A method for modeling vehicle car following at mountainous highway-village intersection considering driving style

Abstract

The present application relates to a kind of mountainous road village-crossing intersection vehicle following modeling method considering driving style, belong to mountainous road village-crossing intersection vehicle following modeling technical field.The method includes: obtaining mountainous road village-crossing intersection vehicle trajectory data and extracting following segment according to preset rule, the speed of following vehicle, the speed difference of front and rear vehicles, following distance and following time interval are calculated;Following segment is extracted based on driving style representation parameter and is clustered to obtain following driving style category, combine the speed of following vehicle, the speed difference of front and rear vehicles, following distance and following time interval as the following state variable, input deep reinforcement learning model based on reward function with the following state variable and train, obtain mountainous road village-crossing intersection vehicle following model, for the following control amount or next time motion state of target vehicle.Output.The present application aims to solve the technical problems that the prior art is difficult to simultaneously represent complex mountain intersection scene characteristics and driving style difference.

Description

Technical Field

[0001] This invention relates to a method for modeling vehicle following at intersections of mountain roads through villages and towns, taking into account driving styles, and belongs to the technical field of vehicle following modeling at intersections of mountain roads through villages and towns. Background Technology

[0002] Mountainous highways crossing villages and towns possess the dual characteristics of ordinary highway intersections and village / township traffic environments. They are often accompanied by pedestrians and non-motorized vehicles randomly crossing, frequent roadside activity, limited visibility, complex road alignments, and significant fluctuations in vehicle speed. Due to these factors, vehicles in such scenarios need to frequently accelerate, decelerate, and adjust their following distances, making following behavior significantly unstable and risky.

[0003] Most existing car-following models are designed for urban roads, highways, or regular traffic environments, typically assuming weak scene disturbances or relatively consistent driver behavior. However, the complex terrain and environmental factors of mountain roads affect driver decision-making processes and vehicle driving characteristics. Traditional car-following models struggle to effectively depict the complex driving decision-making processes under multi-agent interactions and cannot dynamically adapt to the unique characteristics of mountain roads intersecting with villages and towns. Especially when different drivers exhibit significant differences in driving styles, such as aggressive, steady, or conservative, traditional models fail to accurately portray heterogeneous car-following decision-making processes within the same traffic scenario.

[0004] With the development of drone video acquisition, trajectory extraction, and deep reinforcement learning technologies, it has become possible to construct vehicle car-following models for complex scenarios using natural driving trajectory data. However, a complete methodology is still lacking for extracting high-quality car-following segments from measured trajectories at intersections of mountain roads through villages and towns, identifying car-following driving styles, and effectively integrating them into car-following decision-making models.

[0005] Therefore, it is necessary to propose a vehicle following modeling method for mountainous road intersections in villages and towns, while also considering differences in driving styles, in order to improve the model's fitting ability, interpretability, and engineering application value in complex traffic environments. Summary of the Invention

[0006] The purpose of this invention is to provide a vehicle following modeling method for mountain roads through villages and towns that takes into account driving styles, aiming to solve the technical problem that existing technologies cannot simultaneously represent the characteristics of complex mountain intersection scenarios and differences in driving styles.

[0007] To achieve the above objectives, the technical solution of the present invention is: a method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, comprising the following steps:

[0008] S1: Collect video data of vehicles driving at intersections of mountain roads through villages and towns, and extract vehicle trajectory data;

[0009] S2: Extract following segments from the vehicle trajectory data according to preset rules, and calculate the following vehicle speed, the speed difference between the front and rear vehicles, the following distance, and the following time interval;

[0010] S3: Extract driving style characterization parameters based on the car-following segments and perform clustering to obtain car-following driving style categories; wherein, the driving style characterization parameters include minimum headway and minimum collision time;

[0011] S4: Construct a reward function, using the following driving style category, the following vehicle speed, the speed difference between the front and rear vehicles, the following distance, and the following time interval as following state variables. Based on the reward function, input the following state variables into a deep reinforcement learning model for training, to obtain a vehicle following model at mountainous road intersections through villages and towns. Use the vehicle following model at mountainous road intersections through villages and towns to output the following control quantity or the motion state of the target vehicle at the next moment; wherein, the reward function includes a safety reward function, an efficiency reward function, and a comfort reward function.

[0012] Optionally, the preset rule is specifically as follows:

[0013] The two vehicles are in the same lane, and neither the vehicle in front nor the vehicle following behind has changed direction;

[0014] Cars were chosen as the subject of the car-following study.

[0015] The longitudinal distance between the front and rear vehicles is less than a preset distance threshold;

[0016] The duration of the follow-car cycle exceeds a preset time threshold;

[0017] The speed of the following vehicle exceeds a preset speed threshold.

