Keywords

1 Introduction

Driver fatigue has already become a leading factor contributing to traffic crashes around the world [48]. In China, fatigued driving caused 887 (9.26%) of all highway crashes in 2011, resulting in 520 (8.1%) deaths and over RMB 37 million (10.82%) property losses [43]. To decrease the occurrences of traffic crashes and promote road safety, studying drivers’ decision on driving under fatigue condition is urgent. Despite the substantial work in the domain of transportation science by both the industry and academia, there are few studies in applying case-based reasoning approach to modelling the risk of driver fatigue.

Case-based reasoning (CBR) has been used in both cognitive science and artificial intelligence [32], it makes use of the most similar previous cases to solve new problems [27, 30, 32]. Since the early 1980s, CBR has been successfully applied in various fields such as legal reasoning [5, 6, 42, 45], planning [14, 15, 25, 28, 34], E-commerce [18, 41], medical diagnosis [4, 9, 19, 40], incident management [24, 49], and risk analysis [23, 35]. CBR is the best fit for modelling fatigued driving because of its ability to simultaneously process a large number of highly interrelated variables to arrive at a decision [13, 20]. Moreover, CBR models have significant merits as compared to statistical ones (e.g., the widely employed matched case-control logistic regression [2, 31]) and other artificial intelligence models (e.g., neural network) regarding the comprehensibility of output [13].

The novelty of the research work presented lies in the following aspects. First of all, this work explores the potential for using CBR in dealing with driver fatigue management. Secondly, this research examines a highly comprehensive set of risk factors that relate to the driver fatigue. A database of 16,459 real traffic crashes reported from 21 cities in Guangdong province, China from 2006 to 2010 is used, and a total of 50 variables in 21 categories are examined in this study. Compared with the logistic regression model that uses the same variables, our proposed CBR model had better performance in modelling fatigued driving on AUC measures in all simulations. The outcome of our proposed CBR system can help targeting the group of drivers that may intend to drive under fatigue condition, which constitutes the crucial part of driver safety management in an intelligent transportation system.

This paper is organised as follows. Section 2 introduces previous relevant work from the literature. Section 3 describes the methodology. Section 4 presents the case study and empirical evaluation of our approach. Section 5 concludes the study.

2 Related Work

In recent years, there had been an increase in uptake of case-based reasoning (CBR) concepts in the research area of traffic control and management. For instance, Jagannathan et al. [22] developed a CBR prediction system that is capable of differentiating between the accident and non-accident cases. Mounce et al. [36] designed a CBR system to help selection of signal timing plans. Sadek et al. [39] developed a prototype CBR routing system. Li and Zhao [33] applied CBR to intersection control. Kofod-Petersen et al. [26] presented a prototype implementation of a CBR system that predicts traffic flow and calculates signal plans for urban intersections. CBR methods have also been used to model the vehicle control behaviour (i.e., steer, throttle, and brake) of teen drivers [37]. Different from the previous ones, this work aims to explore the potential for using CBR in dealing with driver fatigue management.

The risk of driver fatigue is related to a combination of situational and individual factors. The increased risk may result from a mix of biological, personal, road and environmental related factors [29]. Previous studies on detecting driver’s state of fatigue have used various biological and personal indicators, such as visual indicators (e.g., face images and eye state [10, 12, 17]), physiological sensor signals (e.g., EEG [46] and ECG [7]) and driving states (e.g., accelerate, brake, shift and steer [21]). However, none of them (including the CBR approach of identifying the drivers’ stress state [7]) considered road and environmental features that are crucial towards drivers’ intention to drive under fatigue condition. By using CBR in this research, the risk of driver fatigue is modelled based on past similar circumstances considering not only personal but also road and environmental characteristics. Moreover, most previous studies verify their models through experiments with driving simulators, whereas we validate ours by using the reliable official source of traffic crash data.

3 System Design

Following the work of [27, 33], we propose a case-based system of fatigued driving that consists of three main components: representation module, retrieve module, and reuse module. Figure 1 shows a schematic overview of the system architecture. The proposed system receives traffic information, through the representation module, a target case \(C_t\) and a case-base \(\mathcal {C}\) are constructed based on such information. The retrieve module retrieves the most similar previous case from the case-base, and reuse module further manipulates the retrieved cases for the best outcome of the target case.

Fig. 1.
figure 1

Schematic diagram showing steps involved in modelling fatigued driving

Full size image

3.1 Case Representation

A case in the system represents a previous traffic situation [33], and it is composed of a case number, a case description, and a case outcome. Formally, a case can be represented as follows:

  • Case \(C = \langle N, D, O\rangle \), where,

    • N is the case number;

    • D is the case description;

    • O is the case outcome (i.e., fatigued-driving or non-fatigued-driving);

\(\mathcal {C}\) is used to denote the case base that contains all possible cases.

