A lane global car-following pair matching and extracting method for a video track dataset

Abstract

A lane global car-following pair matching and extracting method of a video track dataset, the vehicle track dataset contains the track point information of each vehicle at each time on the collected road, and has an important role in the research of both macroscopic traffic phenomena and microscopic driving behaviors such as car-following and lane-changing. The method aims to establish a standardized car-following pair matching and extracting method, which can detect and classify the driving behaviors of each track in the track dataset, and can efficiently and quickly identify the car-following pairs in the car-following state within the length range of the entire lane, thereby laying a foundation for in-depth research on driving behaviors.

Description

Technical Field

[0001] This invention relates to the field of microscopic traffic flow data processing, and in particular to a method for matching and extracting lane-wide car-following pairs based on trajectory datasets. Background Technology

[0002] Car-following behavior is the most basic micro-driving behavior, describing the interaction between adjacent vehicles in a convoy traveling on a one-way street with overtaking restrictions. The establishment of car-following models is fundamental to traffic flow theory. Traffic flow research focuses on analyzing the interrelationships between elements such as people, vehicles, roads, and the environment; it is a discipline based on empirical research and requires high-precision real-world data support. [1] Vehicle trajectory datasets, containing trajectory point information for every vehicle on the collected roads at every moment, are crucial for studying both macroscopic traffic phenomena and microscopic driving behaviors such as following and lane changing. Currently, the most widely used trajectory dataset is perhaps NGSIM (Next Generation Simulation program dataset), collected from US highways and urban roads using cameras mounted high on roads, dating back to 2006. However, with the advancement of research, NGSIM, due to limitations in collection time, accuracy, and scene constraints, can no longer meet the needs of more diverse research. In recent years, with the development of data collection technologies such as drones and vehicle-mounted devices, more and more traffic trajectory datasets with richer scenarios have been released as open source, such as the highD dataset collected from German highways in 2018, and the Zen Traffic Data dataset collected from the Hanshin Expressway in Japan in the same year. These datasets may be in the form of videos or images. To conduct traffic flow research, researchers first need to extract vehicle trajectory information from videos and images using various image processing techniques. [2] A framework for vehicle trajectory extraction is proposed, which can extract vehicle trajectory data from video data collected by UAVs and perform denoising operations; [3]Based on deep learning methods, video calibration, vehicle detection and tracking, lane recognition, and vehicle motion feature calculation are performed on the original video data to obtain a high-precision HIGH-SIM (High-Granularity Highway Simulation) dataset. Although these trajectory extraction algorithms are becoming increasingly sophisticated and the accuracy of the extracted trajectory data is improving, many still cannot be directly applied to the study of the most basic driving behavior—car-following behavior. Taking the Zen Traffic Data dataset as an example, each vehicle entry lacks information about the vehicles in front and behind. Because vehicles engage in merging, lane changing, and overtaking behaviors, this significantly interferes with the extraction of car-following behavior. Furthermore, trajectory information obtained directly through deep learning methods sometimes does not accurately reflect actual driving behavior. For instance, relying solely on the "Traffic_lane" field in this dataset to determine lane changing behavior results in a minimum lane change time of only 0.2 seconds.

[0003] In summary, it is believed that scholars will increasingly use drones to collect trajectory data in the future. However, such datasets will encounter the same problems as the Zen Traffic Data dataset. Based on this, this method aims to establish a standardized method for matching and extracting car-following pairs. For the original trajectory dataset mentioned above, the method detects and classifies the driving behaviors present in each trajectory, and can efficiently and quickly identify car-following pairs in a car-following state within the entire lane length range, thus laying the foundation for in-depth research on driving behavior. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for matching and extracting lane-wide car-following pairs in a video trajectory dataset.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for lane-wide car-following pair matching and extraction in video trajectory datasets, comprising the following steps:

