Vehicle high-speed ramp convergence success rate optimization method

Abstract

The invention relates to the technical field of intelligent vehicle control, in particular to a vehicle high-speed ramp convergence success rate optimization method. In order to solve the inherent defects of a DRL method which purely depends on data: 1, a black box characteristic causes poor interpretability and difficult verification of a decision process; and secondly, the model highly depends on the coverage range of training data, and decision-making errors are easy to occur in training of uncovered long-tail scenes, such as extreme driving behaviors. Due to the factors, the training amount of a pure DRL method is large, and the vehicle afflux success rate is low. The method comprises the steps that mode decision is carried out based on LightGBM and remaining afflux time, and a DRL model is dynamically guided to output a scene-adaptive vehicle afflux track. The random traffic flow convergence success rate is effectively improved.

Description

Technical Field

[0001] This invention relates to a method for merging vehicles onto high-speed ramps, and more particularly to a method for optimizing the success rate of merging vehicles onto high-speed ramps. Background Technology

[0002] Highway ramp merging areas are among the most challenging scenarios in road traffic systems. These areas not only pose a high risk of accidents but also severely restrict overall road traffic efficiency. The core challenge stems from the complexity of dynamic, stochastic traffic flow and the strong interactions between vehicles, placing extremely high demands on the decision-making, planning, and trajectory tracking control capabilities of autonomous vehicles.

[0003] At the decision-making and planning level, existing methods can be mainly divided into traditional modular methods and data-driven deep reinforcement learning methods. Traditional trajectory planning methods, such as sampling-based, optimization-based, and rule-based planning methods, often suffer from limited performance and insufficient flexibility and intelligence when dealing with highly dynamic and highly interactive game environments such as highway ramp merging, due to the difficulty in accurately modeling the intentions of other vehicles and handling complex constraints.

[0004] Deep Reinforcement Learning (DRL) offers a novel solution to these problems due to its powerful environmental awareness and ability to learn autonomously through interaction. However, DRL methods that rely purely on data have inherent drawbacks: first, their black-box nature leads to poor interpretability and difficulty in verification of the decision-making process; second, the model is highly dependent on the coverage of the training data, making it prone to decision-making errors in long-tail scenarios (such as extreme driving behaviors) not covered by training. These factors result in a large training load for pure DRL methods and lead to low vehicle merging success rates. Summary of the Invention

[0005] (1) Technical problems to be solved To address the technical problems of DRL methods that rely solely on data for vehicle trajectory generation, such as high training volume and low vehicle merging success rate, this invention provides a method for optimizing the vehicle merging success rate on highway ramps.

[0006] (2) Technical solution In a first aspect, the present invention provides a method for optimizing the success rate of vehicle merging into high-speed ramps, comprising: I. Decision-making steps for import mode The pre-trained LightGBM machine learning model determines whether the vehicle can merge or not based on its own vehicle status and traffic environment status, and outputs a binary classification signal. It calculates the remaining merging time TTM, and selects TTM=2.5s as the merging mode switching point based on the binary classification signal and TTM value, and outputs the merging mode signal decision. II. Decision-making and planning steps Trajectory planning is performed by dynamically adjusting the reward function of the DRL model based on the incoming pattern signal.

[0007] Preferably, the reward function r of the DRL model is:

[0008] Where, r succ Rewards for successful inflows, r loc For location rewards, r safe For security rewards, r track For MPC tracking error reward, r decision Rewards are given for import behaviors under different decisions.

[0009] Preferably, the successful import reward r succ Location reward r loc Security rewards safe MPC tracking error reward r track They are defined as follows:

[0010]

[0011]

[0012] .

[0013] Preferably, the behavioral reward r decision Dynamically adjust based on the import mode signal decision: When decision=0, r decision =5-5∣y real |; When decision=1, r decision =5-∣y real -3.75 | -5 | jerk | -10 | yaw_rate | ; When decision=2, r decision =5-5∣y real -3.75∣-∣jerk∣-∣yaw_rate∣.

