Study on Driver Behavior Pattern in Merging Area under Naturalistic Driving Conditions

,


Introduction
With the increase of trafc fow, trafc pressure in the highway merging area is heavy.Drivers need to fnish complex driving tasks such as acceleration, deceleration, and lane change in merging area.Although the merging area account for a small proportion of the highway, trafc accidents occurred in this area account for a large percentage.For example, during the period from 2014 to 2016, the total length of interchanges in a province in eastern China accounted for about 9.5% of highway in this province but the number of accidents accounted for 34.5% of the total number of highway accidents, and it accounted for 24.95% of the total number of fatalities in the province [1].In a statistical study in the United States, it also showed that accidents occurred in the highway interchanges accounted for 18% of the total number of accidents and the percentage of fatal accidents is 21.8% [2].As an important part of the highway, the merging area is an important "valve" for auxiliary vehicle steering.In the merging area, as vehicles continue to merge into the mainline lanes, driving behaviors such as vehicle acceleration, deceleration, and lane changes occur more frequently, which is likely to cause drivers to take improper driving actions and causes trafc accidents.Terefore, it is important to study speed changes, trajectory changes, and other driving behaviors in merging area to identify and resolve trafc conficts and improve trafc safety.
Due to drivers' personalities, driving styles are diverse, which can be broadly classifed as aggressive, steady, and conservative [3].Early classifcation of driving styles relied on questionnaire surveys, which were subjective [4][5][6].In recent years, with the update of driving simulators and sensors, the evaluation of driving styles through specifc parameter evaluation indicators (e.g., speed, acceleration, etc.) has become more convincing [7].Many scholars used simulations to obtain data such as time headway, acceleration, and steering wheel angle to analyze driving styles or behavioral characteristics.Shi et al. took the lane-changing behavior on highways as the research object and obtained subjective and objective driving data through questionnaires and driving simulators.Te experimental analysis showed that the driving style had a great infuence on drivers' lanechanging behaviors [8].Orfanou et al. studied driving style in trafc congestion and used neural networks to analyze trafc fow parameters.Te results showed that the parameters with the greatest infuence on driving style are distance and acceleration [9].Yang et al. constructed a realtime trafc crash risk prediction model considering the temporal efect diference and explored the relationship between dynamic trafc fow characteristics and real-time trafc crash risk under diferent temporal conditions [10].Zhao et al. collected pedal and speed data under diferent trafc fow densities based on a driving simulator.Feature parameters were extracted, and a driving style recognition model was established considering the infuence of trafc fow density [11].Aguilar et al. established a driving style recognition system based on machine learning algorithms, aiming to improve driving safety [12].In addition, aerial photography technology has been widely used in transportation research and application felds, and some scholars tend to use UAVs to collect real vehicle data [13,14].Constantinescu et al. extracted real-time vehicle motion parameters and proposed a driving style evaluation model based on real vehicle driving data [15].Rodriguez Gonzalez et al. investigated the efects of vehicle speed and lateral and longitudinal acceleration on driving behavior based on data collected from real vehicles and identifed driving behaviors with speed, acceleration, and fuel consumption evaluation indexes [16].Wu et al. utilizes roadside radar to identify and detect targets and classifes them using a two-phase method with a detection accuracy of 89.5% [17].Li et al. took longitudinal vehicle speed and vehicle speed error as input, and took acceleration and brake pedal opening as output.A driver model that can refect driving styles was established based on a neural network algorithm [18].Zhang et al. proposed an improved driver clustering framework by accounting for road types and average speed.Te clustering results were compared with those without considering trafc conditions.Te improved clustering framework performs better in both intraclass aggregation and interclass separation [19].Yang et al. constructed a multidimensional multilevel system for trafc crash analysis, this system was capable of accurately and efciently capturing the mechanics of high-consequence (and possibly low support) highway crashes [20].Wu and Xu analyzed the efects of fve types of driving behaviors on accidents using random forests.Tese included driver age, road alignment, trafc density, road environment, and whether the driver had both hands on the steering wheel [21].Other scholars have analyzed lanechanging behavior.Li et al. established a grouped random parameter logit model with heterogeneity in means and variances (GRPMV) and its baseline models, the GRMPV model has the best model performance and can better capture the unobserved heterogeneity [22].Li et al. investigated the characteristics of discretionary lane change (LC) duration on freeways, established accelerated failure time (AFT) models with fxed parameters, latent classes, and random parameters [23].
It can be found that the current studies related to driving behavior mostly focus on the analysis of driving behavior characteristics of urban roads [24][25][26][27][28][29][30][31][32][33][34][35][36].Tere is a lack of studies focusing on the driving behavior for highway merging areas.Te ability to acquire microscopic driving behaviors such as speed changes and lane changing is insufcient.Terefore, in this paper, a UAV was used to collect data of highway merging areas, data of vehicle speed and vehicle type was extracted based on YOLOv5 and Deep SORT.Te efects of diferent types of merging areas, different lanes, and diferent vehicle types on driving behavior were analyzed.