[0007] S1. Identify the lane where the extracted car-following pair is located, extract the trajectory dataset EE (Ever Existed) from the dataset where trajectory points in the dataset once existed in that lane, and identify the driving behavior present in each trajectory, classifying them into: CF (Car-following) class (all trajectory points in this class are in this lane), LC (Lane-changing) class (trajectory points in this class exhibit lane-changing behavior), and OT (Overtaking) class (trajectory points in this class exhibit overtaking behavior); further, the CF class is divided into the FCF class (Full CF) that meets the requirements and the PCF class (Partial CF) that does not meet the requirements, based on whether the trajectory length meets the requirements;

[0008] S2, LC, OT, and PCF trajectories can alter the preceding and following relationships between vehicles on different road segments, thus affecting the extraction of car-following pairs. To eliminate this influence, m equally spaced road markers are designed. The calculation method for m is as follows:

[0009] Extract all overtaking trajectories in the OT group and calculate the shortest overtaking distance d1 among all overtaking trajectories;

[0010] Calculate the shortest trajectory length d2 in group CF;

[0011] Select the smaller value between d1 and d2, and name it D. min m >= L / D min ;

[0012] S3. Calculate the time when each vehicle in class EE passes each road sign, and sort the vehicles that have passed each road sign in chronological order to obtain T1, T2, ... Tm vehicle time sequence tables, and obtain m sets of vehicle time sequence tables.

[0013] S4. In each time series table, check whether two adjacent vehicles are simultaneously in the FCF class. If so, determine that the two vehicles are in a car-following relationship at the corresponding road sign and extract them into the car-following group of the corresponding road sign. m road signs can obtain m different car-following vehicle pair groups.

[0014] S5. Perform intersection operations on each pair of different car-following pairs in group m to finally obtain the pairs of car-following pairs that follow the car-following ...

[0015] Preferably, the overtaking behavior is determined by detecting whether the vehicle changes lanes at a certain moment and then changes back to the original lane at another moment.

[0016] Preferably, the classification criteria for FCF and PCF are that the trajectory starting point is located at the starting position of the lane and the trajectory length is not less than the lane length (or the set length).

[0017] Preferably, the method for extracting the pairs of vehicles in a following state at each road sign is that two consecutive vehicles in the vehicle time series table calculated at that road sign simultaneously belong to the FCF class.

[0018] Preferably, the time interval between the two times is set to be no less than 1 second, then it is determined that there is an overtaking behavior in the trajectory. The 1-second time interval ensures that the method for determining the overtaking behavior is not overly interfered with by abnormal lane data values.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] (1) Existing technologies focus on extracting vehicle trajectory data from original aerial or roadside camera videos using methods such as neural networks. However, most of the extracted trajectory data cannot be directly used for research in traffic engineering. The reason is that the extracted trajectory data contains a lot of abnormal data that does not conform to actual driving behavior, and the driving behavior in the trajectory data has not been further processed from the perspective of traffic engineering.

[0021] (2) From the perspective of traffic engineering, this method focuses on in-depth processing of the original trajectory dataset obtained by the above methods. It not only identifies driving behaviors such as lane changing and overtaking in each trajectory in the trajectory dataset, but also designs a method to extract pairs of vehicles that are in a car-following relationship throughout the entire lane length range, laying the foundation for in-depth research on car-following and lane changing behaviors.

[0022] This invention provides a method for matching and extracting following pairs of vehicles in front and behind in the original trajectory dataset, laying the foundation for building a micro-traffic flow model that is more consistent with actual driving behavior. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall process of the video trajectory dataset lane-wide car-following vehicle pair matching and extraction method of the present invention.

[0024] Figure 2 Zen Traffic data lane diagram Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Note that the following description of the embodiments is merely illustrative and is not intended to limit the application or use of the invention, nor is the invention limited to the following embodiments.

[0026] Example

[0027] like Figure 1 As shown, a method for lane-wide vehicle pair matching and extraction in video trajectory datasets includes the following steps:

[0028] S1. Determine the lane under study. Based on the lane markings, extract all trajectories that have existed in that lane, defining them as EE (Ever Existed) class. Based on the driving behaviors involved, EE class trajectory data is further divided into three categories: Car-Following (CF), Lane-Changing (LC), and Overtaking (OT). CF class can be further divided into FCF class (fully meeting the requirements) and PCF class (not meeting the requirements) based on whether the trajectory length meets the requirements.