[0018] The affected area is located within the intersection of mountain roads through villages and towns.

[0019] Optionally, S3 specifically includes:

[0020] S31: Minimum headway (THW) and minimum collision time (TTC) are used as parameters to characterize driving style.

[0021] S32: Clustering is performed using the K-means algorithm, expressed as follows:

[0022]

[0023] In the formula, Let S represent the minimization of the sum of squared distances within the total cluster, where S is the number of driving style categories, X is the dataset, and k is the number of style clusters. The eigenvector of the j-th carousel segment. For the i-th cluster, The minimum headway for the j-th following segment. Let be the minimum collision time for the j-th race segment. Let i be the centroid of the i-th cluster in the time-distance dimension at the front of the vehicle. Let be the centroid of the i-th cluster in the collision time dimension.

[0024] Optionally, the security reward function is specifically:

[0025]

[0026] In the formula, Let i be the safety reward function for the i-th car-following driving style, where i = 1, 2, 3, corresponding to the conservative, steady, and aggressive car-following driving styles, respectively. Let c be the minimum collision time (TTC) boundary for the i-th following driving style, where c is a constant.

[0027] Optionally, the efficiency reward function is specifically:

[0028]

[0029] In the formula, Let be the efficiency reward function for the i-th car-following driving style. Let be the fitting constant. , The minimum headway under the i-th following style The parameters of the log-normal distribution function.

[0030] Optionally, the comfort reward function is specifically:

[0031]

[0032] In the formula, Let the comfort reward function be at time k. The jerk at time k Value, C is the base value, where jerk is... The expression is:

[0033]

[0034] In the formula, Let k be the rate of change of acceleration at time k. , Let be the accelerations of the vehicle at times k and k-1, respectively. Let k be the time interval between time k and time k-1.

[0035] Optionally, the deep reinforcement learning model employs the dual-delay deep deterministic policy gradient TD3 framework.

[0036] The beneficial effects of this invention are as follows: First, modeling based on measured UAV trajectory data can more realistically reflect the natural car-following behavior at intersections of mountain roads through villages and towns; second, explicitly introducing car-following driving style into the model training process can enhance the ability to characterize heterogeneous driving behaviors; third, using deep reinforcement learning to construct a stylized car-following model can balance safety, efficiency, and comfort, providing support for traffic safety analysis, micro-simulation, and intelligent driving control on mountain roads. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of data acquisition according to the present invention; Figure 3 This is a diagram showing the results of the car-following segment extraction according to the present invention; Figure 4 This is a clustering result diagram of the driving style following method of the present invention; Figure 5 This is the probability distribution diagram of the minimum headway of the vehicle according to the present invention; Figure 6 This is a probability distribution diagram of the minimum collision time in this invention; Figure 7 This is the framework result of the vehicle car-following model system of the present invention. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0039] Example 1: As Figure 1 As shown, a method for modeling vehicle following at intersections in mountainous areas that consider driving styles at village crossings includes the following steps:

[0040] S1: Collect video data of vehicles driving at intersections of mountain roads through villages and towns, and extract vehicle trajectory data;

[0041] Optionally, this embodiment uses a drone to conduct non-contact data collection over intersections of typical mountain roads through villages and towns to obtain vehicle driving video data. The data collection targets include the continuous following behavior of cars within the intersection and its adjacent road sections, specifically:

[0042] S11: Data Acquisition. Collecting video footage of natural vehicle movement at intersections along mountain roads through villages and towns using drones;

[0043] S12: Vehicle Calibration. Set reference coordinates, scale, and tracking step size to label the preceding and following vehicles as cars based on their salient pixel features;

[0044] S13: Automatic vehicle trajectory tracking. Establishes a dynamic tracking target and automatically identifies and calculates the trajectory information (coordinates, speed, travel distance, etc.) of the calibrated vehicle at a specific tracking step length.

[0045] S2: Extract following segments from the vehicle trajectory data that conform to the operating characteristics of mountain roads passing through villages and towns according to preset rules, and calculate the following vehicle speed, the speed difference between the front and rear vehicles, the following distance, and the following time distance.

[0046] Optionally, within the intersection area, data collection segments exhibiting car-following behavior are initially extracted manually, and invalid data collection segments are filtered out. The specific preset rules are as follows:

[0047] The two vehicles are in the same lane, and neither the vehicle in front nor the vehicle following behind has changed direction;

[0048] Cars were chosen as the research object for car-following to avoid the interference of different vehicle types on driving style and car-following model characteristics;

[0049] The longitudinal distance between the front and rear vehicles is less than a preset distance threshold;

[0050] The duration of the follow-car cycle exceeds a preset time threshold;

[0051] The speed of the following vehicle exceeds a preset speed threshold.