Materials and Methods
2.1.Data Acquisition.Tis paper adopted the method of fxed-point shooting by UAV and monitoring to obtain realtime data, which can record the whole process of following and changing lanes on the main road and ramp within the merging area of the highway.Te use of drones to collect data can avoid interference with the driver and truly refect the driver's driving behavior.
Te videos collected by DJI UAV were divided into two parts.One part was the video data from the actual merging area of the highway, as shown in Figure 1, and the other part was the speed verifcation video data of non-highway sections.Te videos include two types of interchanges and service areas, and the collecting time period for each site was one hour.Te shooting height of the UAV was fxed in the range of 15−20 m, and the shooting area of the merging area was about 450 m.

Vehicle Recognition Detection Model
2.2.1.YOLOv5-Based Vehicle Recognition Model.Firstly, the dataset is classifed and given labels.As shown in Figure 2, the vehicle types in the vehicle recognition dataset are classifed into 3 categories: truck, bus, and car.Classifcation was based on body length, with large vehicles being over 6 m, medium vehicles being 3.5 to 6 m, and small vehicles being under 3.5 m.Te dataset was divided into 3 parts: Train, Test, and Valid, where the Train dataset includes 1488 sample images, the Test dataset includes 31 sample images, and the Valid dataset includes 507 sample images.
YOLOv5 integrates some features of YOLOv3-spp and YOLOv4, with the advantages of small size, extreme speed, and high accuracy.Meanwhile, YOLOv5 adds adaptive anchor frame calculation and mosaic data enhancement, the detection efect of small targets is improved and the overall application efect is better.Terefore, the optimal best weight was derived in this project using YOLOv5 model training.Te weights are brought into the recognition program, and the recognition efect is shown in Figure 3. From the classical formula V � S/T, the speed can be obtained by dividing the distance travelled by the time in one direction in a certain period.t 0 is assumed to be the initial time.s 0 is seen as the initial position.t 1 is supposed to the stopping time and s 1 is regarded as the ending position, then, we have the following equation: d 0 and d 1 are represented as the pixel positions corresponding to the target vehicle at moments t 0 and t 1 in the continuous video sequence, respectively.Ten, (d 1 − d 0 ) is represented as the pixel distance that the target vehicle moves from t 0 to t 1 .If f (x) is a single mapping function that represents the mapping relationship between the actual trafc road coordinate system and the image pixel coordinate system, then the formula can be transformed as shown in the following equation: Currently, we already know the transformation relation f (x) between the image coordinate system and the world coordinate system.Ten, this can be used to calculate the actual moving distance of the target vehicle in the video.Further combined with the time parameter, the actual travelling speed of the vehicle is calculated.

Validation of Trafc Flow Detection Results
. A total of 2 videos of each of the 5 diferent congestion levels were selected for trafc fow verifcation, and the length of the videos were all 3 minutes.At frst, the trafc fow of each video was recorded by manually counting and then compared with the calculation results of the detection model, and the detection accuracy is shown in Table 1.
As shown in Table 1, when the road trafc is running smoothly, the detection accuracy can reach 100%.However, as the level of trafc congestion increases, there was serious occlusion between vehicles, resulting in a decrease in detection accuracy.

Validation of Vehicle Speed Detection.
A total of 5 fxed speed driving videos were used for speed verifcation, including 10 km/h, 20 km/h, 30 km/h, 40 km/h, and 50 km/h.Te data acquisition scenario is shown in Figure 4(a).Te vehicle speed was read by extracting the speed value every 2 frames.Since the videos with diferent speeds have diferent number of frames, there was a diference in the amount of speed value data available for each video.In addition, the vehicle needs an acceleration process to drive to a fxed speed, so the middle 15 values of each video speed data are selected.Te overall average error rate of the vehicle speed was calculated to be 10.03%, as shown in Figure 4(b).Te detection error was about 22.67% at the maximum and 0% at the minimum.Te speed fuctuates more when the vehicle is farther away from the flming location and stabilises the closer it gets.