[0014] Preferably, TTM is the ratio of the remaining length of the acceleration lane to the current speed of the vehicle.

[0015] Preferably, the import mode signal decision is determined according to the following logic: If the binary classification signal indicates that the vehicle cannot merge, then set decision = 0, control the vehicle to adjust its speed and search for a mergeable area; If the binary classification signal is merging and TTM is greater than 2.5s, then set decision = 1 and control the vehicle to execute a smooth merging trajectory. If the binary classification signal is merging and TTM is less than 2.5s, then set decision = 2 and control the vehicle to execute the emergency merging trajectory.

[0016] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for optimizing the success rate of vehicle merging at high-speed ramps as described.

[0017] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for optimizing the success rate of vehicle merging at high-speed ramps as described.

[0018] Fourthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method for optimizing the success rate of vehicle merging at high-speed ramps as described.

[0019] (3) Beneficial effects The vehicle high-speed ramp merging cooperative control method provided by this invention constructs a DRL decision planning and introduces merging mode switching. Based on LightGBM and the remaining merging time (TTM) mode decision module, it provides scene perception signals for DRL, enabling it to generate safer and more reasonable trajectories according to the reward functions of different modes, thereby significantly reducing the risk of collision and lane exit and improving the merging success rate. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method for optimizing the success rate of vehicle merging into high-speed ramps in this invention; Figure 2 This is a flowchart of Example 4; Figure 3 This is the merging situation of the three models in Example 5, where (a) is the lateral displacement and (b) is the longitudinal vehicle speed. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that when a component is said to be "installed on" another component, it can be directly on the other component or it may be in a component that is centered on it. When a component is said to be "set on" another component, it can be directly set on the other component or it may also be in a component that is centered on it. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or it may also be in a component that is centered on it.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] Example 1: Training Method for Pattern Decision Model Based on Imported Real-World Driving Data This example describes how the LightGBM classifier is trained and built using actual input data. This is an offline model preparation process that provides the intelligent core for online decision-making.

[0025] S101. Driver Data Collection A highway ramp merging scenario was constructed in the Town06 map of the CARLA simulator. The road width was set to 3.75 meters, and the acceleration lane length was 180 meters. To obtain training data that could cover various complex situations, a random traffic flow scenario was designed, with specific settings shown in Table 1. Table 1 Random Traffic Flow Settings

[0026] CAV stands for Intelligent Vehicle, which travels at a constant speed, and HV stands for Human-driven Vehicle, which uses an Intelligent Driver Model (IDM). The initial speed of vehicles on the main road is randomly selected between 20 and 30 m / s for each data collection.

[0027] Data was collected using a Logitech G29 driving simulator. The simulated vehicle was controlled via an external Logitech G29 steering wheel and pedals, enabling steering, acceleration, and deceleration control. The CARLA simulator collected data every 0.1 seconds, including the vehicle's speed (v), lateral and longitudinal coordinates (X, Y), and front wheel steering angle (δ). f The distance Δx between the vehicle and the vehicles in front and behind it in the left main lane. f ,Δx r Speed ​​difference Δv f ,Δv r A total of 12,000 sets of data were collected from various traffic flow scenarios.

[0028] S102. Dataset Construction and Model Training Based on data collected during the vehicle's merging process, such as front wheel steering angle and lateral displacement, the merging behavior is determined, dividing the dataset into "merging possible" and "non-merging possible" categories—a multi-input, single-output binary classification problem. The model's merging decision input is... The output is The value “1” indicates that the current location can be merged into the network, and “0” indicates that the current road segment cannot be merged into the network.

