Effects of feature selection on lane-change maneuver recognition: an analysis of naturalistic driving data

Purpose – Feature selection is crucial for machine learning to recognize lane-change (LC) maneuver as there exist a large number of feature candidates. Blindly using feature could take up large storage and excessive computation time, while insufficient feature selection would cause poor performance. Selecting high contributive features to classify LC and lane-keep behavior is effective for maneuver recognition. This paper aims to propose a feature selection method from a statistical view based on an analysis from naturalistic driving data. Design/methodology/approach – In total, 1,375 LC cases are analyzed. To comprehensively select features, the authors extract the feature candidates from both time and frequency domains with various LC scenarios segmented by an occupancy schedule grid. Then the effect size (Cohen’s d) and p-value of every feature are computed to assess their contribution for each scenario. Findings – It has been found that the common lateral features, e.g. yaw rate, lateral acceleration and time-to-lane crossing, are not strong features for recognition of LC maneuver as empirical knowledge. Finally, cross-validation tests are conducted to evaluate model performance using metrics of receiver operating characteristic. Experimental results show that the selected features can achieve better recognition performance than using all the features without purification. Originality/value – In this paper, the authors investigate the contributions of each feature from the perspective of statistics based on big naturalistic driving data. The aim is to comprehensively figure out different types of features in LC maneuvers and select the most contributive features over various LC scenarios.


Introduction
Lane-change (LC) accidents are accounting for about 4-10 per cent of all crashes (Barr and Najm, 2001) and 1.5 per cent of all motor vehicle fatalities in the USA (N.H. T. S. Administration, 2015).With the development of advanced driver assistance systems, functions such as lane departure warning (LDW) and lane-change assist (LCA) hit the market to help avoid LC related accidents (Visvikis et al., 2008).One of the key problems is how to correctly recognize driver's LC maneuver in advance.When an improper LC maneuver is occurring, the driver assistance systems should either give warnings or assist this person aborting the maneuver.
Supervised learning is popularly used for recognizing LC maneuvers.This method raise the challenges of how to select the most contributive or efficient features.Longitudinal features [e.g.time to collision, TTC, (Liebner et al., 2013) (Peng et al., 2015), longitudinal acceleration] and lateral features [e.g.steering angle (Xu et al., 2012), yaw rate (Sivaraman and Trivedi, 2014;Doshi et al., 2011), lateral acceleration (Boubezoul et al., 2009;Kasper et al., 2012)] have been used, with the assumption that they are strong enough for LC maneuver recognition by either intuition or empirical knowledge; however, this assumption is still hanging on and yet comprehensively studied.
In general, LC maneuver can be either discretionary or mandatory.A mandatory LC will occur when a driver must leave a lane due to a lane drop or bypass a blockage, etc.A discretionary lane change occurs when a driver prefers a more efficient adjacent lane (J2944, 2013), for example, passing a slow-moving leading vehicle to maintain the current speed (Lee et al., 2004).So it would require different weighted features to recognize LC maneuver with discretionary and mandatory LC cases.Leonhardt and Wanielik evaluated the effects of various features in different LC driving scenarios (Leonhardt and Wanielik, 2017) and showed that even for the same feature, the weight of overtaking a slow vehicle and merging is different.Thus, feature selection process should also take LC scenarios into account.
In this paper, we propose a feature selection method for predicting driver LC behavior.Our aim is to comprehensively figure out different types of features in LC maneuvers and select the most contributive features over various LC scenarios.The main contribution of our work can be summarized as follows: presenting a feature selection method from the perspective of statistics to investigate the statistical significance of each feature based on big naturalistic driving data; both time-domain and frequency-domain features are considered to fill in gaps in existing works on feature selection; and taking different driving scenarios into consideration in the feature extraction procedure to comprehensively evaluate the extracted features.
The remainder of this paper is organized as follow.Section 2 reviews the related work of feature selection and LC maneuver recognition.Section 3 describes how to model the contextual traffic in each LC scenario.Section 4 details data processing and feature extraction.Section 5 shows the results of feature selection and model performance.Section 6 makes conclusion and discussion.

Related works
2.1 Lane-Change maneuver recognition Machine learning techniques, such as support vector machines (SVM) (Mandalia and Salvucci, 2005;Kumar et al., 2013), Naive Bayes (NB), Decision Tree (DT), k-nearest neighbor (KNN) (Lethaus et al., 2013), artificial neural networks (ANN) (Peng et al., 2015) and Bayesian Networks (BN) (Kasper et al., 2012;Li et al., 2016;Weidl et al., 2018), have been implemented to recognize driver LC maneuvers based on a well-trained classifier using labeled datasets.Then new data are fed to the classifier to determine the classification of either LC or LK maneuver.In this way, data being classified as LC means the driver at the moment is prone to make LC maneuver, otherwise not.
Although most papers have made comparison work to show their effectiveness, it is achieved by using the less contributive features in their proposed model.To overcome this bias, we evaluate model performance using the identical featuresthe comprehensively selected featuresand then give them a relatively objective rate to evaluate their contribution for maneuver recognition.

Feature selection
The goal of feature selection is to reduce the dimension of training datasets by removing redundant information.In general, feature selection methods can be grouped into filter and wrapper methods.Filter methods analyze the intrinsic properties of data, ranking and selecting features without involving learning algorithms.On the contrary, the wrapper method with learning algorithms would get involved to score a given subset of features (Guyon et al., 2008).For wrapper methods, the ranking of features can vary from model to model.Here, we shall find the intrinsic properties of feature candidates related to LC and select the most contributive features rather than ranking and selecting features for a specific learning approach.Therefore, the filter method was selected in this paper.
For LC maneuver recognition, the data collected from sensors are time series data, and the properties of the features in the time domain are the most frequently extracted (Liebner et al., 2013) (Kasper et al., 2012).On the other hand, frequency-domain features have already been used to recognize driver state, for instance, the power spectrum features via wavelet transform were selected for Belief networks (Hajinoroozi et al., 2015;Chen et al., 2015).In other areas of time series recognition such as speech recognition (Thomas et al., 2008) and anomaly detection (Zhang et al., 2008), frequency-domain features play an important role.
In this paper, we consider the properties of the features in both time domain and frequency domain to select the most contributive features for LC maneuver recognition.

Modeling contextual traffic
To capture the most contributive features in each LC scenario, we first model the contextual traffic of the ego vehicle.A potential field diagram (Woo et al., 2016) composed of bubbles with different dynamic sizes was used to describe the dynamic relationship between the ego vehicle and its surrounding vehicles.However, it is not an intuitive way for driving situation analysis.Leonhardt and Wanielik (2017) developed a probabilistic situation assessment model to judge the safety state of the ego vehicle with its surrounding vehicles; however, it only works as behavior recognition model with a single input.
To easily describe the relationship between the ego vehicle and its surroundings, one of the most popular approaches is to segment the surrounding traffic into cell grids.The occupancy state of each cell is represented by a binary value, i.e. occupied or empty (Kasper et al., 2012;Do et al., 2017).We model contextual traffic also based on the cell grid method which is detailed in the next section.