Data Preprocessing.
Model and speed data of 2450 vehicles were extracted using the YOLOv5 + Deep SORT model.Due to problems such as video shooting angle and vehicle occlusion, some vehicle speed changes were incompletely recorded, so the raw detection data needed to be sorted and screened.Finally, 2000 vehicles were selected as sample data for driving behavior analysis, of which, 1000 were vehicle data from interchanges and 1000 from service areas.

Variation of Vehicle Speed in Different
Types of Merging Areas

Trafc Flow and Speed Distribution of Diferent Lanes.
Speed percentage can be more obvious to see the concentration of speed distribution.To better analyze the distribution of vehicle operating speed in the merging area, so as to determine the distribution of speed in each lane, the percentage is used as an indicator for analysis, and the corresponding math expression is as shown in the following equation: In the equation, P j represents the percentage of vehicles in a lane whose speed is in the i speed interval.N j represents the number of vehicles in a lane whose speed lies in the i speed interval.
From vehicle speed data extracted by statistics, it is found that the speed distribution of each lane of the interchange is obviously diferent from that of each lane in the service area merging area, and most speeds are distributed between 60 km/h and 120 km/h.For this reason, the speed range of 0 to 60 km/h is used as the frst speed interval, and then 12 speed intervals are generated from 60 km/h to 120 km/h at an interval of 5 km/h.According to the formula (3), the percentage of vehicle speed of each lane in the corresponding speed interval is counted separately for two diferent types of merging areas.
Both the interchange and the merging area of the service area have four lanes, with three lanes on the main road and a single lane on the ramp, as shown in Figure 5. Te interchange and the merging area of the service area both have four lanes, the main road is three lanes, including the inner lane, middle lane, and outer lane, and the entrance ramp is a single lane.
From the statistical analysis of the 2,000 samples, it can be seen from Figure 6(a) that the middle lanes of the interchange merging area have the highest trafc fow, about 42%.Tis is followed by the inner lanes, and the outer lanes and ramps have similar trafc fows.It can be seen from Figure 6(b) that the service area merge zone has the highest trafc fow in all lanes, about 49%.Tis is followed by the middle lane, and the entrance ramp has a slightly higher trafc fow than the outer lane.Based on the statistical data, the percentage of vehicle travelling speed for each lane in the merged area was plotted.
According to the statistical data, the percentage speed of vehicles travelling in each lane of the merging area is plotted.
For the two types of merging areas, speeds from the inside lanes to the entrance ramps all showed a gradual downward trend.As shown in Figure 7, the average speed in the inner lanes of the mainline tends to be concentrated at 105 km/h or more.Te average speed in the middle lane was mostly concentrated above 90 km/h.Te average speed distribution of the outer lanes of the mainline of the interchange is more dispersed, and the average speed in the service area merging zone was mainly concentrated in 80-100 km/h.In contrast, the biggest diference between the two lies in the speed distribution of the entrance ramps.Te average speed of vehicles on the entrance ramps of the interchanges was mostly below 60 km/h, which is obviously lower than the average speed of the entrance ramp area in the service area.

Distribution of Vehicle Acceleration and Deceleration in
Each Lane.As a factor in the change of vehicle travelling speed, acceleration is one of the main parameters refecting the driving status of the vehicle.It is also a core indicator refecting driving stability, comfort, safety, and driving behavior analysis, which can objectively refect the driver's driving behavior pattern.In this section, the average vehicle acceleration was divided into 10 intervals to analyze the acceleration and deceleration of vehicles in each lane.
Te acceleration distribution showed a trend of high, middle, and low at both ends, mainly distributed in the (−2, 2) m/s 2 interval.Te acceleration and deceleration of vehicles in each lane were basically stable; however, due to lane alignment, driver status, etc., the frequency of sharp acceleration and deceleration of vehicles in the service area merging zone was higher than that in the interchange merging zone.From Figure 8(a), it can be seen that the speed of the middle lane is more stable, and the frequency of acceleration and deceleration in the inner lane is relatively high.Te outer lanes are more closely connected to the entrance ramps, and the acceleration and deceleration frequencies of vehicles were similar and higher.Figure 8(b) shows that the middle lane has a more stable speed, the entrance ramp has a higher frequency of acceleration and deceleration, and the frequency of acute acceleration and deceleration is more frequent.