[0052] The affected area is located within the intersection of mountain roads through villages and towns.

[0053] Optionally, in this embodiment, the longitudinal distance between the front and rear vehicles is less than 100m; the following duration is greater than 10s; and the speed of the following vehicle is greater than 15km / h.

[0054] S3: Extract driving style characterization parameters based on the car-following segments and perform clustering to obtain car-following driving style categories; wherein, the driving style characterization parameters include minimum headway and minimum collision time;

[0055] S31: Minimum headway (THW) and minimum collision time (TTC) are used as parameters to characterize driving style.

[0056] S32: Clustering is performed using the K-means algorithm, expressed as follows:

[0057]

[0058] In the formula, Let S represent the minimization of the sum of squared distances within the total cluster, where S is the number of driving style categories, X is the dataset, and k is the number of style clusters. The eigenvector of the j-th carousel segment. For the i-th cluster, The minimum headway for the j-th following segment. Let be the minimum collision time for the j-th race segment. Let i be the centroid of the i-th cluster in the time-distance dimension at the front of the vehicle. Let be the centroid of the i-th cluster in the collision time dimension.

[0059] S4: Construct a reward function, using the following driving style category, the following vehicle speed, the speed difference between the front and rear vehicles, the following distance, and the following time interval as following state variables. Based on the reward function, input the following state variables into a deep reinforcement learning model for training. Consider the following model of vehicles at intersections in mountainous areas through villages and towns, and use the following model of vehicles at intersections in mountainous areas through villages and towns to output the following control quantity or the motion state of the target vehicle at the next moment. The reward function includes a safety reward function, an efficiency reward function, and a comfort reward function, so that the vehicle following model simultaneously meets the requirements of safety, efficiency, and comfort.

[0060] Optionally, a vehicle car-following model based on deep reinforcement learning is constructed. The input of the vehicle car-following model is the speed of the following vehicle, the speed difference, the following distance, the following time distance, and the following driving style category. The output is the longitudinal control quantity of the target vehicle, specifically the target acceleration and the target speed.

[0061] Optionally, in this embodiment, differentiated reward parameters based on driving style are adopted: different minimum TTC boundaries and different THW distribution fitting parameters are set for aggressive, moderate, and conservative drivers respectively, so that the reward function matches the risk tolerance of drivers with different driving styles.

[0062] Optionally, the safety of car-following behavior is measured using TTC, and the safety reward function is specifically:

[0063]

[0064] In the formula, Let i be the safety reward function for the i-th car-following driving style, where i = 1, 2, 3, corresponding to the conservative, steady, and aggressive car-following driving styles, respectively. Let c be the minimum collision time (TTC) boundary for the i-th following driving style, where c is a constant, and in this embodiment it is taken as 0.3.

[0065] Optionally, the performance in the car-following state is measured by THW, and the efficiency reward function is specifically:

[0066]

[0067] In the formula, Let be the efficiency reward function for the i-th car-following driving style. Let be the fitting constant. , The minimum headway under the i-th following style The parameters of the log-normal distribution function.

[0068] Optionally, the comfort reward function is specifically:

[0069]

[0070] In the formula, Let the comfort reward function be at time k. The jerk at time k The value, C, is the base value used to adjust the numerical range of the reward function, where accelerometer... The expression is:

[0071]

[0072] In the formula, Let k be the rate of change of acceleration at time k. , Let be the accelerations of the vehicle at times k and k-1, respectively. Let k be the time interval between time k and time k-1.

[0073] Optionally, in this embodiment, the deep reinforcement learning model adopts the dual-delay deep deterministic policy gradient TD3 framework. Samples of different car-following driving styles are input into the deep reinforcement learning model for training to obtain car-following policies adapted to different car-following driving styles. Specifically, the training of the deep reinforcement learning model involves:

[0074] Based on the Actor-Critic architecture and experience replay mechanism, the following state, action, and reward samples labeled with driving style categories are stored in the experience replay pool. The dual-critic network is iteratively updated to evaluate the state-action value, and the Actor network is updated with a delay to output the optimal following control quantity. The process is iterated until convergence with the goal of maximizing the cumulative reward, and finally, a following decision strategy adapted to different driving styles is obtained.

[0075] Understandably, existing car-following modeling is generally geared towards regularized scenarios such as highways and urban roads. However, for the special scenario of mountain roads passing through villages and towns, which combines the characteristics of highway traffic with the complex interference of villages and towns (pedestrians and non-motorized vehicles crossing, limited visibility, and large speed fluctuations), this embodiment proposes a complete and feasible technical process from data collection → car-following segment selection → driving style recognition → stylized car-following model construction.

