Driving rule extraction based on cognitive behavior analysis
来源期刊:中南大学学报(英文版)2020年第1期
论文作者:梁军 赵宇橙 陈龙 蔡英凤 姚明 华国栋 朱宁
文章页码:164 - 179
Key words:cognitive driving behavior; driving rule extraction; cognitive theory; integrated algorithm
Abstract: In order to make full use of the driver’s long-term driving experience in the process of perception, interaction and vehicle control of road traffic information, a driving behavior rule extraction algorithm based on artificial neural network interface (ANNI) and its integration is proposed. Firstly, based on the cognitive learning theory, the cognitive driving behavior model is established, and then the cognitive driving behavior is described and analyzed. Next, based on ANNI, the model and the rule extraction algorithm (ANNI-REA) are designed to explain not only the driving behavior but also the non-sequence. Rules have high fidelity and safety during driving without discretizing continuous input variables. The experimental results on the UCI standard data set and on the self-built driving behavior data set, show that the method is about 0.4% more accurate and about 10% less complex than the common C4.5-REA, Neuro-Rule and REFNE. Further, simulation experiments verify the correctness of the extracted driving rules and the effectiveness of the extraction based on cognitive driving behavior rules. In general, the several driving rules extracted fully reflect the execution mechanism of sequential activity of driving comprehensive cognition, which is of great significance for the traffic of mixed traffic flow under the network of vehicles and future research on unmanned driving.
Cite this article as: ZHAO Yu-cheng, LIANG Jun, CHEN Long, CAI Ying-feng, YAO Ming, HUA Guo-dong, ZHU Ning. Driving rule extraction based on cognitive behavior analysis [J]. Journal of Central South University, 2020, 27(1): 164-179. DOI: https://doi.org/10.1007/s11771-020-4286-1.
J. Cent. South Univ. (2020) 27: 164-179
DOI: https://doi.org/10.1007/s11771-020-4286-1
ZHAO Yu-cheng(赵宇橙)1, LIANG Jun(梁军)1, CHEN Long(陈龙)1, CAI Ying-feng(蔡英凤)1,YAO Ming(姚明)2, HUA Guo-dong(华国栋)3, ZHU Ning(朱宁)4
1. Automotive Engineering Research Institute, Jiangsu University, Zhenjiang 212013, China;
2. School of Automotive and Transportation Engineering, Jiangsu University, Zhenjiang 212013, China;
3. Jiangsu Zhixing Future Automobile Research Institute, Nanjing 210000, China;
4. Department of Mechanical Engineering, Shizuoka University of Science and Technology,Shizuoka Bagu 437-0032, Japan
Central South University Press and Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract: In order to make full use of the driver’s long-term driving experience in the process of perception, interaction and vehicle control of road traffic information, a driving behavior rule extraction algorithm based on artificial neural network interface (ANNI) and its integration is proposed. Firstly, based on the cognitive learning theory, the cognitive driving behavior model is established, and then the cognitive driving behavior is described and analyzed. Next, based on ANNI, the model and the rule extraction algorithm (ANNI-REA) are designed to explain not only the driving behavior but also the non-sequence. Rules have high fidelity and safety during driving without discretizing continuous input variables. The experimental results on the UCI standard data set and on the self-built driving behavior data set, show that the method is about 0.4% more accurate and about 10% less complex than the common C4.5-REA, Neuro-Rule and REFNE. Further, simulation experiments verify the correctness of the extracted driving rules and the effectiveness of the extraction based on cognitive driving behavior rules. In general, the several driving rules extracted fully reflect the execution mechanism of sequential activity of driving comprehensive cognition, which is of great significance for the traffic of mixed traffic flow under the network of vehicles and future research on unmanned driving.
Key words: cognitive driving behavior; driving rule extraction; cognitive theory; integrated algorithm
Cite this article as: ZHAO Yu-cheng, LIANG Jun, CHEN Long, CAI Ying-feng, YAO Ming, HUA Guo-dong, ZHU Ning. Driving rule extraction based on cognitive behavior analysis [J]. Journal of Central South University, 2020, 27(1): 164-179. DOI: https://doi.org/10.1007/s11771-020-4286-1.
1 Introduction
Road traffic accidents have become a major public health problem worldwide. According to World Health Organization (WHO), every year, more than 1.25 million people worldwide suffer from road traffic accidents. There are also 20-50 million people who are suffering from non-fatal injuries, many of whom are disabled as a result [1]. The newest statistics of China Highway show that the number of road traffic accidents caused by people’s traffic violations accounted for 95.24%, and the death toll accounted for 95.42% [2]. For people’s traffic violations, bad driving behavior is the main cause, so driving behavior has become a hot spot in the field of traffic safety [3, 4]. In the long-term driving practice, the driver has formed a kind of extremely strong driving instinct to adapt to the environment [5], which is usually called “driving experience” or “driving habit”. These habits themselves have some complexity and ambiguity, and it is difficult to establish accurate models [6]. Therefore, combined with cognitive science [7-9], this paper learns, trains and extracts the corresponding driving behavior rules from the driver’s driving process, clarifies the corresponding rule relationships under different conditions, and reveals the inherent rules of driving habits. As shown in Ref. [10], during the 2016 Grand Cooperative Driving Challenge, the Chalmers Truck Team designed the open source software framework OpenDLV. The design originates from the postulate that the vehicle is fully autonomous to deal with even complex traffic scenarios. Next, it will attempt to comply with, and slightly extend, the current traffic rules and legislation for the purpose of being realistic for full-scale implementation. While using the extracted driving behavior rules, the driverless vehicles facing complex traffic scenarios can make the next stage of driving decisions according to certain prerequisites, which is more humane and effective based on driving experience. In addition, manual vehicle can get the driving decisions from nearby driverless vehicles so as to decide the following behavior, under the mixed traffic flow in the future, which contributes to ameliorating human-machine interface, reducing the traffic accidents, improving driving safety and alleviating traffic congestion.