Analysis of Vehicle Lane Change Behavior of Different Lanes
Lane-changing behavior requires a combination of the surrounding trafc scenario and driving characteristics.By fnding the right time to change lanes, the vehicle reaches its driving destination.Lane-changing behavior is one of the most common behaviors of road vehicles.Especially in complex areas such as highway merging zones, it has a great impact on the road safety and stable operation of vehicles.Terefore, detailed analysis of the characteristics of vehicle lane changes in the merger area is important for improving trafc safety.
To better analyze the frequency of vehicle lane changes in diferent lanes in the merging area, the lane change rate is used as an analysis index.Lane change rate is defned as the ratio of the number of vehicle lane changes in a certain area to the total number of vehicles in that area, as shown in the following equation: where β lcr is the vehicle lane change rate, t lc represents the number of vehicle lane changes in an area, and N tv represents the total number of vehicles in a region.It is stipulated that the behavior of a vehicle entering from one lane to another adjacent lane is a lane change.Te merging area is divided into 10 regions, each with a range of about 50 m, as shown in Figure 9. Vehicle lane change data for each lane in diferent regions is then extracted from it.

Frequency of Vehicle Lane Change in the Main Road Lane.
To study the impact of merge zone entrance ramp vehicle convergence on main road vehicles, this section summarized the frequency of vehicle lane changes for each lane of the arterial.Te results were obtained as shown in Table 2.A total of 115 vehicles changed lanes on the main road in the interchange merging area, and 259 vehicles changed lanes on the main road in the service area merging area.
Since the lane lines between the middle lane and the outer lane in Areas III to VIII are solid lines, lane changing is prohibited for vehicles.So, mainline vehicles change lanes very frequently in Areas I and II to avoid conficts with the merging ramp trafc and the mainline trafc.When the lane lines of the middle lane and the outer lane will become dashed, the lane change rate between the middle lane and the inner lane in Area VIII increased.And the overall lane change rate of each lane in the merging area of the service area was slightly higher than that of the interchange.

Vehicle Change Location on the Entrance Ramp.
To summarize the pattern of where vehicles from the ramps merge onto the main roadway, the number of vehicles merging from the ramps in zones IV-VII was counted.As shown in Table 3, the number of converging vehicles in the interchange merging area ramp is 70, and the number of converging vehicles in the merging area of service area is 63.
It can be seen that 20% of the vehicles in the interchange merging area have merged onto the main road within 50 m of the acceleration lane.Tis is related to the aggressive driving behavior of drivers or the low trafc volume on the mainline.In addition, when the trafc volume of the mainline is too large or the speed increase is slow, it will lead to vehicles not being able to merge into the mainline in time.In this dataset, 10% of vehicles merge onto the main roadway late.Te location of vehicles from service area ramps joining the main road is similar to that of interchanges, with the highest number of vehicles joining within 50−100 m, accounting for 38.65%.
Using the data in Tables 2 and 3, it can be clearly seen that the situation of vehicle lane changing in each lane area, which provides sufcient data support for the subsequent vehicle driving law, confict analysis and early warning, and clarifes the focus research area of vehicle confict location.

The Effect of Different Vehicle Type on Speed and Lane Change Behavior
Tere were 899 cars and 101 trucks in the interchanges in the sample data set.Tere were 816 cars and 184 trucks in the service area.Because of the small number of buses in the sample and their large size, they are categorized among the trucks.Due to the characteristics of diferent car models, there will be a signifcant diference in driving speed and lane changing behavior in the merging zone.Tis section selected special points or sections of the merging zone to analyze the driving patterns of diferent car models.Te merging zones at the collection sites are all single-lane parallel acceleration lanes, and the zoning schematic is shown in Figure 10.It mainly included the front end of ramp confuence, the road segments for acceleration, the confuence point, the segment used for waiting, and the segment used for gradients.

Te Efect of Diferent Vehicle Type on Speed Variation.
To compare the efect of diferent vehicle type on the speed change, this study summarized the average speed change of vehicles of diferent vehicle type on the ramp, as shown in Figure 11.Te abscissa coordinate 1 in the fgure indicated the front end of ramp confuence, 2 indicated the road segments for acceleration, 3 indicated the confuence point, 4 indicated the segment used for waiting, and 5 indicates the segments used for gradients.Each point data is the average of the sample.
As can be seen from Figure 11, in the three mainline lanes, the vehicle speeds in the inner and middle lanes change less.Vehicle speed in the middle lane in the acceleration zone decreased, mainly due to the infuence of the outer lane and converging vehicles.On the entrance ramp, both cars and trucks driving past the confuence point decelerate.After a period of acceleration, speeds stabilise at about the same level as the outer lanes.
Compared to interchanges, the service area ramps have diferent speed trends for diferent vehicle types.As shown in Figure 12, vehicles in the outer lanes accelerate after passing the confuence point.Vehicle speed change diferently on service area merge zone ramps and interchange merge zone ramps.In this case, vehicle speed on the interchange ramps is decreasing, while vehicle speed in the service area is increasing.