Lane-Change scenario modeling
In Do et al. (2017), nine cells and 32 cases (25) were considered in the driving contextual traffic.But the authors did not give the specific boundary of the cells.Kasper and Weidl modeled the cell by carrying the speed-dependent information when a cell will be occupied or will become free (Kasper et al., 2012).But they assumed that the vehicle can move unobstructed toward certain cell, which cannot be satisfied in some situation where a car wants to overtake the ego car.In this paper, we model the cell grid by considering the dynamic relationship between the ego car and the surrounding cars.A three-cell grid is used to model contextual traffic where the ego vehicle executes the LC maneuver, for both left and right LC with eight cases.
As limited by our on-board sensors, which can only detect the traffic in front of the ego vehicle, the traffic situation on back of the ego vehicle is not considered.Despite such limitation, our method of modeling contextual traffic can be extended to more cell grids which can include the traffic on back of the ego vehicle.Here we only model the contextual traffic in front of the ego vehicle as is depicted in Figure 1(a).
We adopt the theory presented in Karim et al. (2013) to define the middle cell (cell m ) and theory in Kesting and Treiber (2013) to define the left (cell l )/right cell (cell r ).The dynamic length of each cell is s Ã 1 , s Ã 2 , s Ã 3 , as shown in Figure 1(a).The length of cell m is defined by a mean safe time gap (MSTG) in Karim et al. (2013) as: where BT EV and BT OV are the brake time of the ego vehicle and object vehicle 1, respectively, RT is the driver's perceptionreaction time and for certain vehicle, the BT is calculated by an empirical equation: where v is the vehicle speed and thus: where _ R is the range rate between the ego vehicle and object vehicle 1.So, the dynamic length of s Ã 1 can be written as: where v is the longitudinal speed of the ego vehicle.
We define cell l and cell r based on the Intelligent Driver Model (IDM) (Kesting and Treiber, 2013).Here, the safe distance is derived from the leading vehicle, driving at a desired speed, or preferring accelerations to be within a comfortable range.Additionally, kinematic aspects are taken into account, such as the quadratic relation between braking distance and speed.First, on the left and right lane, desired distances on the left (s Ã l ) and right (s Ã r ) lane are defined respectively as: where s 0 is the minimum (bumper-to-bumper) gap, T is the safe time gap, a Ã and b Ã are acceleration and comfortable deceleration.R l , _ R l and R r , _ R r are the range and range rate the ego vehicle with object Vehicle 2 and Vehicle 3 in Figure 1 imply the intelligent braking strategy for LLC and RLC cases.Second, based on the desired distance on the left (s Ã l ) and right (s Ã r ) lane, the dynamic safety distance, namely, the length of s Ã 2 and s Ã 3 , can be written as: where az is the longitudinal acceleration of the ego vehicle, Da is the LC threshold and a bias represents the asymmetric property of LLC and RLC.
All the values of the parameters in equations ( 1) and ( 5)-( 8) are listed in Table I (Kesting and Treiber, 2013), and the occupancy states of cells can be given as (Figure 2): Depending on the occupancy state of cell grid, eight scenarios (four scenarios for LLC) can be generated, as depicted in Figure 3: LLC Scenario 0_0: When the ego vehicle makes LLC, there are no object vehicles on both cell m and cell l ; LLC Scenario 0_1: When the ego vehicle makes LLC, there is no object vehicle on cell l but cell m is occupied; LLC Scenario 1_0: When the ego vehicle makes LLC, there is no object vehicle on cell m but cell l is occupied; LLC Scenario 1_1: When the ego vehicle makes LLC, both cell m and cell l are occupied; RLC Scenario 0_0: When the ego vehicle makes RLC, there are no object vehicles on both cell m and cell r ; RLC Scenario 0_1: When the ego vehicle makes LLC, there is no object vehicle on cell m but cell r is occupied; RLC Scenario 1_0: When the ego vehicle makes LLC, there is no object vehicle on cell r but cell m is occupied; and RLC Scenario 1_1: When the ego vehicle makes LLC, both cell m and cell r are occupied.
Here, the name of the LC scenarios such as Scenario 0_1 and Scenario 1_0 is in accordance with the binary states of the occupancy cells illustrated in Figure 3.

Naturalistic driving data
The naturalistic driving data that used in this paper are from the project of the Safety Pilot Model Deployment (SPMD).
The on-road test includes multi-modal traffic, hosting approximately 3,000 vehicles equipped with vehicle-to-vehicle (V2V) communication devices (Henclewood et al., 2014).The data sets we used were extracted from 20 vehicles, driving in the field test including 75 miles of roadway, see Figure 3. Roads that marked as yellow are the route SPMD vehicle driving.Drivers voluntarily joined in SPMD project.They drove the SPMD vehicle completely based on their own driving styles with no restriction on their driving behaviors.Each SPMD vehicle was equipped with data acquisition systems (DAS) such as CAN and GPS and vision system such as Mobileye.All the signals coming from different DAS were time-synchronized and were recorded at 10 Hz.
Finally, 1,375 LC cases (761 LLC and 614 RLC) were analyzed.The distribution of the LC cases with respect to the corresponding LC scenarios in Figure 3 can be seen in Table II.We can see that for LLC, most of the cases took place in LLC Scenario 0_0 (365 cases) and LLC Scenario 0_1 (354 cases).For RLC, the dominating cases are RLC Scenario 0_0 (371 cases) and RLC Scenario 1_0 (214 cases).This result implies that when the driver want to execute left/right LC, he/she tends to wait until the destination lane being empty (cell l /cell r is unoccupied).TTC with the object vehicle in front on the current lane (TTC t ) at time t: where R and _ R [in Figure 1 (b)] are the range and the range rate between the front edge of the ego vehicle and rear edge of the closest object vehicle in the same traveling path as the ego vehicle, respectively.Here, what needs to be mentioned is that TTC is only calculated for the LC case when Cell m = 1, because Cell m = 0 means there is no vehicle in the cell.
TLC at time t (TLC t ): where dx is lateral distance between the front wheel and the lane boundary of the ego vehicle and vx is the lateral speed.
In case that _ R and vx are equal to zero, equations ( 12) and ( 13) approach infinity, we use the inverse of TTC À1 t and TLC À1 t instead.

Time-Window features
Vehicle on-board signals are time series, using time-window (TW) for feature extraction is effective to capture the information during the past few seconds (Thissen et al., 2003;Salfner and Malek, 2007).In the case of LC recognition, different length of TW between 1 and 5 s are selected for feature extraction (Mandalia and Salvucci, 2005).To capture the properties of time series, statistical variables (mean, standard deviation, maximum, minimum and median) are calculated within each TW (Li et al., 2015) as is described in Table III, i.e. feature number 6-80.The number of the top right corner of the feature is the length of TW, so '5' in mean yaw 5 t means 5 s length of TW and '4' in mean yaw 4 t represents 4 s length of TW, see feature # 6 and # 7 as examples.

Frequency-domain features
Frequency-domain feature extraction has already been used in anomaly detection (Chen et al., 2015;Chandola et al., 2009).Fast Fourier transform (FFT) was used to transform timedomain signals into frequency-domain (Heckbert, 1995).The maximum value of FFT coefficients within TW is a good indicator to represent the property of frequency signals (Mörchen, 2003).The description of the frequency-domain features are listed in Table III, with feature number 81-95.

Labeling LC datasets
To evaluate extracted features, both LC and LK datasets should be labeled.Take LLC for example, as shown in Figure 4, the ego vehicle (blue) intends to overtake the slow vehicle (red) by left lane change.The moment that the left wheel of the ego car just crosses the central dotted line is marked as the initial LC time t 0 .Based on the study in Salvucci and Liu (2002), normally drivers tend to start LC maneuver approximately 5 s before actual LC.Thus in this paper, time series between t 0 and 5 s before are labeled as LC behavior.To ensure LK data sets are separation of LC data sets, LK behavior are labeled between 10 and 15 s prior to t 0 .It is the same way to label RLC data sets.

Feature evaluation
In the view of statistics, the p-value is commonly used to test whether there is statistical significance between two groups.In our case, if there is statistical significance between LC data sets and LK data sets, the extracted features are probably good indicators to classify LC and LK maneuvers.However, only using p-value to evaluate significance is insufficient (Sullivan and Feinn, 2012).The effect size, such as Cohen (1988), is also used as an important evaluation metric (Cohen, 1990): where: M 1 = mean of the first group data; M 2 = mean of the second group data; S 1 = standard deviation of the first group data; and S 2 = standard deviation of the second group data.
For each LC maneuver, we label LC and LK data sets and calculate both Cohen'd and p-value for each feature.Then for all the LC cases, we average the Cohen'd and p-values to get the mean for each feature in each scenario.

Models used for feature evaluation
To test if the selected features have advantages for machine learning techniques over all features, the SVM, NB, DT and KNN are chosen to evaluate classification performance.We then built the above learning-based models using the Statistics and Machine Learning Toolbox [1].Here, the SVM model was set with a Gaussian kernel function, and NB with Kernel smoothing density estimation method, DT with the default setting and KNN using empirical prior with k = 1.The datasets used for training the machine learning models are the same, which are labeled by the method presented in Section 4.3.