The study of driving behavior models can be traced back to the driving-field-analysis proposed by GIBSN et al in the 1930s [11, 12]. In the following 50 years, the driving behavior model research did not greatly develop in general, and had not become a research hotspot until the beginning of this century. In 2006, VAA [13] pointed out and analyzed the results of previous driving behavior research, where human-specific cognition and emotion were effective means to ensure the prediction, avoidance and evaluation of dangerous situations during the execution of driving tasks. This confirms the important role of driver cognition in the driving process, but has not studied it in depth. Then, Professor SALVUCCI [14] used cognitive architecture to open up new ideas for driving behavior research, and added the humanization factor of vehicle-driver model in cognitive driving behavior research, which fully reflected the driver’s preference and its dominant position in the driving process. This complemented the cognitive learning style. At present, based on the working memory and continuous attention characteristics of the cognitive structure of driving behavior, GRAEFE [15] simulates the group differences of driving behavior, and studies the individualized and differential driving behaviors of different groups, which pushes the study to a new height. Nevertheless, the research focuses on the differences and personalization of driving behavior, and to some extent ignores some commonalities in driving behavior. WANG et al [16] proposed the driver’s cognitive activity chain. Based on the idea of cognitive activity chain, the fuzzy integration method was used to fuse the driving information of the road driving environment, and the driving mode of the vehicle after the driver’s task agglomeration was obtained, and the driving behavior model was constructed. HAMDAR et al [17] explored drivers’ cognitive and safety parameters in a risk-taking environment from behavioral psychology to acceleration modeling, of which the key point was the heterogeneity of driving behavior. Later, QU [18] proposed a driver behavior modeling method by establishing a driver behavior modeling framework with stochastic characteristics and rolling optimization ideas to simulate various driving characteristics and styles. However, the studies above did not consider the driver’s common driving habits and the dynamic characteristics of driving behavior. In order to address these deficiencies in the above literature, a cognitive driving behavior analysis method based on behavioral learning theory is introduced in the paper.
Cognitive science [19, 20] believes that the creation of artificial neural network (ANN) is one of the most effective means of deciphering human cognition. ANN has its own development characteristics, and it has a strong parallel structure and parallel processing capabilities, which plays an important role in real-time and dynamic control. ANN can grasp the ability of data induction and processing through training, so the problem can be solved when mathematical models and the like are difficult to deal with [21, 22]. In 1990, HANSEN and SALAMON [23] pioneered the neural network integration method. They proved that the neural network system can be significantly improved by simply training multiple neural networks and synthesizing the results. Artificial neural network integration (ANNI) has become an important technical means to solve complex and fuzzy engineering problems. In order to establish the driving behavior rule base in the cognitive driving behavior model, considering dynamic and real-time
driving behavior and the difference and complexity of the acquired data, this paper uses the ANNI method to extract the driving behavior rules [24-26][25].
In order to solve the biggest problem mentioned above that previous literature ignored driver’s common driving habits and the dynamic characteristics,this paper proposes a cognitive driving behavior analysis method based on behavioral learning theory. Driving rules extracted from this method can be applied to driverless vehicles under mixed traffic flow in the near future, which is of great value.
2 Cognitive driving behavior description and analysis
2.1 Cognitive driving behavior description
Behavioral learning theory believes that complex behavior is a compound response based on conditional connection. It is purposeful and is generated by the recognition of the corresponding changing relationship between goals and learning methods in the learning environment. Cognitive driving behavior is essentially an information processing system. The process is to achieve a coherent cognitive process of acquiring, encoding, storing, extracting and manipulating the vehicle according to certain procedures.
Definition 1 (Cognitive learning process): The cognitive learning process Γ is defined as a seven-tuple Γ=(I, C, S, E, P, Δ, γ0), where I is the acquired information and C is the information coding, S is the storage of cognitive information, E is the extraction of knowledge (rules), P is the application of knowledge, Δ is the set of programs, γ0∈Δ, and I×C×S×E×P→γ0 specifically describes a coherent cognitive learning process.
Definition 2 (Cognitive learning method): The habitual information processing method formed by individuals in the process of long-term cognitive learning is called cognitive learning.