[0029] Taking the Zen Traffic data dataset as an example, it is as follows: Figure 2 As shown, this dataset collects vehicle trajectories from segments 4 and 11 of the Hanshin Expressway in Japan. Taking segment 11 as an example, it was collected five times, each time lasting approximately one hour, with each collection corresponding to one file. Segment 11 is approximately 2km long and includes three lanes: 3 represents the entrance ramp, 1 represents the middle lane, and 2 represents the innermost overtaking lane. The trajectory dataset includes fields such as vehicle ID, trajectory point time, vehicle type, speed, lane (traffic_lane), latitude and longitude, travel distance (kilopost), and vehicle length. From these fields, it can be seen that this trajectory dataset alone cannot reveal the specific driving behaviors (lane changes, overtaking, etc.) for each trajectory, nor can it determine the preceding and following driving relationships between vehicles. Therefore, it cannot be directly applied to traffic behavior research. To address this, this method is implemented based on the data visualization method in file 1 of segment 11 of this dataset.

[0030] To eliminate the impact of frequent merging from lane 3 on normally driving vehicles, this embodiment takes extracting the trajectory of vehicles that have previously traveled in lane 2 as an example. The "traffic_lane" field in the trajectory dataset is used for identification; if this field was once equal to 2 in a trajectory, it is extracted, ultimately obtaining the EE class dataset. In the EE class dataset:

[0031] (1) CF Class Extraction: If the “traffic_lane” field is always 2, then the trajectory is of CF class;

[0032] (2) LC class extraction: If the initial time period of the “traffic_lane” field is 2 and the end time period is not equal to 2, or the initial time period is not equal to 2 and the end time period is 2, then it is determined to be LC class;

[0033] (3) OT Class Extraction: If it is determined by “traffic_lane” that the vehicle has undergone two lane changes and both lane changes are related to lane 2, then the trajectory is determined to be OT class. However, due to the fact that some data in the “traffic_lane” field results in the extracted OT class overtaking time being too small (0.2s), which is obviously inconsistent with actual driving behavior, a 1s overtaking threshold is set for the OT class overtaking time. Only overtaking time intervals greater than 1s are counted.

[0034] (4) FCF and PCF Class Discrimination: Further judgment is made on CF class trajectories. The judgment conditions are whether the initial position of the trajectory is the starting displacement of the road segment and whether the trajectory length reaches the length of the road segment. If both requirements are met, it is judged as FCF class; if the requirements are not met, it is judged as PCF class.

[0035] It should be noted that this discrimination method only applies to driving behavior between lane 2 and its adjacent lane 1, and does not apply to driving behavior between lane 3 and lane 1.

[0036] After processing road segment 11 and file 1 in the Zen Traffic data, the number of vehicles of each type is shown in Table 1:

[0037] Table 1 Number of vehicles by type

[0038]

[0039] It should be noted that there are 6 tracks that simultaneously involve lane changing and overtaking. Therefore, the total number of CF, LC, and OT tracks in Table 1 minus 6 is the number of EE tracks.

[0040] S2. Calculate the shortest overtaking distance in the OT category and the shortest trajectory length in the CF category, and take the smaller value D. min To establish an interval of m equally spaced road signs, where m is calculated as m >= L / D min ;

[0041] In the OT category, the shortest overtaking distance is determined by the shortest overtaking time. Calculations show that the shortest overtaking time is the set threshold of 1 second, corresponding to a shortest overtaking distance of 8.2 meters. In the CF category, the shortest trajectory length is 222.7 meters. Therefore, the smaller of the two is m = 2000 / D. min =244

[0042] S3. Calculate the time when each vehicle passes each road sign in class EE to obtain m sets of vehicle time sequence tables.

[0043] Calculate the order in which vehicles travel at each landmark i (i is an integer, and 1 <= i <= m) to obtain the vehicle time sequence table at that landmark. Specifically, the Kilopost field represents the distance between the current displacement of each trajectory point and the starting point. Using this field, detect the datetime value of the field corresponding to the first time each vehicle in the EE class passes landmark i. This value is the time of the corresponding vehicle at landmark i. After sorting the order in which all vehicles pass landmark i, the time sequence table at landmark i is obtained.