Results and analysis
6.1 Analysis from effect size and p value All the evaluation results (Cohen' d and p-value) for each feature can be found in Table AI.A p-value smaller than 0.05  can be regarded as having statistical significance and a Cohen's d value larger than 0.8 has large effect level (Cohen, 1992).By following this two criterion, we mark each feature with Cohen' d larger than 0.8 and p-value smaller than 0.05 as red in Table AI.The red-marked features have great influence on the corresponding LC maneuver (LLC or RLC), and thus can be selected as strong features for LC maneuver recognition.
Overall, based on the features marked as red, we find the following: Although some features (p < 0.05) have shown statistical significance (marked as blue), they have only medium or small effect size (Cohen' d < 0.8).This result also coincides with that only using p-value to evaluate statistical significance is not enough (Sullivan and Feinn, 2012); Original features of yawRate t (#1), az t (#2) and ax t (#3) and compound feature TLC À1 t (#5) are not strong features for LLC case with no items marked as red.For RLC, only az t and TLC À1 t in RLC Scenario 0_1 can be regarded as strong features.This implies that the common empirical knowledge of using these features is not that much convincing.
We mentioned that TTC À1 t is only calculated when the front cell of the ego vehicle is occupied by an object vehicle (Cell m = 1).TTC À1 t is marked as a strong feature in the LLC case, which demonstrates that the potential of rear-end collision does influence drivers' LC decision.In many research, a hypothesisif the driver follows a leading vehicle which is too slow, he/she would probably maneuver a LC to overtake the slow leading vehiclewas made.This analysis from naturalistic driving data proves that this hypothesis is reasonable.
Features #56-#60, which refer to mean_ax, are the least important features for LC maneuver recognition, with no item marked as red.
To analyze the TW features (#6-#95), we take the marked strong features in LLC Scenario 0_0 and LLC Scenario 0_1 for instance.Here, we segment the Table vertically with 5 features in a group, e.g.features.
Features #6-#10 are related to the same feature mean_yaw but with different TW from 5 to 1 s, and so on.The detailed illustration can is shown in Table AI, where the features with the largest Cohen'd and the smallest p-value (marked with 's' and 't', respectively) demonstrate that they have the strongest effect on LC.From these peak and valley values we find that features with the largest Cohen' d are also likely to have the smallest p-values, except for Feature #31 and #32 in LLC Scenario 0_0.We select the final features for each scenario based on the marked peak and valley features, and for the special case like feature #31 and #32, the features with large Cohen'd (e.g.#31) are selected.

Final selected features for each LC scenario
Based on the marked features and results, the final selected features in each LC scenario are listed in Table IV.It can be found that different LC scenarios have different features sets.The number of selected features from all 95 features for each LC scenario ranges from 8 to 16.There is no feature eligible for all LC scenarios.Only using original features and compound features (#2, #4, #5) are far less enough, because no such kinds of features have been selected at all in LLC Scenario 0_0 and LLC Scenario 1_0, RLC Scenario 0_0, and RLC Scenario 1_0.
Although original features related to vehicle's lateral movement (yawRate t (#1), az t (#2) and ax t (#3)) are not contributive as expected, their corresponding TW features have shown large effect sizes.This implies that the property of the original features within certain TW carries more important information regarding to LC maneuvers.In addition, frequency-domain features do have a contribution as expected, with nearly at least one feature falls into the strong feature set (the exception is LLC Scenario 0_1 with no frequency-domain feature eligible).In what follows, we will use these selected features to train models and evaluate their performance.

Performance evaluation using the selected features
To test if the selected features can really improve model performance, we compare the classification results of different models trained with the selected features in Table IV (termed as `Selected') and all features (termed as 'All Features').Data sets used for training are the same as what we used for calculating Cohen'd and p-value.To guarantee that the training data and testing data are disjoint, a cross-validation (CV) method is used to test the performance of these models.The datasets are evenly divided into ten folds.Nine folds are used to train the models and the remaining is used to test the models.
The receiver operating characteristic (ROC) curve is used to access model performance as it has been widely used as a tool to illustrate the performance of binary classifiers by considering the true positive rate (TPR) and false positive rate (FPR) over different thresholds settings (Lethaus et al., 2013;Morris et al., 2011).TPR and FPR are defined as follows: where TP, TN, FP, FN are true positives, true negatives, false positives and false negatives, respectively.A simple way to compare different classifiers is to calculate the value of area under curve (AUC) with a value ranging from 0 to 1.A larger AUC value indicates better performance.All ROC curves are illustrated in Figures 6 and 7.The corresponding AUC values are listed in Table V. Figure 6 demonstrates the classification performance of each classification model in different LLC scenarios.The blue lines are the ROC curves of using selected features for training while the red lines are using all features.Figure 7 represents the same content as Figure 6 but for RLC scenarios (Figure 5).
In Table V, a comparison is made between all features and selected features for each classification model in each LC scenario.We denote ';' as the performance deterioration of using selected features compared with using all features.The improvement in percentage is also shown in Table V.Finally, we find the following results: KNN performs very good in all LC scenarios with AUC values greater than 0.95, so does DT except for the performance in LLC Scenario 0_0 with all features (AUC = 0.82).With this exception, DT can significantly improve the classification performance from 0.82 to 0.98 (an increase of 19.5 per cent) by using the selected features.For DT, using selected features only shows tiny deterioration (1.0 per cent) in LLC Scenario 1_1 and RLC Scenario 1_1.
For SVM, it can greatly improve the classification performance (performance increase between 4.2 and 13.6 per cent) by using the selected features, compared with using all features, but only show declination in LLC Scenario 0_1.
NB represents different pictures.Using selected features cannot improve model performance, compared with using all features (no improvement in all LLC scenarios), except for in RLC Scenario 0_1 and RLC Scenario 1_1 (Figure 6).

Conclusion and future work
In this paper, a statistics-based feature selection method for recognition of LC maneuver is proposed using naturalistic driving data from the time domain and the frequency domain.The extracted features include original features collected from on-board sensors and compound features like TTC, TLC as well as timewindow features.Totally 95 features are extracted as candidate features.We found that for different LC scenarios, the final selected features are different.There is no feature being sufficient for all the LC scenarios.In addition, features refer to vehicle lateral movement which are frequently being used as features regarding to LC, such as yaw rate (yawRate t , #1), lateral acceleration (ax t , #3) as well as TLC (TTC À1 t , #5), do not show statistical significance (except for TTC À1 t in RLC Scenario 0_1).This counter-empirical result makes it more worthwhile to do feature selection work rather than just based on empirical knowledge.
Finally, the classification performance by using the final selected features in each LC scenario is compared to that using all features.The result shows that except for the relatively poor performance of Naive Bayes, the performance of SVM and Decision Tree, as well as KNN, can be improved from different levels by using the selected features in most LC scenarios compared with using all features.Summarily, the high performance achieved by the classification models using all features (95 features) is at the expense of computation time and taking up large storage.
Considering the fact that using the selected features (nearly only ten features) to train the models can still achieve the same performance or even have significant improvement.In future work, a series of on-road experiment will be conducted to recognize LC maneuver to evaluate the recognizing performance in real-time scenarios.Appendix.All the effect sizes regarding LLC and RLC for all extracted features are listed in Table AI.(Brookhuis et al., 2001) and may release people from the boring and distressing task of driving (Stanton and Young, 2005), enhancing traffic safety and fuel economy (Fagnant and Kockelman, 2015).The five levels of autonomous driving as defined by the Society of Automotive Engineers are widely accepted in the automotive industry (SAE, 2016).Nowadays, research on automated driving has mainly focused on the practical use of partial (Level 2; L2) and conditional (Level 3; L3) automated driving systems.In the partial automated driving system (L2) that executes both lateral and longitudinal controls on behalf of the human driver, the driver is required to continuously monitor the road environment and status of the automated system.The conditional automated driving system (L3) executes not only vehicle dynamic controls but also monitors surroundings.This means that the human driver does not need to monitor the driving environment during L3 driving.However, some circumstances (e.g.sensor failures, misunderstanding marked lanes) can make L2 and L3 systems reach their limits; then, the system issues a Request to Intervene (RtI).At this time, the driver should take manual control of the vehicle immediately and appropriately.
Although driving automation may promise such benefits, several human driver-related issues have been raised with regard to automated driving systems.Some researchers reported that automated driving can cause drivers to fall asleep compared to manual driving.Also, Naujoks et al. (2016) pointed out that drivers are more likely to pay attention to nondriving tasks while driving with adaptive cruise control (ACC) and lane keeping assist (LKA).These matters may cause such as out-of-the-loop (OoTL) performance problem (Endsley and Kiris, 1995) where drivers away from control loops cannot adequately respond to system errors.This is partly due to lack of an operator's situation awareness.Merat and Jamson (2009) investigated the awareness and comprehension of driver's peripheral traffic during manual driving and automated driving, and as a result, under automated driving, the driver's response to the dangerous events will be delayed.In addition, there are several studies that show automated driving causes driver drowsiness (Jamson et al., 2013;De Winter et al., 2014), which will lead to a lack of situation awareness of driver.
Many studies have also been conducted on driver response to an RtI under different circumstances.Naujoks et al. (2014) focused on the modality of RtI and reported that a driver responded to an RtI sooner if visual-auditory RtI was used instead of just visual RtI.Zeeb et al. (2016) pointed out that the quality of the takeover can be attributed to driver behavior before the RtI.Louw et al. (2017) proposed that humanmachine interface (HMI) design should focus on helping the driver take evasive action quickly rather than emphasizing the time to takeover.Most studies have focused on the time span and driver behavior required for an RtI.However, there are other issues to consider with regard to RtI.Current vehicles are equipped with assorted features.For instance, car navigation systems are installed in most cars and will tell the driver the current location and destination.However, the driver may be deeply involved in non-driving task or be absentminded, a result based on past studies.Such a driver may not consciously look at the car navigation system and confirm the current location.This situation is not a huge problem as long as an automated driving system is performed without any abnormality.What can be a problem is when a tight RtI is issued.If the time margin given to the driver is small, or the tasks required for the drive is large, such as changing the driving lane after an RtI, the driver may not be able to immediately grasp the situation.Gold et al. (2013) revealed that it takes more than 5 s for the driver to respond to an Rtl owing to (the lack of or decline in) situation awareness.Thus, it is important to investigate the vehicle design such that the driver can grasp the current situation correctly in a limited time.Also, HMI will play a major role when and RtI is issued.Louw et al. (2017) proposed that HMI design should focus on helping the driver take evasive action quickly rather than emphasizing the time to takeover.In L3 automated driving, drivers do not need to do drive tasks and have a supervisory control of the system and environment.As a result, it is possible that the driver will not be able to identify his/her location and does not know which direction to go in a limited time.In fact, in our past research, we confirmed a branch failed case despite presenting the correct branch lane after an RtI to the driver and the driver was able to grasp the steering wheel (NEDO, 2018).It is necessary to further investigate whether this result was caused by automated driving.Thus, the aim of this study is to investigate the influence of automated driving on the driver's localization.In this study, driver's localization means the ability to grasp what position on the route a car is currently travelling to the destination.As HMI design is as an important concern for elderly driving performance (Freund et al., 2005;Körber et al., 2016), this driving simulator study is targeted to elderly drivers.