Cognitive learning has a stable psychological tendency or preference. DGiΓ is the preference level of individual i for cognitive process. PiΓ and MiΓare
positive and negative propensity values.and are the sum of the propensity values, that is
the positive and negative tendency historical factors. Equation (1) represents the tendency or preference decision function under the influence of historical factors:
(1)
The parameters in the formula are normalized dimensionless scalars, λ is the correction factor, and k is the sample capacity.
Usually, individuals form cognitive styles that are different from other individuals because they differ in cognitive style. Drivers with different cognitive styles have certain differences in driving style, driving habits and adaptability to different driving environments. These differences will have an impact on driving behavior and traffic safety.
Cognitive driving behavior is a general term for the driver’s perception of the driving environment and the manipulation of the vehicle during the execution of the driving task. As shown in Figure 1, the cognitive driving behavior can be generally described as a series of detailed driving behavior cognitions, the execution of the driving subtasks, and the driver taking the driving components for completing the driving subtasks, during the execution of the driving task [27]. A driving task is a subset of a series of basic and simple driving subtasks that are repeated again and again.
The driver usually needs to remain awake in the driving situation throughout the driving process. Firstly, it is expressed as a comprehensive, dynamic, accurate and clear perception of the status and attribute information of each relevant element in the driving environment [28], mainly including the information of the vehicle itself, such as the speed of the vehicle, the condition of the vehicle, the change of the vehicle itself and the surrounding environment. There are also road traffic signals,signs, and so on. On this basis, the disorganized information is synthesized,and the meaning of the situation information is analyzed according to different driving purposes. At the same time, the driver’s own driving experience is combined to make predictions on possible situations, make reasonable response decisions on possible events, and maintain or change the expected traffic events.
Figure 1 Cognitive model driving behavior diagram
2.2 Cognitive driving behavior analysis
Generally speaking, crisis situations do not occur frequently in the driving process, but in fact, most of the traffic accidents occur in crisis situations. All along, the researchers believe that the driving behavior of the driver in a dangerous situation [29], such as the steering wheel or brake pedal operation, will exceed the operating range of normal driving, where the driver’s reaction is instinctive and subjects to human physical and psychological restriction of reaction ability. During the driving process [30], most of the dangerous situations occur suddenly, and traffic accidents are very likely to occur later. Therefore, only when a dangerous situation occurs, can the active safety system quickly make a correct judgment and urgently take corresponding technical measures to effectively prevent the occurrence of traffic accidents or reduce the loss of accidents.
In a word, cognitive driving behavior is to break down the driving process into several driving tasks. The driving task is triggered by a corresponding driving behavior rule, and the execution result is sent to the driving behavior motion buffer zone. The steering task of the running vehicle is realized by operating the steering, throttle and brake light operating components of the vehicle. The driving environment sensing module transmits the continuously sensed road traffic, weather, vehicle and other road traffic internal and external information to the driving task detection module, and implements driving rules generation through corresponding buffer storage, information processing, machine learning technologies, and so on. After the query and matching with the driving behavior rule base, the information that cannot be successfully matched is discarded, and a rule is triggered to be successful, and the execution result is sent to the driving behavior motion buffer, and then the driving control task is realized by the manipulation component. The decision task module collects and analyzes the driving data and queries the matching success information, obtains the driving subtask in the buffer through the target module, triggers one or more driving rules, and delivers the driving behavior to the dynamic buffer of the driving behavior, and then the operating module enables the release of driving decision tasks. Cognitive driving behavior is carried out in such a reciprocating manner.
3 Cognitive driving behavior rule extraction method
3.1 Basic theories
The above description and analysis of cognitive driving behaviors indicate that it is necessary to design a model and algorithm. This model and algorithm can explain the driving behavior, generate non-sequential rules, do not need to discretize continuous input variables, have high fidelity, and have high safety during driving. In order to meet these needs, this section designs the ANNI model and its integration model. The integration of ANNI is mainly to improve the ubiquitous performance of the network. The biggest feature of the ANNI integration model is that the integrated network composed of the model for network integration is still an ANNI model, but with multiple hidden layers added. Therefore, the ANNI rule extraction algorithm is still adapted to the ANNI integrated neural network.
Definition 3 (ANNI structural model): Let the input layer vector be x={x0, x1, …, xn}, the activation vector h(M+l) of the output layer (where M is the number of hidden layers, in particular, h(1) is the first hidden layer activation vector, and all other hidden layers are δ functions, except that the activation function of the first hidden layer is a step function), wijm is the weight between the (m-1)th hidden layer and the m-th hidden layer (i is the i-th neuron, the offset value of the virtual neuron is 0. In particular, wij0 is the weight between the input layer and the first hidden layer), and the number of the first hidden layer node is an integer multiple of the number of input layer nodes, which is called the ANNI structural model.
As shown in Figure 2, the ANNI model has two hidden layers. It can be seen that there is no full connection between the input layer and the first hidden layer in the ANNI model, and the first hidden layer uses the step function as its activation function. In fact, it can also be seen as δ function. It can be proved that the ANNI model has the following properties: 1) The activation value of the output layer neurons is constant; 2) A virtual hyperplane parallel to the coordinate axis is a possible decision boundary; 3) The virtual hyperplane is located in 0≤k≤q.
