Driving skill classification in curve driving scenes using machine learning

To understand the driving skill of a given driver, her continuous process, which is characterized as the Perception–Decision–Action cycle, has to be taken into account. A wide variety of research on driver modeling has been reported [5‐18]. However, very few studies have been explicitly characterized driving skill in terms of driving action [14‐20] and applied this using machine learning methods.

Table 1 shows a detailed comparison of machine learning methods used in prior studies of classification of driving maneuver skill. Tang et al. [17] and Zhang et al. [16] have dealt with analyzing driving data in the lateral direction (steering angle only) by setting the speed of the driving simulator to a constant. In contrast, Chandrasiri et al. [15, 18] have dealt with a more natural case: motions in both longitudinal and lateral directions of the driving simulator/real vehicle are considered with multiple attributes (specifically, speed is not fixed, and the driver can control the vehicle in both longitudinal and lateral directions). In making this possible, a novel approach has been used, converting time-indexed data into distance-indexed data. The details of this are discussed in Sect. 3.1.

Most of the target driving scenes involve curves or double lane changes, although one includes a narrow straight lane [15]. Each scene is intended to elicit a display of maneuvering skill. Another novel aspect of our research is that it presents scenes with curves of different radiuses and angular changes. Analysis of results from these scenes can determine which require more maneuvering skill.

In using machine learning to characterize driving behavior, there are two main steps.

Various features and classifiers have been used. Tang et al. used wavelet transforms (WTs) and WTs with a discrete Fourier transform (WT + DFT), applying this to the steering angle and using a neural network and support vector machine (SVM) for learning. Zhang et al. have applied DFT to the steering angle and used a neural network, SVM, and decision trees for learning [14, 16]. Chandrasiri et al. have used DFT with probabilistic neural networks and SVM in their work [15].

In our work, we have experimented with a combination of multiple features that cover both lateral and longitudinal controls as discriminative features. For dealing with multivariate time-series data, we first apply the well-known principal component analysis (PCA) to reduce the dimensionality of the data. In doing so, we also aim to overcome the problem of data sparseness.

In this work, we use the k-nearest neighbor (k-NN) classifier, which is simple to implement, and SVM, which is known to be powerful. The driving scenarios and data segmentation methods that are used in our experiment distinguish this study from Refs. [16, 17]. In those studies, a full set of data from single runs in driving simulator scenes involving a double lane change and a lane change in a curve (a single curve of common radius and angular change is used for all trials) is used for learning and prediction. In our work, we simulate typical curves to elicit typical driving maneuvers. Specifically, we use full curve data from six different curves of distinct radii and angular changes. By choosing driving scenes in this way, we can analyze differences in driving skill that are demonstrated in response to the difficulty of the driving scene, which can vary according to the radius and angular change of the curve. We also analyzed driving data after dividing the curves into small segments. This is a challenging task because the data from a small segment of a curve must be applied to analyzing the skill level of the driver. Despite the difficulty in finding it, knowing the driving skill displayed from a small driving segment would benefit driver support and infotainment applications.

As an example application, information overload could be prevented by taking instantaneous driver skill into account. Even for a highly skilled driver, real-time fluctuation in the classification of driving skill can be expected for reasons such as fatigue and driving conditions.

We collected data on driving in curve-containing scenes using the driving simulator shown in Fig. 2. The collected data include steering angle, speed, longitudinal acceleration, lateral acceleration, yaw rate, accelerator control, brake control, lateral displacement, longitudinal displacement, accelerator control speed, and brake control speed. These data were sampled at 60 Hz. Sixteen adult participants (men and women 20–40 years of age) took part in the experiment. As shown in Fig. 3, the driving course consisted of 6 curves (radius = 50, 100, or 200 m; angular change = 45° or 90°). In the simulator, each driver ran up to 12 trials in an urban scene with 2 trials per scene. In total, 960 (16 × 10 × 6) runs of driving data from curves were used for the analysis. Traffic signs indicating a speed limit of 60 km/h are set at 100 m and 50 m before the starting point of each curve (Fig. 4). This controls the speed at which drivers enter the curves. We asked drivers to drive in the left lane and not to cross the yellow center line.