Participants
Seventeen elderly participants (men, 12; women, 5), of age 66 to 82 years (M = 72.0,SD = 4.0), participated in the experiment.The participants were recruited via a local human resource agency.All of them hold a valid driver's license and drove daily.The experiment was conducted after obtaining approval from the University of Tsukuba Research Ethics Committee.

Experimental design
In the experiment, one factor in the "existence or nonexistence of automated driving" was taken into consideration.It was a betweensubject design with two levels of autonomous operation (AO) and manual operation (MO).The participants were randomly assigned to each group.The average age and composition of each group were as follows (AO: M = 73.6,men = 6, women = 3; MO: M = 71.3,men = 6, women = 2).

Driving simulator
As presented in Figure 1, the experiment was conducted with a motion-based driving simulator (Honda Motor Co., Ltd.).The simulator has three LCD monitors that display the left and right mirror image and rearview mirror image.The front field-of-view was approximately 120°horizontally and 45°vertically.Two LCD monitors for displaying the HMI (automated status) and Internet TV (AbemaTV: https://abema.tv/) were mounted on the dashboard (Figure 2).We used the TV programs of AbemaTV with their permission.The simulator provided vehicle traveling sounds.The sounds were not too high to enable the drivers to hear the TV and an RtI request.To prevent motion sickness, the motion function was not activated.

Automated driving system
The automated driving system can be activated by pressing a button.The system maintained the vehicle's speed at 70 km/h and kept the vehicle at the center of the driving lane.Note that an automatic lane-changing function was not installed.Drivers did not have to hold the steering wheel during autonomous operation.Drivers could disengage the autonomous operation by steadily pushing the brake pedal more than 30 per cent the amount of depression or moving the steering wheel more than the absolute value of 30°.

Human-machine interface
The HMI was used to display automation status using an 8-inch LCD monitor (Figure 2).When automation was activated, a green image appeared; when deactivated, an orange image with an "off" sign appeared, as shown in Figure 3(a) and 3(b), respectively.Figure 3(c) shows that the automated driving system required the driver to resume control of the vehicle.The RtI was issued with two consecutive beeping sounds.Even if the driver did not take any action following the RtI, the system was deactivated 10 s later.

Driving task
All participants were instructed to proceed to a junction or a branch on an expressway specified in advance by an experimenter.They were also asked to drive safely.In addition, the MO participants were asked to keep the vehicle's speed at approximately 70 km/h and stay in the left lane (this is the driving lane in Japan).The AO participants were asked to watch Internet TV during automated driving and to resume control of the vehicle as needed.
The experimental driving course was based on a section of the Kita-Kanto expressway in Japan (Figure 4).It has two lanes on each side and no other car existed.The starting point was the same in all scenarios, which was the Kaminokawa IC.Since the automated driving system of this experiment operated only a single lane, and it was impossible to change lanes automatically.Thus, in the AO, the RtI request was issued approximately 200 m before the Tsuga IC exit, 250 m before Tochigi-Tsuga JCT.This means that because the car was travelling at 70 km/h, it gave the driver approximately 10 to 13 s of grace time before branching.Because the Tochigi-Tsuga JCT had a large left curve before branching, an RtI was issued in a straight section 250 m before branch considering the smooth handover.The scenarios were as follows (Figures 4 and 5): get off the expressway at Tsuga IC; proceed to the right of Tochigi-Tsuga JCT; and proceed to the left of Tochigi-Tsuga JCT.

Procedure
Each participant signed an informed consent form after receiving an explanation of the experiment's purpose and overview.A practice drive was given to each participant to get used to the simulator (approximately 5 min).Subsequently, the AO participants received an explanation about the automated system and HMI, and practiced automated driving for approximately 5 min.They also performed exercises on how to engage and disengage the automated system by pressing the brake pedal or moving the steering wheel.
Afterwards, each participant drove under three scenarios.Before each driving scenario, they received instructions from the experimenter about the route using a map.There were no navigation devices to indicate the vehicle's location and route to participants while driving.Thus, they had to remember the destination.
Figure 5 presents a detailed outline of each scenario with an actual simulator photo and illustrations.It took approximately 15 to 20 min to complete each driving scenario, and the order of drives was randomized for each participant.All participants took a 10-min break in between each drive.After all the drives, the participants were interviewed and took the Trail Making Test A/B (TMT A/B) (Tombaugh, 2004).TMT is a kind of cognitive function test used to confirm whether participants were assigned uniformly to each group.The total time of this experiment per participant was within 2 h.

Dependent measures
To investigate our hypothesis, i.e. drivers may have difficulty in appropriately driving to the instructed destination during automation compared to manual operation, the data collected from the simulator were analyzed with following dependent measures.
Results of the TMT A/B.This was used as supplemental data to confirm whether participants were equally allocated to each group.In this study, a significant decline in cognitive function Figure 4 Outline map of the driving course may cause differences in results because participants were tasked with memorizing their destination.Thus, it is necessary to ensure that the experimental results are reliable.
Number of participants who drove appropriately to the instructed destination.For each driving condition, all participants experienced three scenarios.Whether they drove to the instructed direction was counted to investigate the driver's ability to understand the location of the host vehicle.
Time elapsed to hold the steering wheel.This is the elapsed time from when the system asked the driver to take control of the vehicle until the driver takes hold of the steering wheel in the AO condition.To investigate why the driver failed to reach the destination, the elapsed time was calculated.
Point where lane change began.This indicates how far before the lane change was made from the branch point in Scenario B, in which drivers needed to change lanes to the right.This was used as an indicator to show how much the driver's maneuvering attitude against branching differs between the AO and the MO conditions.
In this experiment, we focused on the driver's ability to grasp driver's own location, and thus we will analyze the precise analysis of vehicle behavior as necessary in the future.

Results of TMT-A/B
The results of TMT-A and TMT-B are shown in Figures 6 and 7, respectively.Error bars indicate the standard deviation.The result of one-factor analysis of variance (ANOVA) on the "existence or nonexistence of automated driving" shows that the main effect was not significant [TMT-A: F(1,8) = 5.32, p > 0.05, TMT-B: F(1,8) = 2.81, p > 0.05].Therefore, it can be considered that participants were homogeneous between the two groups and the experimental results are worthy of analysis.