The description of a case depicts the situation when the case appears [26, 33]. Four categories of attributes are needed to describe the case. The first category contains the demographic information of the driver. The second one describes the vehicle information. The third category consists of road situation factors. The fourth category contains some environmental factors such as the weather situation. Formally, we have:

  • Case description \(D = \langle DD, VI, R, E \rangle \), where,

    • DD is the demographic information of the driver such as gender;

    • VI is the vehicle information such as vehicle type;

    • R is the road situation such as road type;

    • E is the environmental factors such as bad weather.

For example, a simple case (No. 0) that a crash occurred on the expressway at about 8 am on Saturday was caused mainly because the fatigued driving of the male motorcyclist would be represented as \(\langle 0, \langle \{\text {male}\}, \{\text {motorcycle}\}, \{\text {expressway}\}, \{\text {moring, week-end} \} \rangle , \text {fatigued-driving}\rangle \).

3.2 Case Retrieval and Reuse

The crucial part of the retrieval module is the similarity measure, which calculates the similarity between the target case and the cases in the case-base [22]. For each target case, the similarity between the target case and the cases in the case-base is measured by calculating the Euclidean distance (the most widely used distance metric in CBR) between them. Given a target case t and a case c in the case-base, the similarity between both cases is:

$$\begin{aligned} S(t,c)=\sqrt{\sum _{a}(D_{t,a}-D_{c,a})^2} \end{aligned}$$

where \(D_{t,a}\) is the normalised value of attribute a \(\in DD, VI, R, E\) in the target case t and \(D_{c,a}\) is the normalised value of attribute a in the case c from the case base. Note that all the attributes are of equal importance in this paper for simplicity reasons.

With the above similarity measure, the most similar cases will be retrieved by using the K-nearest neighbour (KNN) algorithm [11], where K refers to the number of neighbours. The KNN algorithm, widely used in classification problems to assign objects to classes, matches each attribute in the target case to its corresponding attribute in the retrieved case [22]. The predicted outcome of the target case is then obtained by the majority vote of its K nearest neighbours, i.e., the most common outcome of the K retrieved cases are reused and assigned to the target case directly.

Table 1. Case structure
Full size table

4 Case Study: Based on the Traffic Crash Data in China

In this study, we use the traffic crash data for the period 2006–2010 in Guangdong Province, China. These data are extracted from the Traffic Management Sector-Specific Incident Case Data Report, the Road Traffic Accident Database of China’s Public Security Department. They are the only officially available source of traffic crash data in China. Data are recorded and reported by the traffic police on-scene who conducted assessments and provided feedback immediately to the headquarters of the Traffic Management Department. These reports include characteristics of drivers, vehicle features, road conditions, the time of crashes, the environmental context for each crash and the cause of crashes such as traffic violations like driving under fatigue condition [47]. To experiment with the proposed approach, we set up a case-based system in which each of the samples in our datasets was a case. Table 1 shows the case structure used in the experiment where relevant binary attributes are selected based on [48]. According to working time patterns and peoples’ lifestyles in China, we classify five groups of time.

4.1 Evaluation Settings

To evaluate our proposed case-based system of drivers’ risk of fatigued driving, the data set was partitioned into a training set and a test set. To be aligned with the literature (e.g., [3]), we varied the percentage of the training and test data to study the possible variations of performance based on different partitions. In particular, ratios 10:90, 30:70, 50:50, 70:30 and 90:10 were used as the portions of training and test data. Each case in the test set was consecutively made the target case with the cases in the training set constituting the case-base. For each portion of training and test data, KNN algorithms with different numbers of K were run. Ties may occur in a binary classification problem if \(K > 1\), to avoid ties, the most similar case (\(K=1\)), the three most similar cases (\(K=3\)), and the five most similar cases (\(K=5\)) were retrieved in the current study. In comparison, the widely employed logistic regression was also conducted for each partition. The dependent variable indicated the occurrence of driving under fatigue condition in a crash, while the independent variables under consideration were the same as the CBR setting.

The performance of a computational intelligent system can be measured in many different ways. Typically, the models’ predictive performances are measured regarding overall accuracy, e.g., the percentage of fatigued-driving and non-fatigued-driving classified accurately in the current research [16]. However, other measures like the true positive rate (TPR, also called recall or sensitivity) tends to be the key concern when infrequent events (as fatigued driving in the current study) are to be predicted [44]. Complementary to TPR that measures the proportion of positives that are correctly identified as such, another primary evaluator is the false positive rate (FPR, also known as fall-out or false alarm ratio) that calculates the ratio of negatives wrongly categorised as positive. Based on pairs of TPR and FPR, a receiver operating characteristic (ROC) curve could be obtained to compare the performance of different models. We calculate the area under the ROC curve based on the work of Cantor and Kattan [8], and use IBM SPSS Statistics 23 to perform all the statistical analysis.