The flow of data analysis in this paper is depicted in Fig. 6. We used driving data acquired by a driving simulator as explained in Sect. 3. Ten driving data attributes analyzed here are those previously mentioned (steering angle, speed, longitudinal acceleration, lateral acceleration, yaw rate, accelerator control, brake control, lateral displacement, accelerator control speed, and brake control speed), which covers both longitudinal and lateral control of the vehicle. In the pre-processing step, the collected driving data are converted from time-index data into distance-indexed data. This enables comparison among different runs of drivers at the same point on a curve (see Sect. 3.1 for details). The segmentation method is depicted in Fig. 7. As shown in figure, segments 1 and 5 are included in the data analysis along with the curve Sects. 2–4. In those end segments, the vehicle enters and leaves the curve. Therefore, in the full curve case, analysis of driving skill via machine learning considers transition effects from straight segments to curves and vice versa. In our future experiments, the design of the road could be smoother, such as using clothoids.

To merge the data from different sensors into a single feature vector, some normalization method should be employed. For our analysis, we use the standardized score (z-scores) normalization at the points in the distance-indexed data. Feature extraction using PCA and driving skill classification using k-NN and SVM are explained in Sects. 4.1 and 4.2, respectively (Fig. 8).

In this section, we show and discuss some of the results from the analysis of driving skill that was performed using the experimental method discussed in Sect. 4. Driving course contained six curves with different radiuses and intersection angles as explained in Sect. 3. In this paper, we report classification accuracy for six full curves and curve-segmented cases.

Figure 10 shows how combinations of different attributes from feature extraction and two driving skill classifiers contribute to the accuracy of driving skill classification. In previous research, except for those using our proposed methods (this work and [18]), which are shown in Table 1, only steering angle data were used. We wanted to compare how combinations of multiple parameters work, and this can be seen in Fig. 10. We selected a combination of parameters that can be easily acquired in a real driving scenario using currently available sensor data, such as driver’s control data (steering angle, acceleration pedal control, and brake pedal control) and basic parameters that account for both longitudinal and lateral movement of the vehicle (steering angle, speed, longitudinal acceleration, lateral acceleration, and yaw rate). Additionally, we also included lateral displacement, which accurately characterizes lateral movement of the vehicle in a driving simulator scene.

In terms of the classifier, SVM performs better than k-NN in all cases, which is not surprising. However, since we have extracted low-dimensional discriminative features via PCA, k-NN does not offer much worse performance.

In analysis of the full curve data, using all ten attributes gives the highest accuracy rates among tested cases. Overall, low/average-skilled and high-skilled drivers were classified with 95.7 % accuracy using SVM.

In analysis of the segmented curves, an average accuracy of 89 % was obtained, using the same classifier. These results show the feasibility of using only a segment of a curve for classification, in spite of the small data window. Figure 11 illustrates how driving skill recognition accuracy depends on segment number. Using SVM for classification, the accuracy varies from 85 % to 91 % in different segments of the curves, with the maximum accuracy of 91 % recorded at the fourth segment. This segment is near the curve exit, which may require higher use of skill, particularly if errors had accumulated in earlier segments.

An example of PCA-based projection of the driving data onto a low number of dimensions (here, 2) is depicted in Fig. 12 for the segmented curve case. Clustering of drivers by skill level is visible, even in two dimensions. However, we cannot directly compare the accuracies of driving skill through the qualitative analysis of data for each segment in Fig. 11 because we use several different dimensionalities, with the components to use determined from the cumulative contribution ratio of the principal components.

Table 1 shows a comparison of machine learning methods for classifying driving skill at maneuvers. In the analysis by the proposed method, both lateral and longitudinal control of a vehicle are considered. In contrast, only lateral control is taken into account in the works by Zang et al. [16] and Tang [17].

Our work deals with a simple scenario of driving on a curve, and a highest accuracy in classifying driving skill compares favorably with results from prior research.

Driving skill classification results for curves of different radii and angular change [intersection angle (IA)] are depicted in Figs. 13 and 14. There is a tendency toward higher accuracy when the radius is smaller. This could be because the difficulty of driving increases as the radius becomes smaller. For curves of the same radius but different angular change, a change by 90° has higher difficulty than a change by 45° due to the longer distance of the curve in the first case. The above tendency can be seen explicitly in the analysis of the segmented curves. With the full curves, differences between the 50-m and 200-m radii are marked. From these results, we can select curves with a smaller radius and a larger angular change to increase the accuracy of driving skill classification.

In our experiment, all curves were to the left for convenience in building the simulator scenario. This may affect both the potential learning effects and the anticipation of control input needed. We intend to address these issues in future research.

However, this work and [18] are attempts to analyze driving skill on multiple curves and multiple segments and to compare driving skill in curves of different radii and angular change (as shown in Figs. 13, 14) using machine learning methods.