[0029] If we consider "1" samples as eligible samples and "0" samples as non-eligible samples, then n t_t This indicates the number of samples that the model correctly identified as eligible for import, n. f_f This represents the number of samples that the model correctly identified as unacceptable for import, n. f_t The number of samples that the model incorrectly identifies as eligible samples, n. t_f This represents the number of samples that the model identifies as non-integrable. The model input is I, and the model output is O, with a training set to test set ratio of 3:1. The performance of each model can be evaluated using accuracy, precision, recall, and F1-core. Accuracy = Precision is the proportion of correct predictions out of the total number of predictions, measuring the accuracy of the model's predictions; Precision = Recall is the proportion of samples predicted as eligible for inclusion that are actually eligible for inclusion; it measures the reliability of the model's predictions of eligible samples. This refers to the proportion of samples that are actually eligible for inclusion but are predicted to be eligible for inclusion. It is used to measure the model's ability to find eligible samples. The precision and recall of the models are used to comprehensively evaluate their performance. The performance of each model is shown in Table 2. Table 2 Performance Comparison of Different Classification Models

[0030] As shown in Table 2, the LightGBM model performed best in all four metrics: accuracy, precision, recall, and F1-score, and was therefore selected as the final model for incorporating conditional judgments.

[0031] S103. Decision logic considering remaining inflow time (TTM) The trained LightGBM model was deployed to the online decision-making system. During online decision-making, the system not only relies on the model's classification results but also introduces the concept of Time-To-Merge (TTM) to enhance the timeliness of the decision. The formula for calculating TTM is: TTM = Remaining length of the acceleration lane / Current vehicle speed. A TTM of 2.5 seconds was chosen as the merging mode switching point.

[0032] The final decision-making logic for the online import mode is as follows: When the merge decision model outputs a signal indicating that merge is not possible, the merge mode decision=0. At this time, the main goal of the vehicle is to adjust its speed and quickly find and drive towards the area that the model determines is a mergeable area on the acceleration lane.

[0033] When the input model output signal is available for input and TTM>2.5s, the input mode decision=1, the input is feasible and the input time is sufficient. The trajectory planning pursues smoothness and accuracy in order to achieve a natural and smooth input.

[0034] When the input model output signal is input-available and TTM ≤ 2.5s, the input mode decision = 2, indicating that input is feasible but time is tight.

[0035] Example 2: A collaborative control method for merging vehicles on high-speed ramps, including optimization methods for the success rate of merging vehicles on high-speed ramps, S201-S202.

[0036] This embodiment describes in detail a cooperative control method for merging vehicles onto high-speed ramps, mainly including the following steps: S201. Import Mode Decision Steps This step corresponds to the online decision logic trained and deployed in Example 1.

[0037] The LightGBM machine learning model is used to determine the current environmental state (including the state of the vehicle and the vehicle itself) and output a signal indicating whether the vehicle can be merged or not.

[0038] Calculate the remaining merging time TTM, where TTM = remaining length of the acceleration lane / current vehicle speed.

[0039] Based on the above judgment results and TTM values, the final inflow mode decision signal (decision) is determined: If it cannot be imported, then decision = 0; If import is possible and TTM>2.5s, then decision=1 (normal import). If it can be imported and TTM≤2.5s, then decision=2 (urgent import).

[0040] S202. DRL Trajectory Planning Steps Based on the vehicle's state, traffic environment state, and merging pattern signal decision, trajectory planning is performed using a deep reinforcement learning (DRL) model.

[0041] Using the SAC algorithm, reinforcement learning learns the optimal policy through interaction with the environment. The state space is the foundation for ensuring the feasibility of the policy, and it is defined as follows:

[0042] Where v is the vehicle speed, X and Y are the lateral and longitudinal displacements of the vehicle with the position of entering the acceleration lane as the origin, and ψ is the vehicle's heading angle. The distance between your car and the cars in front and behind. To the speed difference between vehicles in front and behind on the main road These represent the lateral displacement tracking error and the longitudinal vehicle speed tracking error, while decision is the merge mode signal.

[0043] The motion space is: , used to output the increment of the target trajectory, where The target lateral displacement increment for the next moment is set to a range of 1. , The target longitudinal speed increment for the next moment is set to a range of 100%. .

[0044] The reward function is: The specific calculation method for each reward item is related to the input mode, ensuring that different trajectory planning strategies are used in different modes.

[0045] The reward function includes the following sub-items: 1. Successful remittance reward r succ The reward indicates whether the vehicle has completed the merging mission at the end of the round.