[0076] The invention will be further illustrated by a specific example below.

[0077] Step 1: Data collection and trajectory extraction.

[0078] Step 1.1: Based on factors such as traffic flow, town and village size, and roadside environment, and taking into account the fit between the surveyed road section and the research needs, this embodiment selects a certain township intersection as the research road section. Its main road traffic flow is 120 vehicles / hour, the intersection slope is about 0.45°, and there is a deceleration sign 150m before the main road entrance. If there is no lateral interference, passing vehicles generally pass through the intersection area at a constant speed or quickly.

[0079] Step 1.2: As Figure 2 As shown, drones were deployed to take aerial photos of the intersection and its upstream and downstream affected areas, collecting video data on natural vehicle movement. Clear weather and unobstructed times were preferred during data collection to ensure video clarity and trajectory continuity. The data collection primarily focused on car following behavior, and the videos were designed to identify at least the vehicle's position, direction of travel, relationships between adjacent vehicles, and lateral interference around the intersection.

[0080] Step 1.3: Using Tracker video analysis software, establish the conversion relationship between pixel coordinates and actual road coordinates, and extract vehicle trajectory data frame by frame. The trajectory data includes at least basic information such as time, longitudinal position, and lateral position. After frame-by-frame debugging, the optimal frame-by-frame extraction step size was determined to be 5, achieving 5 frames to mark a single trajectory point, meaning the time granularity of the vehicle trajectory data is 0.167 seconds. Initially, 523 sets of vehicle trajectory data were extracted.

[0081] Step 1.4: Perform error analysis and quality control on the extracted original trajectory, remove obvious outliers, severely missing segments and discontinuous segments; then use a one-dimensional convolution smoothing method to reduce noise in the longitudinal displacement sequence and reconstruct the vehicle speed and acceleration sequence to obtain high-quality trajectory data that can be used for subsequent analysis.

[0082] Step 2: Extracting the carousel segments.

[0083] Step 2.1: Based on the characteristics of mountain roads crossing village intersections, formulate rules for extracting car-following segments. Specifically: during car-following, both vehicles are in the same lane, and neither the leading nor following vehicle changes direction; cars are selected as the car-following study object; the longitudinal distance between the leading and trailing vehicles is less than 100m; the car-following duration is greater than 10s; and the speed of the following vehicle is greater than 15km / h. Figure 3 As shown, based on the above extraction rules, a total of 410 car-following segments were extracted, and the total trajectory data was 38,864, including 234 uphill segments and 176 downhill segments.

[0084] Step 2.2: Calculate the following vehicle speed, the speed difference between the vehicles in front and behind, the following distance, and the following time interval to describe the dynamic following characteristics within the segment.

[0085] Step 3: Follow-up driving style recognition.

[0086] Step 3.1: Extract car-following driving style parameters with safety characteristics from the car-following segments, and use minimum headway (THW) and minimum collision time (TTC) as style classification indicators.

[0087] Step 3.2: As Figure 4 As shown, the K-means algorithm is used to cluster the samples, classifying drivers or following segments into aggressive, robust, and conservative types.

[0088] Step 3.3: As Figure 5 and Figure 6 As shown, the probability distributions of THW and TTC for each category were statistically analyzed. Using the 10% percentile of the cumulative TTC probability as the minimum TTC boundary for the safety reward function, the minimum TTC time was 9.3s for the aggressive following driving style, 18.1s for the conservative following driving style, and 26.8s for the moderate following driving style. Meanwhile, the optimal THW values ​​for the aggressive, conservative, and moderate following driving styles were 2.13s, 2.65s, and 3.26s, respectively.

[0089] Step 4: Carry-following model construction.

[0090] Step 4.1: Construct a vehicle car-following model based on deep reinforcement learning, taking the following vehicle speed, speed difference, following distance, following time distance, and driving style category as state inputs, and outputting the longitudinal control quantity of the target vehicle.

[0091] Step 4.2: As Figure 7 As shown, the model employs a dual-delay deep deterministic strategy gradient TD3 framework. The reward function comprehensively considers safety, driving efficiency, and comfort indicators, enabling the model to simultaneously meet the requirements of safety, efficiency, and comfort.

[0092] Step 4.3: Input samples of different car-following driving styles into the model for training to obtain stylized car-following strategies adapted to different driving styles.

