Analysis of driving trajectories is one of the heated topics in the field of traffic safety. Mauriello et al. explored driving behavior on horizontal curves of two-lane rural highways based on a driving simulator experiment. Six major classes and twenty-one sub-classes were defined [6]. In addition, their study pointed out that driving trajectories are a promising surrogate measure of safety as highlighted by the correlation between the trajectories identified as dangerous and the radii of the curves. Another study analyzed driving trajectories with little lane discipline [7]. By studying lateral distance-keeping behavior, lane keeping behavior and the lateral behavior of vehicles, the research explored the non-lane-based behavior of traffic. The researchers also compared lateral clearances of various vehicle types in the study.

Lane keeping and lane changing are maneuvers that commonly happen in driving trajectories. Galante et al. studied lane keeping by analyzing the mean and standard deviation of lateral position [8]. They demonstrated several effective treatments which were strongly advised to be tested on the real road. Wonho Suh et al. developed an index related to lane keeping named Deviation of Lateral Placement (DLP) which represents a driver’s steering behavior along a given section of highway [9]. The index was able to demonstrate a vehicle’s overall lateral stability. Among the studies on lane changing, some researchers have used lateral position and vehicle heading angle to study lane variation [10]. They created a new ADAS system to support drivers when overtaking cyclists.

Lane changing maneuvers’ recognition has been used to study traffic safety for many years. Many studies have focused on this area to test new theories, improve the accuracy of lane change identification and reduce traffic accidents. The Gipps decision method model based on logic rules was the first to be proposed [11]. However, this model was too idealistic to be applied in practice. Researchers have optimized and improved the classical model from different aspects and built new models, such as MITSIM and CORSIM [12,13]. Other models have also been proposed by researchers. One group of researchers built an integrated lane changing model combining a sine function and a constant velocity migration function [14]. A multilevel analysis model of vehicles’ longitudinal and lateral acceleration during lane change was established as well [15]. Other researchers set up a safe lane-change model based on a quintic polynomial trajectory [16].

With the development of technology, researchers used new methods to analyze lane change maneuvers. Machine learning was proven to be useful for estimating, classifying and predicting lane changing maneuvers. Using steering wheel angle and angular velocity as inputs to an HMM model, researchers developed an algorithm for recognizing lane changes [17]. Based on their method, the accuracy of lane change left (LCL), lane change right (LCR) and lane keeping (LK) classifications was 84%, 88% and 94%, respectively. Other researchers established a decision model of an autonomous lane change by analyzing the influencing factors of autonomous lane change [18]. Additionally, with the help of a support vector machine (SVM), they solved the multi-parameter and nonlinear problems. Another RVM model used 200 sensor signals [19]. It was able to classify a driver’s intentions three seconds before the actual lane change occurred. With long short-term memory networks (LSTM) and deep belief networks (DBN), researchers modeled both lane keeping and lane changing [20]. Compared with the classical model, this model could accurately estimate both lane keeping and lane change maneuvers. The model had higher lane-change prediction accuracy as well. Though all these methods have achieved good results, none of them have used a distinct set of physical data as input.

Meanwhile, vehicle status data are also derived from a variety of sources. Researchers used different methods to obtain naturalistic driving data (NDD). One group of researchers developed a data collection platform called POSS-V (PKU Omni Smart Sensing - Vehicle) to collect real human driving data in urban street scenarios [21]. The POSS-V included GPS/IMU, a steering angle sensor and a panoramic camera. Other researchers made use of UAVs to collect natural driving-track data [22]. A fixed-base driving simulator was also utilized by researchers [23]. Compared with collecting data individually or using virtual data, it is more convenient and efficient to use datasets from the real world. Both NGSIM and HighD [24], two high-resolution natural vehicle-trajectory datasets, had a good effect on machine learning [25,26]. However, among the lane-changing maneuver recognition models which were based on a dataset, a distinct set of physical data was not used as the input for machine learning.