Driving appropriately to the instructed direction
Figure 8 shows "the proportion of drivers who proceeded to the appropriate course.Between groups, a Friedman test of success rate provided a chi-square value of 0.33, which was not significant (p > 0.1).In the MO, the ratio is almost equal in each scenario.On the other hand, approximately half of the AO participants failed only in Scenario B (to proceed to the right of the JCT), but had no errors in the other scenarios.A chi-square test of the success rate between Scenarios A-B and B-C in the AO provided a chi-square value of 2.89, which was marginally significant (p < 0.1).Although branch failures were certainly confirmed even in MO, an approximately half of AO drivers failed in Scenario B where a driver should proceed to the right lane.The driver may be able to properly deal with an RtI in the case of going straight to the left or leaving to the left lane like Scenarios A and C. On the other hand, in Scenario B where lane change was required to the right lane after an RtI, the driver may have proceeded to the left lane as they were trying to grasp the situation.Despite the relatively long grace period after an RtI, it is noteworthy that some drivers did not proceed to correct lane.This matter may have been more significant in more critical situations.Failure to do the appropriate lane change may cause traffic accidents.This finding suggests that automated driving hinders the driver's situation awareness and as a result may affect the ability to grasp self-location.

Elapsed time until holding the steering wheel
Table I shows the classification of the number of participants who held the steering wheel in the AO condition before and after the RtI.Note that data set of one participant in Scenarios A and B is missing.
Figure 9 shows the time from issuing the RtI to taking over the wheel in the AO (0 s = RtI issue point).Error bars indicate the standard deviation.As a result of a one-factor ANOVA with three levels of Scenarios A, B and C, the main effect was not significant [F(2, 21) = 0.99, p > 0.05].However, as seen in Figure 9 and Table I, half of the participants gripped the steering wheel before the RtI and prepared for the coming situation.The average time was À18.6, À6.8 and À5.0 s, from the left of the graph.Some studies have revealed that a driver took preliminary action for an RtI if the circumstances in which the RtI was issued is predictable (Merat et al., 2014); the results of this study support this conclusion.Especially in Scenario A, which involves leaving the expressway, the tendency to grasp the steering wheel earlier than in other scenarios was confirmed.On the other hand, even if the driver held the steering wheel before the RtI, a case of making a mistake in branching was confirmed.This also occurred in our previous experiments.Therefore, when suddenly assuming control from automatic operation, the driver may not distinguish the course that he/she should go to; as a result, there is a risk of creating a dangerous situation.

Point where lane change began
In this experiment, although participants were not explicitly instructed to do so, almost all drivers used the turn signal when they began to change lanes.Thus, the point where the turn signal was made was analyzed as the starting point of lane changing.Six out of eight participants in the MO and five out of nine participants in the AO condition succeeded in branching in Scenario B.Even if the driver did not take control, the automated driving system was disengaged at the beginning of the zebra zone.The zebra zone was approximately 300 m.It was also possible to change lanes within the area: however, it is highly recommended to change lanes before entering the zebra zone.
Figure 10 shows the point where the driver who successfully changed lanes began to change lanes in Scenario B (AO: six participants, MO: five participants).The result of F-test showed that the two populations are not equally distributed [F(4,5) = 84.325,p < 0.001)].The result of Welch's test showed no significant difference between the two groups [t(4.07)= À1.269,p > 0.05].However, most participants in MO seemed to start

Discussions
The aim of this study was to investigate the influence of automated driving on the driver's own localization.The driving simulator experiment involved two types of operationmanual and autonomous operationand all participants experienced three different road environments that simulated a Japanese highway.By forcing participants to choose an instructed lane before junctions or interchanges in all driving scenarios, we investigated the impact on the ability of the driver to grasp own position.
The statistical results suggest that the driver is more likely to rely on the automated driving system.Although there is no significant difference between the two groups, it seems that automated driving influenced the advancement/movement toward the correct course and the position at which the lane change was made.For instance, in Scenario B, lane changing was required after an RtI, but drivers were unable to quickly resume control owing to the expectation, which caused approximately half of failures.Also, it was suggested that it may take longer time to execute lane change under such a situation.On the other hand, in the other two scenarios where lane change was not required, there is no driver who failed to take over.These results suggest that if the driver does not know why the RtI was issued and how to take over, a failed takeover may occur more frequently.In addition, despite holding the steering wheel after an RtI, cases where the lane change was made incorrectly or where the lane change was made in the zebra zone were confirmed.This result suggests that even if the driver responds to an RtI, the driver may not always understand the situation.Further investigation in a more complex road environment will be necessary.It would be useful to explain a driver via HMI why an RtI was issued and what to do next.
There are also several limitations in this experiment.The number of participants in this experiment is not sufficient to discuss the impact/effects of the results.Thus, it is important to consider which information can be beneficial or them in terms of designing an HMI for such circumstances.Moreover, even though the drivers were asked to watch TV during automated driving, most drivers were not always looking at the TV; they sometimes monitored their surroundings.This may be because one trial for each condition lasted at most approximately 20 min.If drivers are driving for a longer time, the driver may fall into the OoTL more.However, several failures at branch occurred even in the MO group.We need to keep in mind that there could be a participant who just cannot read the letters of the sign because of a problem with the resolution of the simulator.In this experiment, to alleviate this problem, the speed of the host vehicle was limited to 70 km/h, and the participants received supplementary explanation during instructions about the route (e.g.get off at the third exit to the left/go to the fourth branch to the right).
Some drivers who failed to branch remarked that they missed the branch point unintentionally.This could be due to their driving experience or highway experience.Thus, it is necessary to analyze in detail whether the experience of manual driving will affect branch failures.
To design an HMI that allows drivers to take over safely and to recognize which direction they should go, it will be important to investigate whether the driver's maneuvers and gaze will change depending on the manner of presenting the course.Larsson et al. (2015) suggested that a simpler HMI would be preferable for smooth RtI.
From the viewpoint of safety, if measures based on HMI alone are insufficient, then methods using tactile sense such as directional and vibrotactile RtI (Fricke et al., 2015;Petermeijer et al., 2016;Petermeijer et al., 2017) or haptic shared control (Abbink et al., 2012) are also conceivable.These approaches can convey information to the driver through tactile.It is necessary to further investigate how to design an HMI for automated driving and the limitations of RtI.
It is also important to design a system that does not issue the RtI at the scene where the driver is forced to change lanes suddenly.However, there are various situations when a system must suddenly issue an RtI (e.g.accident ahead, road construction, system failure).Even if driving authority is suddenly transferred, an HMI design that allows safe handover that/allows safe handover requires further investigation.

Conclusions
Automated driving may lower a driver's situation awareness; therefore, the main objective of this study was to investigate the influence of automation on understanding of driver's own localization.The driving simulator experiment involved two types of operation: manual and automated operations, and all participants experienced three different scenarios.All drivers were instructed how to go to the course in advance.The car was not equipped with a navigation system; thus, the driver had to remember the destination.
Furthermore, the automated operation group was instructed to watch an Internet TV program during automated driving, and to take over the operation if an RtI was issued.The experimental results suggested that the driver may not drive to the expected destination when lane changing is forced right after resuming manual control from automated driving.In the scene of the junction where a lane change to the right was required after an RtI, nearly half of the drivers could not change course correctly.This result suggests that it is insufficient to just issue an RtI, and shows the necessity of presenting what the driver should do afterwards.It also seems to be important that the HMI should provide information on where a driver is and where to go upon regaining control of the driving task.
It would be important to investigate how much information is appropriate and how much instruction should be provided, while changing these levels step by step.Furthermore, it will be necessary to investigate what should be presented to the driver not only during the RtI but also during normal automated driving via the HMI.