Efects of Diferent Vehicle Types on Lane Change Behavior.
Te merge area merge section was divided into 4 segments of 50 m.Te sample consisted of 179 cars and 54 trucks merging from the ramp to the main road.In this study, the trajectories of vehicles merging into the main road from the ramps in the sample data were recorded.Te trajectories are shown in Figure 13.
According to the statistics, the lane change positions of cars are concentrated in the interval of 0, 50 m, accounting for 76% of the total number of cars.Te lane change positions of trucks are concentrated in the interval of 50-150 m, accounting for 85% of the total number of trucks.Infuenced by body length, vehicle acceleration, and other factors, we can see that the overall position of cars changing lanes to the ramp is more forward, from the lane change trajectory diagram, the trajectory of trucks changing lanes is more smooth.

Driving Style Clustering
Driving style not only afects the interaction behavior between vehicles in diferent trafc scenarios, but also has a subtle impact on the decision-making, planning, and control of future autonomous vehicles.Diferent driving styles may perform diferently in diferent trafc scenarios.Diferent types of driving styles are also inconsistent in terms of trafc safety hazards.Studying driving styles can to some extent reduce the likelihood of trafc accidents.

Characteristic Variables of Driving Style Clustering.
To cluster driving style, it is frst necessary to obtain features that can represent driving style.Tis study mainly extracted features from the perspective of representing driving safety  Journal of Advanced Transportation

Driving Style Feature Dimensionality Reduction.
Given that there is a certain correlation between the above extracted features, and there is redundant information between the features.If all the features are directly brought into the clustering model processing, it will increase the model training cost.It even afects the comprehensive performance of the model and reduces the accuracy.Terefore, in this paper, the proposed features are frst downscaled before clustering the driving styles.In view of the excellent performance of PCA in this this paper frstly carries out normalisation and PCA dimensionality reduction on the features.Normalisation eliminates the misjudgment of feature importance by the model, which is caused by the nonuniformity of the feature scale.PCA can generate new features through a series of linear transformations, and these new features are principal components.Te principal components are generated based on the importance of the original feature information, i.e., the cumulative contribution rate.In practice, it is usually enough to extract the frst few principal components whose cumulative contribution rate reaches more than 80%.
Based on Python for feature normalisation and PCA dimensionality reduction, the cumulative contribution rate of principal components can be obtained, as shown in Figure 14.It can be seen that the frst three principal components can already represent more than 85% of the original feature information.Terefore, this article selected the frst three principal components for subsequent analysis.

Driving Style Clustering Model.
Clustering algorithms can divide a series of data into meaningful clusters according to certain laws.Combined with the PCA dimensionality reduction results above, it can be found that the results obtained only have features without labels, so unsupervised algorithms can be used for subsequent analysis.Te K-means algorithm is the most commonly used unsupervised classifcation algorithm.Te traditional K-means algorithm uses a random method to draw samples from the sample points as the initial center of mass, which has some limitations.Te improved K-means++ algorithm based on the traditional K-means algorithm can make the initial centers of mass far away from each other, so as to obtain more reliable results than the random initialisation.In this section, the practical efects of the two algorithms are compared and evaluated using evaluation metrics.

Cluster Assessment Indicators.
Compared with the silhouette coefcient, the Calinski Harabasz Score (CH Score) were calculated quickly.It is better for more concentrated samples with clearer "separation boundaries."Te CH score was used to evaluate the clustering efect, generally when the value of the CH score is larger the clustering efect is better.Te equation is as follows: where K is the number of clusters, N is the number of samples, T r (B k ) and T r (W k ) are the traces of the intercluster deviation matrix (i.e., covariance matrix) and the traces of the intraclass deviation matrix, respectively, and W k and B k are computed as shown in the following equations:     where C q is the sample points in a particular cluster q, c is the center of all datasets, and n q is the total number of samples in a particular cluster q.As can be seen from the CH score, clustering is best when the number of clusters is 3. On this basis, the dimensionality reduction feature data were clustered.Te results after dimensionality reduction were clustered using k-means and k-means++ clustering, as shown in Figure 16, Te axes in the fgure are the coordinates of the center of mass points of the eigenvalues.
Compared to manually labeled driving styles, the accuracy rate of k-means clustering results is 86.45%.Te accuracy rate of k-means++ is 97.76%.It can be seen that there was a signifcant diference between the driver style clustering between the service area and the interchange and merging area.Te driver style classifcation in the service area confuence area was more balanced, and the driver style of the interchange merging area was more aggressive.