Figure 2 ANNI model with two hidden layers
Definition 4 (Expected output error): Suppose p is the training sample, i is the index of the output neuron, si is the target value, and E is the minimum error sum expected. Then E can be expressed as:
(2)
It should be noted that when learning using the BP algorithm, the formula (2) cannot be derived under the step function, so the compromise method is adopted, that is, in the training phase, the function is used to solve the gradient of each layer (except the first layer) [31]. Using the step function to calculate the sum of squared errors in the first layer associated with the stopping criteria. Find the weight change of the square of the error associated with the gradient, as shown in Eq. (2):
(3)
where i is the number of repeated indexes; η and μ are the learning and momentum factors; Flat is the flat point elimination parameter.
Definition 5 (Hyperplane correlation): Suppose ρ is the correlation of hyperplanar Lijm, and Ri1, i2, …, in is a hyperrectangle in (q+l)n cells constructed by n input neurons of order q, ij=0, 1, …, q. Then P can be described as:
(4)
where Pi denotes a data sample; D is an ANNI network; K is an ANNI network output layer classification representation function [32].
3.2 Rule extraction algorithm based on ANNI
3.2.1 Description of ANNI rule extraction algorithm
The description of the ANNI rule extraction algorithm is shown in Figure 3.
3.2.2 Analysis of ANNI-REA
1) Complexity analysis: Suppose the input number is n, the sample tree is p, the activation function step number is q, and the first hidden layer has m neurons (m is an integer multiple of n), and then all virtual hyperplane numbers are mnq.
Therefore, the algorithm execution efficiency of each step is analyzed [33, 34]. As shown in Table 1, it can be seen that the complexity of the ANNI rule extraction algorithm is a polynomial of m, n, p, q, where n and p are related to the size of the classification problem, and m is related to network size. In other words, the efficiency of algorithm execution in the worst case is the sample, attribute, and step three-factor polynomial time. Therefore, the algorithm can be used for appropriately scaled ANNI networks and classification problems of appropriate scale.
2) Fidelity analysis: ANNI can obtain 100% rule fidelity for the sample data in the rule extraction stage. It is not guaranteed for unknown data, but it can be extracted by the unknown network by the trained network. In other words, when classifying unknown data, if there are one or more rules containing the sample, and the classification given by a certain network, it will stop; if there are one or more rules with different classification results, then it will add the sample as a new sample to the training set to re-examine the rule [35, 36]. The shortcoming of this method is that the rule extraction algorithm needs to be executed multiple times, and the time overhead is large, but for the driver’s driving behavior learning and driving safety, a slightly longer time is also acceptable.
Figure 3 ANNI rule extraction algorithm description
Table 1 Algorithm complexity analysis
3.2.3 ANNI example
It can be known from the nature of ANNI that the input layer and the first hidden layer determine the predecessor of the rule, and the inequality of the inserted predecessor is “>” or “<”, and the weight between the input neuron and the first hidden layer neuron is assigned for 1 or -1. Besides, the closed value in the inequality is converted into the weight of the virtual neuron, as shown in Figure 4.
Figure 4 Rule antecedent virtual neuron weights
Each neuron of the second hidden layer represents the “And” of the rule predecessor. The output layer represents the “Or” of the rule predecessor. The “And” of the predecessor sets the weight as w between the first hidden layer and the second hidden layer. The weight of bias neurons is (-y+0.5) times that of w, and y is the number of regular predecessors. The “Or” of the predecessor is to set the weight value between the second hidden layer and the output layer as w. The corresponding offset value is set as -w/2. For example, in the rule extraction of overtaking and changing lanes, the hidden layer is used to train feature quantities related to relative speed, accelerating distance, time headway, signal light, driving habits, lane change, etc. As shown in Figures 5 and 6.
Figure 5 “And” rule extraction
Figure 6 “Or” rule extraction
Figure 7 shows an ANNI network that represents a rule set with two rules (Rule 1, Rule 2) and a default rule (Rule 3). The activation value of the first hidden layer is binary, and the step (t) function is the activation function of the first hidden layer, so it can be encoded in KBANN mode.
Take the two classification problems of a two-dimensional space as an example, the ANNI rule extraction process is further illustrated. First, the virtual hyperplane (such as Figure 8(a)) is drawn, and a decision tree (such as Figure 8(b)) is constructed according to the definition of the hyperplane correlation (such as Figure 8(c)), in which a decision rule is formed from the roots to the leaves. Therefore, a total of 6 initial rules can be extracted for Rule 1-Rule 6 (as shown in column 2 of Table 2). Then, these rules are further analyzed, and the rules are tailored and deleted. Obviously, “x2
Figure 7 Extraction rules from ANNI network
Similarly, two antecedents from Rule 1 can be deleted. In addition, according to the maximum coverage cutting standard rule, the number of samples covered by Rule 3 increases to 14 when the precursor “x2
3.3 Rule extraction algorithm based on ANNI integration
The description of the ANNI integrated extraction algorithm is shown in Figure 9.