Study on driver's braking intention identification based on functional near-infrared spectroscopy 1. Introduction
Compared with the traditional driver assistance system, the cooperative driving system is required to be more intelligent and to have the capacity to identify drivers' driving intentions in real time.It can also adjust the condition of the vehicle to adapt to the driver so that it can achieve coordinated control and relieve The current issue and full text archive of this journal is available on Emerald Insight at: www.emeraldinsight.com/2399-9802.htmthe stress of the driver.Driver's intention identification is an extremely challenging problem in the field of cooperative driving and it now leads to a proliferation of studies.There are two methods to identify a driver's driving intention: one is based on driving operation data (Klingelschmitt et al., 2014;Jin et al., 2012;Dang et al., 2013) and the other on brain activity data (Ikenishi and Kamada, 2014;Ikenishi and Kamada, 2015).Many researches have studied drivers' driving intentions based on driving-operation data (Zhang et al., 2011;Xiong et al., 2016;Xin, et al., 2017;Su et al., 2017;Bocklisch, et al., 2017).Li et al. (2016) proposed a novel algorithm that combined the hidden Markov model (HMM) and Bayesian filtering (BF) techniques to recognize a driver's lane-changing intention.Zhou and Wu (2011) studied a recognition method for driver's intention based on genetic algorithm and ant colony optimization.Frederik et al. (2015) described the development of a driver intention detection algorithm for automated emergency braking systems .Schmidt et al. (2015) presented a mathematical model of the steering wheel angle, which was supposed to contribute to predicting lane-change maneuver.Although a driver's driving intention can be accurately identified based on driving data, referring to a posteriori method, the intentions may not be confirmed before the driving operation.Driving is a process that involves perception, judgment and operation and is associated with brain activities.Researchers suggested that different driving operations were related to different activities in the Brodmann areas of the cortex (Oka et al., 2015;Orino et al., 2017).Accordingly, we can identify a driver's driving intention directly through the application of technologies that are used to measure brain activities (Foy et al., 2016), such as functional magnetic resonance imaging (fMRI), electroencephalogram (EEG) and functional near-infrared spectroscopy (fNIRS).
EEG is the most traditional measurement method and is frequently used in fatigue driving detection and driving intention identification (Ikenishi et al., 2010;Ikenishi et al., 2013;Vakulin et al., 2015).However, EEG is insufficient in spatial resolution and susceptibility to the electromagnetic environment.With the introduction of fMRI and fNIRS, the researchers studied the relationship between brain function and driving behavior in a detailed way.You et al. (2012) and Maguire (2012) used fMRI to study the effects of brainmemory-related areas on the driving behavior of taxi drivers.Calhoun and Pearlson (2012) adopted simulated-driving paradigms to study both the healthy brain and the effects of acute alcohol administration on functional connectivity during such paradigms.Schweizer et al. (2013) identified the brain areas involved while performing different real-world driving maneuvers and assessed the effects of driving while distracted.Takahashi et al. (2010) studied the driver's brain regions associated with recognition of signals at intersections, using NIRS.Yoshino et al. (2013aYoshino et al. ( , 2013b) ) used fNIRS to study the relationship between driving behavior and cerebral cortex activity.However, fMRI has some disadvantages because of the harsh experimental environment in which the participants are made to lie down.In addition, although it has a good spatial resolution, it also has the problem of poor temporal resolution (Kato, 2004).Therefore, fMRI is not suitable for real-time identification of driving intentions.Compared with the fMRI device, the fNIRS device is more flexible as it allows the driver to maintain a normal driving posture during the experiment and complete the driving operation flexibly.
For better cooperative driving, the driver's driving intention must be perceived with the Advanced Driver Assistance System (ADAS) in advance.This study aims to establish a driver's driving intention identification model.Therefore, we used the NIRx device to measure the data of cerebral cortex activities to identify driver's braking intentions.The experiment was carried out in a virtual reality environment.During the experiment, the driving simulator (DS) recorded the driving data and the fNIRS device recorded the data of hemoglobin concentration in the cerebral cortex.After the experiment, the driver's braking intention identification model was established through principal component analysis (PCA) and back propagation neural network (BPNN).The purpose of this study is to establish a model that can accurately identify a driver's driving intention in real time, and to provide a new research method for driving intention identification in the field of cooperative driving.

Methodology
The flow diagram of experimental data acquisition is shown in Figure 1.Firstly, we designed an experiment for the study of driver's braking intention, which was implemented on the DS.During the experiment, the DS could collect driving data in real time, such as speed, engine speed, acceleration, lateral acceleration, steering wheel angle, front wheel angle, brake pedal pressure and acceleration pedal pressure.The fNIRS device could record the oxy-hemoglobin concentration changes (DHbO) and deoxy-hemoglobin concentration changes (DHb) of the driver's cerebral cortex in real time.After that, we established the driver's braking intention recognition model through these experimental data.

Experimental design based on driving simulator and functional near-infrared spectroscopy
The experiment was conducted in a virtual reality (VR) lab (Figure 2).A total of 52 participants were invited to participate in the experiment: 10 female and 42 male.All participants possessed a valid Chinese driving license.Their ages ranged from 19 to 38 years, with an average age of 26.5 years.Before the experiment, each driver needed to close his/her eyes and sit in the cab for about 2 min.The experiment was started only after each driver's cerebral cortex blood oxygen concentration stabilized.The value of the change in cerebral blood oxygen concentration at that time was used as reference for cortical activity.At the Figure 1 The flow diagram of experimental data acquisition beginning of the experiment, participants sat in a realistic vehicle mock-up and controlled it in the DS.The drivers had to keep quiet and calm during the whole experiment.While the experiment was going on, the DS recorded driving-related data, including the vehicle speed, brake pedal pressure, acceleration pedal pressure and steering wheel angle, and the fNIRS device measured the participants' brain activities.Each experimental condition was introduced as follows.
The NIRx device (i.e.fNIRS device) used in the experiment is shown in Figure 3 and it was provided by NIRx Medical Technologies, LLC.The NIRx device adopted a unique measurement strategy wherein every possible combination of sources and detectors formed a measurement channel.It relies on the optical determination of changes in hemoglobin concentrations in the cerebral cortex, which results from increased regional cerebral blood flow.Two wavelengths were set at 760 and 850 nm for all the recording channels.The frequency in the sample was set at 7.8152 Hz (record data every 0.128 s).The NIRx device channel distribution and collection channel area numbers are shown in Figure 4. We can see that the NIRS cap contains 16 light source points (S1-S16) and 16 light detector points (D1-D16).The data acquisition channel is located between the light source and light detector points.There are 41 data acquisition channels on the NIRx device, which can record the DHbO and DHb.With this experiment, we could calculate the total hemoglobin concentration changes (DTH) based on DHbO and DHb.DTH is the sum of DHbO and DHb.
During the experiment, the drivers needed to drive according to the road signs.The experimental roads included sections of constant speed, acceleration and deceleration.As shown in Figure 5, the accelerated section road was located before the 70 speed limit sign, the deceleration section road was between the 70 and 50 speed limit signs and the constant speed section road was behind the 50 speed limit sign.Therefore, if a participant drove on this road section and if his vehicle speed was more than 50 km/h, he/she had to slow down the vehicle.If the vehicle speed was no more than 50 km/h, no deceleration operation was required.During the experiment, each driver passed the deceleration section roads six times.After the experiment, about 300 deceleration section road samples were created.

Experimental data preprocessing
In order to establish the identification model of driver's brake intention, the experimental data was divided into two groups: braking group and non-braking group (see Figure 6).The braking group had brake operations, with no steering operations.The non-braking group had no brake operations with no steering operations in a time window, of which, traveling at a constant speed of not less than 5 s.After the experiment, a total of 152 braking group samples and 115 nonbraking group samples were created.On analyzing the experimental data recorded by the NIRx device, it was found that the DTHR (DTH range, i.e. the difference between the reflected the intensity of cerebral cortex activity.DTHR was very small when the driver kept calm.Conversely, DTHR would increase when he/she started thinking and judging.Therefore, DTHR could reflect the brain's process from a static state to a state of thinking and judging.Also, we could use DTHR as an input for driving intention identification classifier.

Establishment of braking intention identification model
In this experiment, a total of 266 valid sample data were obtained, including 152 samples of braking group and 115 samples of non-braking group.About 80 per cent (120 brake group samples and 95 non-brake group samples) of the total sample was used as training data and about 20 per cent (31 brake group samples and 20 non-brake group samples) was used as testing data.The NIRx device used in this study contained 41 data acquisition channels, which meant that each sample had 41 features.However, after the model was built, too many features and very few samples resulted in overfitting.Therefore, PCA was used to reduce the dimensionality of sample data and the driver's braking intention identification model was established through BPNN.PCA is one of the most commonly used dimensionality-reduction methods that aims to transform the multi index into a few comprehensive indexes (i.e. the principal component) by using the idea of dimensionality reduction.Each principal component of PCA can reflect most of the information of the original variable, and the information contained is not repeated.PCA processing is as follows: Standardization of raw data: All samples of braking and non-braking groups are put into matrix X. Equation ( 1) is the standardizing process, zij is the value of the standardized matrix after the standardization processing, xij is the value of row i and column j of matrix X and n is the row number of matrix X.
Calculating the correlation coefficient matrix R for the standard matrix Z: equation ( 5) is the calculation of rij.