[0046]

[0047] 2. Position reward r loc To characterize the positional distribution reward between the vehicle and the vehicles in front and behind during the merging process,

[0048] 3. Safety Rewards safe To characterize the safety bonus for the autonomous vehicle during the merging process,

[0049] 4. Tracking error reward r track It is a key reward used to evaluate whether the DRL decision planning trajectory meets the vehicle dynamics requirements.

[0050]

[0051] 5. Behavioral reward r decision The reward for vehicle merging behavior varies depending on the model and is dynamically adjusted based on the merging mode. When decision=0, r decision =5-5∣y real |; When decision=1, r decision =5-∣y real -3.75 | -5 | jerk | -10 | yaw_rate | ; When decision=2, r decision =5-5∣y real -3.75∣-∣jerk∣-∣yaw_rate∣.

[0052] Example 3 like Figure 1 As shown, this embodiment describes a vehicle trajectory tracking method based on a weighted adaptive model. This method is applicable to high-speed scenarios, maintaining trajectory tracking accuracy and vehicle stability when the vehicle steering actuator experiences front wheel steering angle loss.

[0053] First, an MPC framework based on a 3-DOF vehicle dynamics model is established, considering the vehicle's lateral, longitudinal, and yaw motions, and selecting state variables. The control variable is The nonlinear dynamics model of the vehicle is Its discretized state-space equation can be expressed as:

[0054] in , , QUOTE T is the sampling time.

[0055] Within the MPC framework, at each sampling time, the controller optimizes its objective function based on the current state and the reference trajectory for the next N steps:

[0056] in, These are the predicted output and the reference output, respectively. To control the increment, Q and R are weight matrices. The Q matrix is ​​used to change the controller's tracking performance on the reference trajectory, N p For predicting the time domain, N c To control the time domain, the R matrix adjusts the magnitude of the system's control input, and ε is the relaxation factor, which ensures that the quadratic programming problem has an optimal solution. Combining the above, this problem can be transformed into a quadratic programming problem: .

[0057] The specific implementation steps include: S301. Obtain real-time vehicle data Obtain real-time longitudinal vehicle speed v x and real-time lateral tracking error e y Among them, e y This refers to the lateral deviation between the vehicle's current lateral position and the reference trajectory.

[0058] S302. Perform adaptive weight adjustment 1. Weighting based on vehicle speed First, the baseline weight matrix is ​​adjusted based on the real-time vehicle speed. The baseline weight matrix includes the tracking error baseline weight matrix Q0 and the control increment baseline weight matrix R0. Adaptive speed adjustment is performed using the following formula:

[0059] Where λ is the error weight attenuation coefficient and κ is the control weight enhancement coefficient. This adjustment appropriately reduces the tracking error weight and increases the control increment weight during high-speed driving, thereby enhancing system stability.

[0060] 2. Dynamic adjustment based on tracking error Next, based on the real-time lateral tracking error e y and its rate of change Calculate the dynamic adjustment factor :

[0061] Where, β min β max These are the maximum and minimum values ​​of the preset adjustment factor, respectively; e danger This is the dangerous limit; e safe The safety limit is represented by ε, where ε is the lateral error sensitivity coefficient. ; e danger Take 0.4m; e safe The value is set to 0.05m; this dynamic adjustment factor can reflect the urgency of the current tracking status.

[0062] 3. Calculate the final adaptive weights The vehicle speed adaptive weights are finally adjusted using a dynamic adjustment factor:

[0063] This design ensures that during vehicle operation, the base weights, Q(V), are first adjusted based on the vehicle speed. x ), R(V xAs vehicle speed increases, the tracking error weight Q decreases, while the control increment weight R increases to prevent instability caused by overly aggressive operation. When the controller begins to merge the tracking trajectory, a steering loss occurs, increasing the vehicle's lateral tracking error. At this point, the tracking error weight Q increases, while the control increment weight R decreases, allowing for a larger front wheel steering angle control input to compensate for the tracking error caused by the steering loss.