The ANNI integration rule extraction process is exemplified in Figure 10. The first hidden layer determines the precise position of the hyperplane, and other subsequent hidden layers use 0 or 1 as the weight of the hyperplane in a specific area, and then construct a decision tree according to the hyperplane list and hyperplane correlation, using rule extraction [37-39]. The algorithm extracts the rules in the ANNI network integration. Table 3 lists the four rules ultimately extracted from Figure 10.
Figure 8 List of rules of initial rule and reduction
Table 2 List of rules of initial rule and reduction
Figure 9 ANNI integrated extraction algorithm description
Figure 10 ANNI integrated rule extraction process
Table 3 List of rules by ANNI integrated
4 Experimental results and analysis
4.1 Comparison on UCI standard dataset
Two rule extraction algorithms, ANNI Integration and Decision Tree Integration (C4.5-REA), were validated and compared on the UCI standard dataset (shown in Table 4). The number of individual networks used for ANNI network integration and decision tree integration is 25. The experiment obtained average prediction accuracy based on 10 cross-validation. For each cross-validation experiment, the data ratios of training set, verification set and test set were 8:1:1. Figure 11 shows the average prediction accuracy of the two models, and Figure 12 shows the average prediction accuracy of the extraction rules. From this we can easily find that the ANNI integration achieves better algorithm performance than the C4.5-REA on most UCI standard data sets.
Table 4 Standard UCI data set
Figure 11 Comparison of average prediction accuracy
Figure 12 Average prediction accuracy of obtained rules
4.2 Comparison on data set of self-built driving behavior
The experiment uses the three driving behaviors of free driving, vehicle following and lane change, which are respectively recorded as X1, X2 and X3. The ANNI integration and neuro-rule were used to test and compare the performance of classification and rule extraction on the self-built driving behavior data set.
Table 5 shows the parameter settings. Among the experimental parameters, implications of influence actors of driving behavior are as follows. Besides, X represents driving decision (DD); X1 represents free driving (FD); X2 represents vehicle following (VF); X21 represents constant speed following (CF); X2 represents acceleration following (AF); X23 represents deceleration following (DF);X3 is lane changing (LC).
Table 5 Experimental parameters
The data set contains a total of 2683 samples, of which 1836 vehicles follow the car, 565 lane changes, 282 free driving, and 1541 samples are randomly selected as the training set. Some training samples are shown in Table 6, and the rest samples are used as the test set for testing.
Figure 13 shows the comparison results of the accuracy of the extraction of the two methods of ANNI-REA and Neuro-Rule. It can be seen that ANNI integration has a slight advantage over Neuro-Rule in general. In particular, Neuro-Rule has higher rule extraction accuracy on the training set, while ANNI-REA is 0.4% more accurate than Neuro-rule. In addition, comparing rule extraction accuracy of training set and test set, obviously ANNI-REA is more stable.
The ANNI integration and REFNE algorithm are compared with the rule extraction precision experiment (as shown in Figure 14). Figure 14 shows that the average accuracy of ANNI-REA and REFNE extraction rules is difficult to distinguish, including their own accuracy(AC) and their rule accuracy(RA) , but its complexity(CO) is reduced by about 10% and its fidelity(FI) rises to 99.1%.
Table 7 lists the driving behavior rules extracted using ANNI integration, such as rule 27: If Y5=“-1/0” and Y8=“-1” and Y6=“1” and Y4=“-1” and Y3 = “-1” then X=“X2”, which means that when the vehicle is within 8 s (Y5=“-1/0”), the traffic sign is identifiable (Y8=“-1”), the distance from the preceding vehicle is greater than the safety distance (Y4=“-1”), the difference between the preceding vehicle and the vehicle is less than the threshold (Y3=“-1”), but the distance from the front vehicle is within the dangerous distance (Y6=“1”), the driving strategy adopted is “follow the car” (X=“X2”), which is consistent with the driving style adopted by ordinary drivers. It is not difficult to find out the driving behavior rules extracted in Table 7 one by one. It is easy to be stimulated and influenced by multiple factors such as driver, vehicle, road and environment during driving, and the driving strategy is formed under these factors. It also confirms the correctness of cognitive driving behavior description.
In addition, it is not difficult to find some behavioral rules in accordance with human driving experience from Table 7. Rule 19 and Rule 20 show that If Y5=“-1/0” and Y8=“1” and Y1=“1” and Y6=“1” and Y4=“1” and Y3=“0/1” then X=“X23”; If Y5=“-1/0” and Y8=“-1” and Y6=“-1” and Y7=-1 then X=“X3”. It can be seen that when the current vehicle faces the same lane and the front lane is not safe (Y4=“1”), the driver generally chooses to slow down or change lanes (X=“X23” or X=“X3”). However, when the warning indicates the stimulus (Y8=“-1”), the driver usually prefers the lane change operation (X=“X3”). This is the same as that in actual driving, the driver will consider changing the speed of the car, which will lead to traffic delays and increasing fuel consumption.