[0093] Step 4.4: After training, input the real-time or simulated vehicle following state into the model, and it will output the next moment's motion state of the target vehicle or longitudinal control commands. This model can be applied to vehicle following behavior prediction, micro-traffic simulation, driver assistance control, and safety risk analysis in mountainous road intersections through villages and towns.

[0094] In summary, this invention collects natural driving video data of vehicles at intersections along mountain roads through villages and towns, and extracts vehicle trajectory data. It extracts following segments that fit the research scenario based on the longitudinal positional relationship between preceding and following vehicles, speed differences, following distance, and following duration. Based on safety representation parameters such as minimum headway and minimum collision time, it clusters following driving styles to obtain different driving style categories. The driving style categories and following state variables are jointly input into a deep reinforcement learning model for training, constructing a vehicle following model suitable for mountain road intersection scenarios. The trained model outputs the following control quantity or the next moment's motion state of the target vehicle, achieving vehicle following prediction and decision-making in mountain road intersection scenarios. This invention can incorporate the characteristics of complex mountain intersection scenarios and differences in driver style into the vehicle following modeling process, improving the model's ability to represent heterogeneous following behaviors and adapt to complex scenarios. It can provide technical support for traffic safety analysis, microscopic traffic simulation, and intelligent driving assistance control on mountain roads.

[0095] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, characterized in that... The method includes the following steps: S1: Collect video data of vehicles driving at intersections of mountain roads through villages and towns, and extract vehicle trajectory data; S2: Extract following segments from the vehicle trajectory data according to preset rules, and calculate the following vehicle speed, the speed difference between the front and rear vehicles, the following distance, and the following time interval; S3: Extract driving style characterization parameters based on the car-following segments and perform clustering to obtain car-following driving style categories; wherein, the driving style characterization parameters include minimum headway and minimum collision time; S4: Construct a reward function, using the following driving style category, the following vehicle speed, the speed difference between the front and rear vehicles, the following distance, and the following time interval as following state variables. Based on the reward function, input the following state variables into a deep reinforcement learning model for training, to obtain a vehicle following model at mountainous road intersections through villages and towns. Use the vehicle following model at mountainous road intersections through villages and towns to output the following control quantity or the motion state of the target vehicle at the next moment; wherein, the reward function includes a safety reward function, an efficiency reward function, and a comfort reward function.

2. The method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, as described in claim 1, is characterized in that... The preset rules are specifically as follows: The two vehicles are in the same lane, and neither the vehicle in front nor the vehicle following behind has changed direction; Cars were chosen as the subject of the car-following study. The longitudinal distance between the front and rear vehicles is less than a preset distance threshold; The duration of the follow-car cycle exceeds a preset time threshold; The speed of the following vehicle exceeds a preset speed threshold. The affected area is located within the intersection of mountain roads through villages and towns.

3. The method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, as described in claim 1, is characterized in that... Specifically, S3 is: S31: Minimum headway (THW) and minimum collision time (TTC) are used as parameters to characterize driving style. S32: Clustering is performed using the K-means algorithm, expressed as follows: ； In the formula, Let S represent the minimization of the sum of squared distances within the total cluster, where S is the number of driving style categories, X is the dataset, and k is the number of style clusters. The eigenvector of the j-th carousel segment. For the i-th cluster, The minimum headway for the j-th following segment. Let be the minimum collision time for the j-th race segment. Let i be the centroid of the i-th cluster in the time-distance dimension at the front of the vehicle. Let be the centroid of the i-th cluster in the collision time dimension.

4. The method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, as described in claim 1, is characterized in that... The security reward function is specifically as follows: ； In the formula, Let i be the safety reward function for the i-th car-following driving style, where i = 1, 2, 3, corresponding to the conservative, steady, and aggressive car-following driving styles, respectively. Let c be the minimum collision time (TTC) boundary for the i-th following driving style, where c is a constant.

5. A method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, as described in claim 1, characterized in that... The efficiency reward function is specifically as follows: ； In the formula, Let be the efficiency reward function for the i-th car-following driving style. Let be the fitting constant. , The minimum headway under the i-th car-following style The parameters of the log-normal distribution function.

6. The method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, as described in claim 1, is characterized in that... The comfort reward function is specifically as follows: ； In the formula, Let the comfort reward function be at time k. The jerk at time k Value, C is the base value, where jerk is... The expression is: ； In the formula, Let k be the rate of change of acceleration at time k. , Let be the accelerations of the vehicle at times k and k-1, respectively. Let k be the time interval between time k and time k-1.

7. A method for modeling vehicle following at intersections of mountain roads through villages and towns, considering driving styles, as described in claim 1, is characterized in that... The deep reinforcement learning model adopts the dual-delay deep deterministic policy gradient TD3 framework.

Images