Fig. 2.
figure 2

Accuracy of the different models

Full size image

4.2 Results and Discussion

The available data set contained 16,459 traffic crashes, of which 384 (2.33%) were fatigue-related. Among all the fatigue-related crashes, 99% of drivers were male, 88% of them had more than two years of driving experience, and 72.4% came from urban areas. Drivers that were involved in fatigue-related crashes included self-employed workers (23.7%), migrant workers (18.5%), farmers (17%), clerks (8%), and employees from other professions (32.8%). Concerning vehicle information, trucks and motorcycles constituted 56% and 16.7%, respectively. For road situation, 65.9% of these fatigue-related crashes occurred in separate lanes, and 38% occurred in expressway. Considering environmental features, 51.8% of the fatigue-related crashes occurred during 00:00–06:59, and 38% took place at daytime. Seasonal distribution of the fatigue-related crashes: 32.3%, 26.3%, 24.7%, and 16.7% occurred in the summer, winter, spring, and fall, respectively.

Table 2. Evaluation results (TPR for true positive rate and FPR for false positive rate)
Full size table

Figure 2 shows the comparison of the overall accuracy for all simulations. In overall, all simulations received very high accuracy in modelling fatigued driving, i.e., more than 96%. Logistic regression (LR) had a better performance in comparison to the proposed CBR method (no matter what the value of K) for most training/test partitions. When 70% or above of the dataset was used for training, the CBR method with K = 3 achieved slightly higher accuracy than LR.

Table 2 summarises the results of true positive rate (TPR) and false positive rate (FPR) from all simulations. In all scenarios, we observed that the FPR is nearly zero, indicating that both LR and the proposed CBR method have excellent performance in correctly identifying non-fatigued-driving cases. However, LR received zero TPR for all the training/test partitions, which reflects that LR is relatively biased towards the majority class in our dataset (i.e., the non-fatigued-driving case). Such bias may be caused by the imbalance situation of our dataset (i.e., only 2.33% cases were fatigue-related). Recall that TPR reflects the hits of fatigued driving reality, in such circumstances, although LR got very high accuracy and very low FPR, it is meaningless [44]. In contrast, our proposed CBR models with K = 1 or K = 3 had shown their merits by receiving non-zero TPR in all the training/test partitions.

Figure 3 illustrates the trade-off between the TPR and the FPR of each simulation. Our proposed CBR models (no matter what the value of K) had better performance than LR in modelling fatigued driving on AUC measures. Simulation of K = 1 outperformed the rest as the percentage ratio of the training and test data were varied.

To sum up, we can see that the proposed CBR model with K = 1 and that with K = 3 have their strengths in modelling fatigued driving. To further differentiate them, we calculated the precision values in various partitions of training and test data. Precision measures the probability that a driver with a positive screening test indeed conducted fatigued driving. As shown in Fig. 4, when 10% of the data set were used for training, the K = 1 model had higher precision value, whereas the K = 3 model had better performance when more than 10% were used for training. Further empirical validation is required to study the optimal value of K for modelling fatigued driving.

Fig. 3.
figure 3

AUC of the different models

Full size image
Fig. 4.
figure 4

Precision of the K = 1 and K = 3 models

Full size image

4.3 Further Implications

Fatigue-related crashes are one of the major threats to road safety worldwide. The important implication of our research is: if driving under fatigue condition could be reduced or controlled successfully through the early detection, the occurrences of a crash would be reduced accordingly. When designing road traffic interventions to decrease incidences, it is well established that a change in the driver’s attitude is of utmost importance. The outcome of our proposed CBR model can help targeting the group of drivers that may have an intention to drive under fatigue condition. In particular, our proposed CBR model can highlight the important features of such group of drivers in the aspects of various personal, vehicle, road, and environmental conditions. Specific countermeasures by integrating these highlighted features into a driver fatigue detection component of the intelligent transportation systems will be effective in improving traffic safety [38]. The results obtained in this study may likely be generalised to other provinces in China because the crash data in Guangdong Province and features considered are rather representative and comprehensive. The fact that Guangdong has a high percentage of residents migrating from other provinces, and is having one of the largest numbers of vehicles among all provinces in China reinforces the generalisability [48]. Furthermore, the findings of this study can also contribute as a reference to future road safety research for other countries.

5 Limitations and Future Work

A live case-based reasoning system has a “4 REs” cycle (i.e., retrieve, reuse, revise, and retain) [1]. Since our proposed case-based system is not operating on live data, there is no component designed in the current study for the standard revise and retain process. Our proposed system could be extended to deal with live data by adding a new part of revise and retain in the future.

The advantage of using case-based reasoning in modelling fatigued driving has been demonstrated through comparing with logistic regression. In getting higher values of recall, precision and AUC, an important consideration is case attribute selection and weighting. The attributes we have chosen are based on previous work. However, it is likely that fatigued driving also occurs due to other conditions that have not yet been studied. Moreover, the contribution of each attribute to the similarity between the target case and the cases in case-base could differ as well. Thus, it is worth testing out different attributes and weighted assignments for attributes considered.

The empirical experiments show that the CBR system under development is capable of differentiating between fatigued-driving and non-fatigued-driving cases to some extent based on certain personal and environmental traffic characteristics. The evaluation of our proposed method is not exhaustive and could be expanded in several directions, for instance, to compare with other artificial intelligence techniques (e.g., neural network). It would also be worth evaluating whether or not our proposed approach could be fruitfully applied to other risky behaviours rather than the kind we have considered.