[0064] S303. Generate control commands The adaptive weight matrix Q adapt and R adapt The optimal control command, including the front wheel steering angle command δ, is obtained by solving the objective function applied to the MPC controller. f and longitudinal acceleration command a x And send it to the vehicle actuator.

[0065] Simultaneously, the lateral displacement tracking error Δy and the longitudinal vehicle speed tracking error Δv generated in the trajectory tracking control are also included. x The data is fed back to the DRL model in real time, forming a closed-loop control structure.

[0066] This feedback enables the DRL model to perceive the tracking performance of the lower-level control, especially when the tracking error increases due to actuator angle loss. This prompts the DRL to generate a more easily trackable trajectory in subsequent planning, thereby forming a closed-loop control that coordinates the upper and lower levels.

[0067] Example 4: Vehicle Merging into High-Speed ​​Ramp Cooperative Control System like Figure 2 As shown in the figure, this embodiment describes a vehicle high-speed ramp merging cooperative control system.

[0068] 1. System Hardware Components Perception module: Includes various vehicle sensors (such as cameras, radar, LiDAR, GPS / IMU) to acquire vehicle status (such as vehicle speed and pose) and traffic environment status (such as the position and speed of surrounding vehicles).

[0069] Computing Unit: A high-performance onboard computer or domain controller used to run the core algorithms in the following software architecture.

[0070] Actuators: including steering, drive and braking systems, are used to execute control commands issued by the computing unit.

[0071] 2. System Software Architecture The computing unit is configured to execute computer programs, thereby implementing the following functional modules: Inflow pattern decision module: configured to execute based on data from the perception module and a pre-trained LightGBM model, outputting an inflow pattern signal decision.

[0072] DRL trajectory planning module: Receives sensing data and pattern decision signals, runs the DRL algorithm, and generates the target trajectory.

[0073] Weighted Adaptive MPC Trajectory Tracking Module: Receives the target trajectory, runs the weighted adaptive MPC algorithm, and generates vehicle control commands (front wheel steering angle and longitudinal acceleration).

[0074] Closed-loop feedback path: Within the software architecture, the MPC trajectory tracking module generates tracking errors Δy, Δv. x The data is transmitted to the DRL trajectory planning module in real time.

[0075] Example 5: Simulation and Analysis Joint simulation analysis was performed using the Matlab / Simulink, Carsim, and Prescan co-simulation platforms.

[0076] Table 3 Vehicle Parameters

[0077] 1. Import Decision Module To evaluate the merging performance of the two models after training, 1000 simulation tests were conducted in randomly generated traffic flow scenarios. The results are summarized in Table 4 below.

[0078] Table 4 Import Results

[0079] As can be seen from Table 4, the DRL-MPC import model with integrated import mode decision module significantly increases the import success rate.

[0080] 2. DRL-MPC Merging Model Considering Cornering Loss The above simulation studies compared and verified the merging behavior of the proposed merging decision module and the tracking performance of the adaptive MPC. Now, the DRL-MPC module considering the cornering loss problem is validated. The merging trajectory determined by DRL under normal conditions, the merging trajectory of DRL without considering the cornering loss problem under 50% steering loss, and the merging trajectory of DRL considering the cornering loss problem under 50% cornering loss are compared under different MPC conditions. These are represented by Model 1, Model 2, and Model 3, respectively. The merging trajectory and vehicle speed change under the three models are shown below. Figure 3 As shown.

[0081] like Figure 3As shown in (a), when the DRL does not consider the corner loss problem during training, if this problem actually occurs, the planned merging trajectory (Model 2) is similar to the normal situation (Model 1). This means that when corner loss occurs, MPC needs to output a large corner to compensate for the loss, which may lead to significant tracking errors and yaw rate fluctuations. However, when this problem is taken into account during training, a smoother curve can be planned within the merging range (Model 3), meaning that MPC only needs a small corner compensation to achieve good tracking performance. Conservative weights under the three models ( Adaptive weights and aggressive weights The trajectory tracking performance under these conditions is shown in Table 5.