Conclusion
Tis paper analyzed the driving behavior patterns of vehicles in two types of merging zones.Te results are summarized as follows: (1) Compared to interchange merging areas, vehicle speed distribution was more decentralized across the lanes in the service area.In interchange merging area, about 80% of vehicles were travelling at speeds of ≤60 km/h.In the service area merging area, the average speeds of the vehicles in each lane were concentrated in the area of (85, 120) km/h.Te speed gap was mostly in the range of 20 km/h within the same lane.(2) For two types of merging areas, the acceleration distribution of vehicles showed a trend of high in the middle and low at both ends.Acceleration was mainly distributed within −2∼2 m/s 2 .Vehicles in the interchange merge area accelerate and decelerate sharply more frequently.(3) Vehicle lane change rates were higher in the service area merge area.It exceeded the interchange lane change rate by 21.7%.Te average speed of large vehicles in each lane is lower than that of small vehicles, by about 10%.When conducting vehicle confict warning in the merging area, the monitoring of the service area should be strengthened to further clarify the speed limits and lane change behavioral specifcations for diferent vehicle types near the service area.(4) When small vehicles merge into the main road from the ramp, most of them chose to merge in the range of 0-50 m.However, most of the large vehicles chose to merge in the range of 50-150 m.In the next study of vehicle confict locations, the focus could be on enhancing the characterization of vehicle driving in these two areas.(5) Tere were clear diferences in the style of drivers in diferent types of merging areas.Compared to the service area, drivers in the interchange merge area have more aggressive driving styles.
Compared with existing studies, this study considers a variety of scenarios and describes more specifcally the driving patterns of vehicles in the merging area of interchanges and service area ramps.Te behavioral patterns of drivers in diferent merging zones can provide a theoretical basis for planning highway merging zones, such as speed limit setting, road structure, and trafc confict warning.Further, corresponding management methods are proposed for diferent types of merging zones or diferent car models.It is of great signifcance to improve road trafc safety.In this study, the extracted feature data types were relatively single due to the limitation of equipment conditions.To improve the accuracy of the driving style clustering model.Te next step will take into account the physiological and psychological characteristics of drivers.Combined with radar point cloud data of vehicles on real roads, the driver's behavioral patterns will be analyzed in depth.

2 Journal of Advanced Transportation 2 . 2 . 2 .
Vehicle Tracking Detection Model Based on Deep SORT.Te Deep SORT model can track the relevant attributes of the target of the current frame, including the center point x coordinate, center point y coordinate, target serial no., vehicle type, vehicle pixel width, etc.

Figure 6 :Figure 7 :Figure 8 :Figure 9 :
Figure 6: (a) Te trafc volume distribution of each lane in the merging area of the interchange; (b) the trafc volume distribution of each lane in the merging area of the service area.(a) Interchange and (b) Rest area.

Figure 10 :Figure 11 :Figure 12 :
Figure 10: Special point and segment distribution in the merging area.

Figure 13 :
Figure 13: Vehicle lane change trajectory of diferent vehicle types in the merging area.(a) Car.(b) Truck.

Figure 14 :
Figure 14: Schematic diagram of PCA cumulative contribution rate.

Figure 15 :
Figure 15: CH assessment scores for diferent numbers of clustered clusters.

Figure 16 :
Figure 16: Clustering results of vehicle driving styles in the merging area.(a) Te clustering efect of interchange merging area; (b) the clustering efect of service area merging area.

6. 5 .
Clustering Results and Analyses.For both k-means and k-means++ algorithms, CH scores were computed for different numbers of clusters.Te results are shown in Figure 15.

Table 2 :
Statistics on the frequency of vehicle lane change in diferent areas.

Table 3 :
Lane change rates for entrance ramps in each region.
to cluster driving style.We used a sample of 1,000 vehicles in an interchange versus 1,000 vehicles in a service area.Te main extracted features include speed, acceleration, and variable acceleration.Te commonly used statistical values of these indicators are calculated as features, and a total of 6 features are obtained, as shown in Table4.

Table 4 :
Characteristics of driving style clustering.
3Standard deviation of absolute value of variable acceleration