Calculation of eigenvalues and eigenvectors of matrix R:
The eigenvalues of matrix R can be obtained by using equation ( 6).The eigenvalues are arranged according to the values.(l 1 = l 2 = [. ..] = l n = 0).Then, the corresponding eigenvectors of each eigenvalue are calculated.
Calculation of principal component contribution rate and cumulative contribution rate: equation ( 7) is the calculation of a single vector's contribution.Equation ( 8) is the calculation of the cumulative contribution.Eigenvalue l 1 , l 2 ,[. ..] and l m correspond to the 1st, 2nd, [. ..] and m-th principal components.In general, the eigenvectors corresponding to the eigenvalues with a cumulative contribution rate of 90 per cent are selected as the eigenvector matrix E [see equation ( 9)].
Mapping.z ij ¼ x ij À x j s j (1) By mapping the sample matrix X to the selected eigenvector E, we can obtain the reduced dimensional matrix T [equation ( 10)].In this study, the dimensionally reduced samples were used as input samples for the classifier BPNN.BPNN is a multi-layer feed-forward neural network.It is mainly characterized by that the signal propagates forward and the error propagates backward.The learning process is mainly divided into two stages.The first stage is the forward propagation of the signal.The signal passes through the hidden layer in the input layer to the last output layer.The second stage is the back propagation of the error.The signal goes from the output layer to the hidden layer and finally reaches the input layer to adjust the weight and bias of each layer.The typical BPNN structure is shown in Figure 7.It contains an input

Model validation
Three drivers' driving data under the deceleration condition road were randomly selected to test and verify the model established in this paper.The process is shown in Figure 9.The input data of the braking intention identification model were the DTH vector of the driver's cerebral cortex at some point, and the output data were the testing result of the model to the driver's braking intention.The internal processing of the model is as follows: In the first place, the vector of DTHR with a time window of 3.84 s was calculated based on the raw DTH vector.
After that, the calculated vector DTHR was added to matrix A to obtain a new matrix D.Then, the PCA algorithm was applied to reduce the dimension of matrix D. The reduced vector E was worked out as the input of BPNN.Finally, the output of the model was whether the driver had a braking intent at the moment.indicates the driver has no braking operation at the time and 1 indicates the driver has a braking operation.As shown in Figure 11, the output of the model was greater than 1 before the driver's braking operation, which indicated that the model could identify the driver's braking intent prior to his braking operation.

Discussion
ADAS has raised expectations for the reduction of human error while driving and released or freed people from the task of driving.We called the pattern that ADAS sharing controlling with human driver and completing driving task together as cooperative driving.To achieve better cooperative driving, the driver's driving intention must be perceived by ADAS in advance.Driver's driving intention identification is of great significance to automatic driving, and identifying driver's driving intention with ADAS in advance is helpful for vehicle safety, stability and comfort.This study aimed to establish a driver's driving intention identification model.To this end, we designed an experiment based on DS and NIRx devices to study the driver's braking intention.The simulation experiment was carried out on the deceleration road (Figure 5).During the experiment, the DS collected driving-related data and the NIRx device collected cortical activity data.Then, the machine-learning algorithm was used to establish the identification model of driver's braking intention.
In cooperative driving, the accuracy and timeliness of driver intention recognition are particularly important for ADAS.Studies (Li et al., 2016) have shown that drivers' driving intentions can be identified accurately from driver's driving data.However, this method can identify the driver's driving intention only after the driver has a corresponding driving operation.If the driver's driving intention can be identified by ADAS in advance, it will be better to realize cooperative driving.To this end, we explored the method of identifying driver's driving intention by brain activity measurement.In this study, the NIRx device was used to measure the activity of the driver's cerebral cortex.
At present, researchers have used fNIRS to study the relationship between driving behavior and cerebral cortex activity, such as the relationship between activity in the prefrontal cortex and vehicle speed (Yoshino et al., 2013a(Yoshino et al., , 2013b)).Oka et al. (2015) studied the respective active regions of the brain when turning left and right through curve conditions experiments.Orino et al. (2017) studied the relationship between the activity of the entire brain and the behavior of driving operations during actual road travel.However, a few researchers have used fNIRS to predict driver's driving intentions in advance.In this study, based on the DS and the fNIRS experiments, the prediction model of driver's braking intention was established.The test accuracy of the model with the driver's braking intention was 80.39 per cent.Although the recognition rate in this paper was lower than the recognition rate of the model built by Ikenishi and Kamada (2015) (about 88 per cent), fNIRS was used to predict driving intention for the first time, which opened a new research direction for driving intention identification.
It is worth noting that the driving intention identification model established in this study can recognize the driver's braking intention only from constant speed driving to deceleration driving.The limitation of this study was that the experimental environment was ideal and did not consider the surrounding traffic.At the same time, other actions of the driver were not taken into account when establishing the braking intention recognition model.Besides, the validation data were gathered from 52 drivers' experimental data, and the verification results obtained in this paper could only reflect the results of some drivers' identification of braking intention.Because of the individual differences among the drivers, the identification model of the driver's braking intention established in this paper cannot accurately identify every driver's braking intention in real time.The individual differences in the brain activities of the drivers while driving are mainly reflected in the slight differences in the changes of hemoglobin concentrations in the cerebral cortex when the driver has a braking intention.Therefore, in the future research, we will establish a driving intention identification model under complex driving conditions and improve the identification accuracy of the model for driving intention to achieve the effect of real-time identification of driver's braking intention.In addition, we will consider differences in the brain activities of drivers when they have the same driving intention.

Conclusion
In this study, the driver's braking intention identification model was established through PCA and BPNN.In addition, the data of three drivers driving under deceleration condition were gathered randomly to verify the model.The research results showed that the accuracy of the model established in this paper was 80.39 per cent.This study can be used as a reference for future research on driving intention through fNIRS, and it has a positive effect on the research of brain-controlled driving.At the same time, it has developed new frontiers for intention recognition of cooperative driving.
Figure 11 Verification results of the braking intention identification model adopted principal component analysis (PCA) to extract critical characteristics.And then, the discriminant model of driver's driving skill was established by using SVM, K-nearest neighbor (KNN) and probabilistic neural networks (PNN) (Chandrasiri et al., 2010(Chandrasiri et al., , 2012(Chandrasiri et al., , 2016)).Ly et al. (2013) 2017) used a k-means clustering method for drivers' labeling and applied a semi-supervised approach, namely, a semi-supervised support machine (S3VM), to classify various driving styles, the data labeling required a prior is greatly reduced and S3VM improved classification accuracy by about 10 per cent.Li et al. (2013Li et al. ( , 2014) ) studied drivers' driving skills under a specific curve by using wavelet analysis to extract critical features and established the algorithm of experienced driver's behavior extraction based on AdaBoost.The above three studies were based on curved roads, using indirect features that reflected the potential specifics of practiced drivers and unpracticed drivers as candidate features.The studies analyzed drivers' lateral driving traits and longitudinal driving characteristics at the same time.Although drivers' driving skills can be better reflected in lateral and vertical operations under the cornering condition, the method of generating candidate feature results in a driving skill analysis only based on several single features, which cannot reflect driving skill on drivers' co-occurrence of driving operation; although signals of different frequency components can be found in the same feature, it is still limited to a single feature.This paper took advantage of candidate combined features reflecting the consistency of driving operations; critical features were extracted using AdaBoost at the same time.Section 1 of this paper introduces the main achievements in terms of drivers' driving level.Section 2 involves a battery of experiments designed for driving data collection based on driving simulator.Section 3 describes data processing method and data analyzing approaches.Section 4 discusses relevant data analyzing result.Section 5 states the conclusions.The main research process is shown in Figure 1.

Experiment
This experiment was carried out with a driving simulator (DS) (Figure 2), which consisted of a visual system with a field of view of 140°around, a sound system and a dynamic model.The driving environment for the experiment (Figure 3) was a city road with six curves with left turn, and these curves, with different radiuses and lengths, were numbered 1-6 according to the travel direction (Figure 4).The speed limit of 60 km/h at 50 and 100 m before the start of each curve required drivers to maintain a speed of about 60 km/h before entering the curve.The collected data contained the position of accelerator and brake, front wheel angle, vehicle speed, lateral acceleration, longitudinal acceleration and yaw rate, with a sampling frequency of 60 Hz.To obtain sufficient experimental data, a total of 16 drivers of different driving levels participated in the experiment.Each driver completed 12 laps, the first two of which were test drives.Basic information of drivers is shown in Table I.

Data normalization
All data collected on the basis of time were normalized with a certain distance according to the travel direction utilizing liner interpolation so that the same curve at different laps had comparability.The normalized data with the same data length are shown in Figure 5. Murphey et al. (2009) suggested that a smaller jerk or steady driving process would result in less fuel consumption and higher safety.This means that the smaller the jerk, the higher the driving skill.Complex jerk on behalf of a changed rate of acceleration at a distance was used for showing driver's driving skill in a curve.J, representing the complex jerk, is given in equation ( 1).In the condition of the same average speed as described above, the bigger the variate J, the lower the driving skill:

Driving skill labeled
where J lateralÀi and J longitudeÀi stand for lateral and longitudinal accelerations, respectively, at the i-th point in one curve and the variate N is the total number of standard points in the same curve.