[0044] S4. Based on the timetable of each road sign, extract the pairs of vehicles that are in a following state at each road sign to obtain m different groups of following vehicles.

[0045] For each vehicle timetable, sequentially check whether two adjacent vehicles belong to the FCF class. If so, extract them as the vehicle following pairs at that road sign. Taking road sign i as an example, the vehicle IDs at that location, sorted according to the vehicle timetable, are assumed to be 1, 3, 2, 4, 6, and 8. Among them, only vehicles with vehicle IDs 3, 2, 6, and 8 belong to the FCF class. Since vehicle 3 and vehicle 2 are adjacent, 3 and 2 form a following pair at road sign i, and the same applies to 6 and 8. Perform this operation sequentially for each road sign to obtain m sets of not completely identical following pairs.

[0046] S5. Perform intersection operations on each pair of different car-following pairs in group m to finally obtain the pairs of car-following pairs that follow the car-following ...

[0047] After performing an intersection operation on m groups of car-following pairs, 246 car-following pairs with car-following each other along the entire length of the lane were obtained.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0049] The above embodiments are merely illustrative and do not constitute a limitation on the scope of the present invention. These embodiments can also be implemented in various other ways, and various omissions, substitutions, and modifications can be made without departing from the technical spirit of the present invention.

[0050] References

[0051] [1]LI L,JIANG R,HE Z,et al.Trajectory data-based traffic flows studies:a revisit[J],Transportation Research Part C:Emerging Technologies,2020,114:225.

[0052] [2]CHEN

[0053] [3]SHI X,ZHAO D,YAO H,et al.Video-based trajectory extraction withdeep learning for high-granularity highway simulation(high-sim)[J],Communications in transportation research,2021,1:100014.

Claims

1. A method for lane-wide car-following pair matching and extraction in video trajectory datasets, comprising the following steps: S1. Identify the lane where the extracted car-following pairs are located, extract the trajectory dataset EE (Ever Existed) of the trajectory points in the dataset that once existed in that lane, and identify the driving behavior of each trajectory, classifying them into: Car-following (CF), Lane-changing (LC), and Overtaking (OT); CF is further divided into Full CF (FCF) and Partial CF (PCF) based on whether the trajectory length meets the requirements. For S2, LC, OT, and PCF trajectories, design m equally spaced road markers. The method for calculating m is as follows: Extract all overtaking trajectories from the OT group and calculate the shortest overtaking distance among all overtaking trajectories. ; Calculate the shortest trajectory length in the CF group. ; Select as well as The smaller value in is named , ; S3. Calculate the time when each vehicle passes each road sign in the trajectory dataset EE, and sort the vehicles that passed each road sign in chronological order to obtain T1, T2, ..., Tm vehicle time sequence tables, resulting in m sets of vehicle time sequence tables; S4. In each time series table, check whether two adjacent vehicles are simultaneously in the FCF class. If so, determine that the two vehicles are in a car-following relationship at the corresponding road sign and extract them into the car-following group of the corresponding road sign. m road signs can obtain m different car-following vehicle pair groups. S5. Perform intersection operations on each pair of different car-following pairs in group m to finally obtain the pairs of car-following pairs that follow the car-following ...

2. The matching and extraction method as described in claim 1, characterized in that: The method for determining the overtaking behavior is to detect whether the vehicle changes lanes at a certain moment and then changes back to the original lane at another moment.

3. The matching and extraction method as described in claim 1, characterized in that: The classification criteria for FCF and PCF are that the trajectory start point is located at the starting position of the lane, and the trajectory length is not less than the lane length or the set length.

4. The matching and extraction method as described in claim 1, characterized in that: The method for extracting pairs of vehicles in a following state at each road sign is that two consecutive vehicles in the vehicle time series table calculated at that road sign belong to the FCF class at the same time.

5. The matching and extraction method as described in claim 1, characterized in that: If the time interval between two time points is set to be no less than 1 second, it is determined that there is an overtaking behavior on the trajectory. The 1-second time interval ensures that the method for determining overtaking behavior is not overly interfered with by abnormal lane data values.

Images