4.3 Experimental verification of driving behavior rules
The 8 km two-lane experimental road was built by PreScan, and the road ratio was 1:1, which simulates the actual road driving situation. Add various vehicles, warning signs, sensors and other modules in the traffic scene. The target vehicle is Audi_A8_1. The driver can manipulate the simulator to control the target vehicle and complete the driving behavior decision in the specific traffic scene jointly constructed by PreScan-Simulink X21, X22, X23.
Table 6 Part of training sample
Figure 13 Rule extraction accuracy comparison of ANNI-REA and Neuro-Rule
Figure 14 Rule extraction accuracy comparison of ANNI-REA and REFNE
Table7 Extracted driving behavior rules
Experiment 1: On the expressway, the front car is driven at the speed Velocity 1 in the upper left picture, and the initial speed is set to 60 km/h, and the initial distance between the target vehicle and the preceding vehicle is 100 m. The driver manipulates the simulator to follow the car. The corresponding driving decisions at different times are shown in the upper right “unapplied 1” in Figure 15. The results corresponding to the target vehicles with the extracted driving rules are shown in Figure 15, upper right “applied 1”.
Experiment 2: On the urban expressway, the front car is driven at the speed Velocity 2 at the bottom left, setting the initial speed to 30 km/h, and the initial distance between the target vehicle and the preceding vehicle is 70 m. The target vehicle is added with ACC control, and the desired vehicle speed is set to 70 km/h. The corresponding driving decisions at different times are shown in the lower right “unapplied 2” in Figure 15. On this basis, the result of the target vehicle using extracted driving rules is shown in the lower right “applied 2” of Figure 15.
It can be seen from Figure 15 that in different road scenes, the driving behavior decisions of the target vehicle using extracted driving behavior rules compared with that adding ACC control or driver free to drive during different periods have a slight delay difference, and the others are basically consistent, which verifies the correctness of the extracted driving rules. During driving process, driving behavior is easily stimulated and influenced by various factors such as vehicle, road and environment, and driving strategies are formed under these factors. Therefore, the validity of the extraction based on cognitive driving behavior rules is also verified.
5 Conclusions
1) In this paper, a driving behavior rule extraction algorithm based on ANNI and its integration is proposed. In the process, the cognitive driving behavior model is established, and then the cognitive driving behavior is described and analyzed according to cognitive learning theory. Such works combines the advantages of human- specific cognition in avoiding dangerous situations and ANNI in solving complex and fuzzy engineering problems. The extracted driving rules reveal the inner law of driving habit, and solve the difficulty in formalizing driving experience and being fully utilized.
Figure 15 Front car speed and driving rules verification results:
2) This paper is a new attempt to propose driving behavior extraction method based on cognitive behavior analysis. The implication of the proposed method to the literature is twofold. Firstly, compared with common C4.5-REA, Neuro-Rule and REFNE, relevant experimental results show that the method is about 0.4% more accurate and about 10% less complex. Secondly, designed simulation experiments where extracted driving rules are applied to targeting vehicle in two different traffic situations, verify the correctness of the extracted rules and the effectiveness of rule extraction based on cognitive driving behavior.
3) Except for ANNI integration algorithm based on cognitive behavior analysis proposed in the paper, other algorithms can be considered on this basis to improve the accuracy of driving extraction. In addition, more other particular rules under more complex traffic environment should be researched and discussed in the future to extend the driving rule database.
References
[1] World Health Organization. Road traffic injuries. [EB/OL]. [2018-02-19]. http://www.who.int/zh/news-room/fact-sheets/ detail/road-traffic-injuries.
[2] China Highway. Analysis on the main causes and characteristics of road traffic accidents in China. [EB/OL]. [2018-04-23]. http://www.chinahighway.com/news/2018/116 9754.php.
[3] ARBABZADEH N, JAFARI M. A data-driven approach for driving safety risk prediction using driver behavior and roadway information data [J]. IEEE Transactions on Intelligent Transportation Systems, 2017, 19(2): 17541832. DOI: 10.1109/TITS.2017.2700869.
[4] PAVLOU D, YANNIS G. Road safety behavior of drivers with neurological diseases affecting cognitive functions: An interdisciplinary structural equation model analysis approach [J]. Advances in Transportation Studies, 2018, 44: 133-150. DOI: 10.1016/j.cbpra.2018.03.002.
[5] WANG J, YU X P, LIU Q, YANG Z. Research on key technologies of intelligent transportation based on image recognition and anti-fatigue driving [J]. J Image Video Proc, 2019: 33. DOI: 10.1186/ s13640-018-0403-6.
[6] GAO X, GAO L, DONG G. Research on intelligent driving behavior based on cognitive science and scene simulation [C]// International Conference on Intelligence Science & Information Engineering. IEEE, 2011. DOI: 10.1109/ ISIE.2011.15.
[7] AYOUBI S, LIMAM N, SALAHUDDIN M A, SHAHRIAR N, BOUTABA R, ESTRADA-SOLANO F, RENDON O M C. Machine learning for cognitive network management [J]. IEEE Communications Magazine, 2018, 56(1): 158-165. DOI: 10.1109/MCOM.2018.1700560.