Method for data processing
3.1 Generation method for candidate combined features Candidate combined features were decided by driving data, including steering wheel angle, accelerator petal position, brake petal position and corresponding operation and vehicle speeds.We chose the average in a distance of 9 m, which contained 30 standard points as candidate features, to decrease the error caused by operating occasionality, and the averages were extracted every other point: where variable y represents the change of single feature P.This paper referred to the feature co-occurrence for face detection (Mita T et al., 2005), which combined two or more different features into one feature, called the combined feature.The following gave the combined principle of two features at the same point: for a single feature P, the current feature P i 1 1 was compared with the previous adjacent feature P i , and a threshold value D was set empirically for each kind of the feature P.Then, ternary numbers 2, 1 and 0 were used to indicate that the difference of P i 1 1 and P i was greater than D, equal to D and less than D, respectively.The variable y for a sample P is figured in equation ( 2).
With the above processing for two features at a certain point of a certain curve, we could obtain a two-dimensional N Â 2 array, according to the rule of converting a ternary number into a decimal number [equation (3)], the ternary array of n Â 2 was converted to the decimal array of n Â 1 and the decimal array only contained the elements of 0-8, representing the nine kinds of candidate combined features, as shown in Table II: where D 1 and D 2 are both ternary numbers.

Method for feature extraction
The feature extraction processing using AdaBoost is shown in Feature extraction processing using AdaBoost: 1. Given example of labeled data (x 1 , y 1 ), (x 2 , y 2 ), . .., (x n , yn), where x i [ X, y i [ {À1, 11} 2. Initialize weight w i;t ¼ 1 N , y i = 0,1, where 0 and 1 are on behalf of experienced driver and  Þ was a liner combination of a group of T weak classifiers.An optimal operation feature would be extracted per iteration until reaching the error threshold of classifier in step 3.

Feature extraction
The relationship between number of weak classifiers and error rate of strong classifiers in Figure 6 was a critical step for deciding the number of weak classifiers using AdaBoost.The number of weak classifiers was the number of features extracted.We found that the error rate of strong classifiers was less than three per cent as the number of weak classifiers reached 15.This paper stipulated that when the accuracy of classifiers satisfied 97 per cent, the process for feature extraction was completed.Figure 7 shows the concrete locations of a part of the 15 features extracted.A major difference between skilled and unskilled drivers was obvious at the entrance.The details of those features are provided in Table III.For example, the first combined feature consisting of velocity and steering angle appeared at the site that was 62.7 m away from the origin of 1-th curve, and the feature value of this combination was bigger than 1.5 as to inexperienced drivers.

Features distribution characteristics
Curves were divided into five parts, including 50 m before curve, 50 m after curve and trisection of the remaining curve in Figure 8.They were named sections AB, BC, CD, DE and EF along the travel direction.Figure 9 shows all features' distribution on the five sections of curved proposed above.Most features occurred at the entrance and exit, which were in line with actual driving as drivers got used to adjusting driving operations at those parts.In contrast, there were a few operations in the middle of the curves, seen in section CD.Combined features of "steering wheel operation speed and accelerator operation speed" and "accelerator petal position and steering wheel operation speed" were the most frequently extracted, which meant that the difference between the two groups of drivers was mainly in these two combined features.
In section AB, it was found that the combined feature of steering wheel angle and accelerator operation speed was more frequently extracted.In fact, drivers changed the steering wheel angle and velocity constantly at the entrance to adapt to the

Conclusion
This paper proposed a method for driving operations characteristics analysis, using AdaBoost and feature cooccurrence.When the driving operations went through the curves at a special course, they were studied based on DS.In the end, all features corresponding to relevant curves were selected and extracted using the proposed method.The result illustrated that most features came out at the entrance and exit of all curves, which conformed to actual behavior when drivers entered or left curves.We just studied driving feature extraction, which was a part of fundamental research in the field of driving operations characteristics.In the future, we plan to enrich the driving environment and not keep it restricted to courses consisting of curves alone.We are also keen to develop a driving assistant system that will help improve inexperienced drivers' driving skills through driving behavior analysis, so as to decrease traffic accidents.

Figure 1
Figure 1 Illustration of the occupancy cells of LC scenarios

Figure 2 Figure 3
Figure 2 Illustration of the occupancy cells LC scenarios

Figure 1 Figure 2
Figure 1 Driving simulator used in the experiment

Figure 3
Figure 3 (a) Status of the automation engaged; (b) status of the automation disengaged and (c) issue of RtI (was written in Japanese)

Figure 8 Figure 9
Figure 8 Proportion that proceeded to the instructed direction

Figure 10
Figure 10 Point where lane changing began in scenario B

Figure 2 Figure 3 Figure 4 Figure 5
Figure2The DS used in this study

Figure 6
Figure 6 Grouping of experimental data

Figure 10 Figure 7 Figure 8 Figure 9
Figure 10 shows the relationship between the number of sample features and the test accuracy of BPNN: the x-axis represents the number of sample features after PCA dimensionality reduction and the y-axis represents the test accuracy of BPNN.It can be seen that the samples have different feature quantities corresponding to the different test accuracy of BPNN.When the sample feature number was 15, the network achieved the best test accuracy of 80.39 per cent.The verification result of the braking intention identification model is shown in Figure 11, in which the x-axis represents the time and the y-axis represents the verification result of the model.In this figure, the blue dashed line indicates the identification result of the model to the driver's braking intention.A result of more than 0 means that the driver has a also used a support vector machine (SVM) to recognize driving styles based on the labeled information of the vehicle's inertial sensors.To model and analyze driving styles semantically, Wang et al. (2017) gave a new framework for driving style analysis using primitive driving patterns with Bayesian nonparametric methods, a hierarchical structure (HDP-HSMM) was developed by combining hierarchical Dirichlet process (HDP) and hidden semi-Markov model (HSMM), which could learn a set of expected primitive driving patterns in car-following behaviors.Wang et al. (

Figure 4
Figure 4 Driving route of experiments

Figure 6 Figure 7
Figure 6 Error rate of strong classifier at 1-th curve

Table I
Values of the defined cell grid Original features from on-board sensors Vehicle yaw rate and lateral acceleration are usually used as strong features of vehicle lateral behavior.Together with longitudinal acceleration, the above signals are necessary for recognition, prediction and modeling vehicle lateral behaviors(Leonhardt and  Wanielik, 2017; Higgs and Abbas, 2015; Li et al., 2015; Luo et al.,  2016).Here we also choose the following signals directly collected from on-board sensors as our candidate features: ).If a small TTC indicates the driver may execute LC to overtake the slow leading vehicle.Thus the TTC can be regarded as a valuable feature to recognize LC maneuver(Kasper et al., 2012).Time-to-lane crossing (TLC) represents the time available for a driver until the moment at which any part of the vehicle reaches one of the lane boundaries(Godthelp et al., 1984).It is a parameter to estimate if the ego vehicle is going to cross the lane.Based on (J2944, 2013), TTC and TLC are given by.

Table IV
Final selected strong features for each LC scenario

Table V
AUC values of comparison results with different models using the selected features and all features in each LC scenario Figure5ROC curves of comparison results with different models using the selected features and all features in LLC scenarios Figure6ROC curves of comparison results with different models using the selected features and all features in RLC scenarios

Table AI
Full-scale effect size of the featuresFor instructions on how to order reprints of this article, please visit our website: www.emeraldgrouppublishing.com/licensing/reprints.htm Or contact us for further details: permissions@emeraldinsight.com

Table I
Information of driversFigure 5 Data before/after normalizationinexperienced driver respectively.3. Iteration times t = 1,2,3,. .., T 3.1 The t-th weak classifier H(x): X !{À1, 11}, its error rate « t is evaluated with respect to w t (i): Weight of the t-th weak classifier a t ¼ 1 2 ln 1-e t corresponded to the label of variate x i .As initialized weight was 1/N, weight would update once per iteration and be used in the next iteration.The last strong classifier H x ð Þ ¼ sign a t h t x ð Þ ð

Table II
Combined feature methodNotes: Ã 2 means P i 1 1 À P i > D, feature P increased; 1 means jP i 1 1 À P i j D, feature P unchanged; 0 means P i 1 1 À P i < ÀD, feature P decreased

Table III
Basic information of features extracted at 1-th curve