[8] HANSEN P, OXOBY R J. An earned presence: Studying the effect of multi-task improvisation systems on cognitive and learning capacity [J]. Connection Science, 2017, 29(1): 77-93. DOI: 10.1080/09540091.2016.1277692.
[9] BEETZ M, BUSS M, WOLLHERR D. Cognitive technical systems-what is the role of artificial intelligence? [C]// Annual Conference on Artificial Intelligence. Berlin, Heidelberg: Springer, 2007. DOI: 10.1007/978-3-540- 74565-5_3.
[10] BENDERIUS O, BERGER C, LUNDGREN V M. The best rated human–machine interface design for autonomous vehicles in the 2016 grand cooperative driving challenge [J]. IEEE Transactions on Intelligent Transportation Systems, 2018, 19(4): 1302-1307. DOI: 10.1109/TITS.2017.2749970.
[11] POPA A, MCDOWELL J J. Behavioral variability in an evolutionary theory of behavior dynamics [J]. Journal of the Experimental Analysis of Behavior, 2016, 105(2): 270-290. DOI: 10.1002/jeab.199.
[12] HUI J, HANNA E, GATTO C L, PAGE T L, BHUVA B, BROADIE K. A fully automated Drosophila, olfactory classical conditioning and testing system for behavioral learning and memory assessment [J]. Journal of Neuroscience Methods, 2016, 261(3): 62-74. DOI: 10.1016/j.jneumeth.2015.11.030.
[13] VAA T. Modelling driver behaviour on basis of emotions and feelings: Intelligent transport systems and behavioural adaptations [M]// Modelling Driver Behaviour in Automotive Environments. London: Springer, 2007. DOI: 10.1007/978-1-84628-618-6_12.
[14] SALVUCCI D D. Modeling driver behavior in a cognitive architecture [J]. Human Factors, 2006, 48(2): 362-380. DOI: 10.1518/001872006777724417.
[15] GRAEFE A C. The role of cognition in simulated driving behavior in young adults with attention-deficit/hyperactivity disorder [D]. Philadelphia: Drexel University, 2015.
[16] WANG Xiao-yuan, WANG Lei. Cooperative simulation model for driving behavior [J]. Journal of Southwest Jiaotong University, 2007, 42(2): 238-242. (in Chinese)
[17] HAMDAR S H, MAHMASSANI H S, TREIBER M. From behavioral psychology to acceleration modeling: Calibration, validation, and exploration of drivers’ cognitive and safety parameters in a risk-taking environment [J]. Transportation Research Part B: Methodological, 2015, 78: 32-53. DOI: 10.1016/j.trb.2015.03.011.
[18] QU Ting. Driver behavior modeling based on stochastic model predictive control [D]. Changchun: Jilin University, 2015. DOI: 10.3969/j.issn.0258-2724.2007.02.021. (in Chinese)
[19] WATANABE T, SASAKI Y, SHIBATA K, KAWATO M. Advances in fMRI Real-Time Neurofeedback [J]. Trends in Cognitive Sciences, 2017, 21(12): 14-15. DOI: 10.1016/j.tics.2017. 09.010.
[20] HOVHANNISYAN G. Humanistic cognitive science [J]. Humanistic Psychologist, 2018. DOI: 10.1037/hum0000074.
[21] DELBROEK T, VERMEYLEN W, SPILDOOREN J. The effect of cognitive-motor dual task training with the biorescue force platform on cognition, balance and dual task performance in institutionalized older adults: A randomized controlled trial [J]. Journal of Physical Therapy Science, 2017, 29(7): 1137-1143. DOI: 10.1589/jpts.29.1137.
[22] HRAMOV A E, FROLOV N S, MAKSIMENKO V A, et al. Artificial neural network detects human uncertainty [J]. Chaos, 2018, 28(3): 1-10. DOI: 10.1063/1.5002892.
[23] HANSEN L K, SALAMON P. Neural network ensembles [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 1990, 12(10): 993-1001. DOI: 10.1109/34. 58871.
[24] PEZESHKI Z, MAZINANI S M. Comparison of artificial neural networks, fuzzy logic and neuro fuzzy for predicting optimization of building thermal consumption: A survey [J]. Artificial Intelligence Review, 2018(819): 1-31. DOI: 10.1007/s10462-018-9630-6.
[25] LIANG Jun, SHA Zhi-qiang, CHEN Long. Dynamic neural network-based integrated learning algorithm for driver behavior [J]. Journal of Transportation Systems Engineering and Information Technology, 2012, 12(2): 34-40. DOI: 10.3969/j.issn.1009-6744.2012.02.006. (in Chinese)
[26] LIN Ming-long. Incremental learning based on neural networks ensemble [D]. Hefei: University of Science and Technology of China, 2012. DOI: 10.7666/d.y2126157. (in Chinese)
[27] QU W N, GE Y, XIONG Y X, CARCIOFO R, ZHAO W G, ZHANG K. The relationship between mind wandering and dangerous driving behavior among Chinese drivers [J]. Safety Science, 2015, 78: 41-48. DOI: 10.1016/j.ssci.2015.04.016
[28] ZHANG L, LIU P, GULLA J A. Dynamic attention- integrated neural network for session-based news recommendation [J]. Machine Learning, 2019(6): 1-25. DOI: 10.1007/s10994-018-05777-9.
[29] LEDESMA R D, MONTES S A, POO F M, LOPEZ-RAMON M F. Individual differences in driver inattention: The attention-related driving errors scale [J]. Traffic Injury Prevention, 2010, 11(2): 142-150. DOI: 10.1080/153895 80903497139.
[30] KANG X J, NAMGUNG M, FUJIWARA A, KIM W C, WANG W J. Analysis of vehicle maneuverability and driving characteristics on a curved road condition [J]. KSCE Journal of Civil Engineering, 2018, 23(6): 420-432. DOI: 10.1007/s12205- 018-1803-y.
[31] MA C, TSANG I W, SHEN F M, LIU C C. Error correcting input and output hashing [J]. IEEE Transactions on Cybernetics, 2019, 49(3): 781-791. DOI: 10.1109/TCYB.2017.2785621.
[32] BADNAVA B, MOZAYANI N. A new potential-based reward shaping for reinforcement learning agent [J]. arXiv, 2019: 1902.06239.
[33] JANSEN T, ZARGES C. Analysis of evolutionary algorithms: From computational complexity analysis to algorithm engineering [C]// Foundations of Genetic Algorithms, International Workshop, Foga 2011. Schwarzenberg, Austria. 2011: 1-14. DOI: 10.1145/1967654.1967656.
[34] HAESER G, LIU H, YE Y. Optimality condition and complexity analysis for linearly-constrained optimization without differentiability on the boundary [J]. Mathematical Programming, 2018(5): 1-37. DOI: 10.1007/s10107-018- 1290-4
[35] DEBATTISTA K, BASHFORD-ROGERS T, HARVEY C, WATERFIELD B, CHALMERS A. Subjective evaluation of high-fidelity virtual environments for driving simulations [J]. IEEE Transactions on Human-Machine Systems, 2018, 48(1): 30-40. DOI: 10.1109/THMS.2017.2762632.
[36] AZADEH A, BONAB N A, SALEHI V, ZARRIN M. A unique algorithm for the assessment and improvement of job satisfaction by resilience engineering: Hazardous labs [J]. International Journal of Industrial Ergonomics, 2015, 49(9): 68-77. DOI: 10.1016/j.ergon.2015.06.002.
[37] GHOSH J, DEUSER L, BECK S D. A neural network based hybrid system for detection, characterization, and classification of short-duration oceanic signals [J]. IEEE Journal of Oceanic Engineering, 2002, 17(4): 351-363. DOI: 10.1109/48.180304.
[38] ASADI S, SHAHRABI J. Complexity-based parallel rule induction for multiclass classification [J]. Information Sciences, 2017, 380(2): 53-73. DOI: 10.1016/j.ins.2016. 10.047.
[39] FERRETTI E, TAMARGO L H, GARCIA A J, ERRECALDE M L, SIMARI G R. An approach to decision making based on dynamic argumentation systems [J]. Artificial Intelligence, 2017, 242(C): 107-131. DOI: 10.1016/j.artint.2016.10.004.
(Edited by HE Yun-bin)
中文导读
基于认知型行为分析的驾驶规则抽取
摘要:为了充分利用驾驶员在道路交通信息感知、交互和车辆控制过程中的长期驾驶经验,提出了一种基于ANNI及其集成的驾驶行为规则抽取算法。首先,基于认知学习理论,建立认知型驾驶行为模型,并进行描述和分析。然后,基于ANNI的模型和规则抽取算法(ANNI-REA)来解释驾驶行为及离散性。在不离散连续输入变量的情况下,规则具有较高的保真度和安全性。UCI标准数据集和自建驾驶行为数据集上的实验结果表明,该方法比常见的C4.5-REA、神经规则和REFNE方法准确度提高了约0.4%,简化了约 10%。进一步地,仿真实验验证了所提取的驾驶规则的正确性和基于认知驾驶行为规则提取的有效性。总的来说,抽取的多条驾驶规则充分体现了驾驶综合认知序列活动的执行机制,这对于车联网下混合交通流的交通以及未来无人驾驶的研究具有重要意义。
关键词:认知型驾驶行为; 驾驶规则抽取;认知理论;集成算法
Foundation item: Project(2017YFB0102503) supported by the National Key Research and Development Program of China; Projects(U1664258, 51875255, 61601203) supported by the National Natural Science Foundation of China; Projects(DZXX-048, 2018-TD-GDZB-022) supported by the Jiangsu Province’s Six Talent Peak, China; Project(18KJA580002) supported by Major Natural Science Research Project of Higher Learning in Jiangsu Province, China
Received date: 2018-10-09; Accepted date: 2019-08-23
Corresponding author: LIANG Jun, PhD, Professor; Tel: +86-13952924831; E-mail: liangjun@ujs.edu.cn; ORCID: 0000-0001-8942- 0391