[0082] Table 5 Tracking Performance Indicators

[0083] The results show that the proposed DRL-MPC merging model considering corner loss can effectively address the corner loss problem. The MPC controller improves trajectory tracking accuracy through real-time adaptive compensation. As shown in Table 5, although the maximum and average errors of Model 3 under adaptive weights increase by 50.8% and 58.8% respectively compared to Model 1 due to corner loss, the reasonable trajectory planning of DRL reduces the maximum yaw rate and maximum yaw acceleration by 28.6% and 43% respectively. Compared to Model 2, which does not consider corner loss but suffers from steering loss, Model 1 under adaptive weights reduces the maximum and average errors by 45.7% and 44.9% respectively. Simultaneously, the maximum yaw rate and maximum yaw acceleration decrease by 49.35% and 54.4% respectively. In summary, this model, through a closed-loop feedback mechanism and adaptive weight strategy, can reasonably plan merging trajectories with smaller curvatures in the event of corner loss, and coordinate and optimize tracking accuracy and stability, ultimately achieving safe and stable merging.

[0084] Based on the above embodiments, the innovations are as follows: (1) A hybrid merging architecture that integrates data-driven and model-driven approaches is proposed. The upper layer uses deep reinforcement learning (DRL) for trajectory planning, and the lower layer uses model predictive control (MPC) for trajectory tracking. This framework comprehensively considers the expected functional safety and functional safety issues caused by DRL decision errors and steering actuator angle loss, thereby improving the safety and reliability of vehicle merging at the system level.

[0085] (2) An inflow condition judgment and mode switching mechanism based on LightGBM was designed. By evaluating the traffic environment and remaining inflow time in real time, the reward function and decision mode of DRL are dynamically adjusted. The addition of this module improves the efficiency of model training. At the same time, the model inflow success rate is greatly improved before and after the addition.

[0086] (3) An adaptive MPC weight adjustment strategy based on vehicle speed and tracking error is proposed. When the actuator suffers cornering loss, the control target weight can be dynamically optimized, which improves the tracking performance and stability. At the same time, the tracking effect of the lower layer is fed back to the DRL decision layer to form a closed-loop interaction, which enhances the coordination and safety of the system under abnormal conditions.

[0087] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0088] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for optimizing the success rate of vehicle merging at high-speed ramps, characterized in that, include: I. Decision-making steps for import mode The pre-trained LightGBM machine learning model determines whether the vehicle can merge or not based on its own vehicle status and traffic environment status, and outputs a binary classification signal. It calculates the remaining merging time TTM, and selects TTM=2.5s as the merging mode switching point based on the binary classification signal and TTM value, and outputs the merging mode signal decision. II. Decision-making and planning steps Trajectory planning is performed by dynamically adjusting the reward function of the DRL model based on the incoming pattern signal.

2. The method for optimizing the success rate of vehicle merging at high-speed ramps according to claim 1, characterized in that, The reward function r of the DRL model is: ， in, Rewards for successful remittances For location rewards, For security rewards, For MPC tracking error reward, Rewards are given for import behaviors under different decisions.

3. The method for optimizing the success rate of vehicle merging at high-speed ramps according to claim 2, characterized in that, The reward for successful import Location rewards Security rewards MPC tracking error reward They are defined as follows: ， ， ， 。 4. The method for optimizing the success rate of vehicle merging at high-speed ramps according to claim 2, characterized in that, The behavior reward r decision Dynamically adjust based on the import mode signal decision: When decision=0, r decision =5-5∣y real |; When decision=1, r decision =5-∣y real -3.75 | -5 | jerk | -10 | yaw_rate | ; When decision=2, r decision =5-5∣y real -3.75∣-∣jerk∣-∣yaw_rate∣.

5. The method for optimizing the success rate of vehicle merging at high-speed ramps according to claim 1, characterized in that, The TTM is the ratio of the remaining length of the acceleration lane to the current speed of the vehicle.

6. The method for optimizing the success rate of vehicle merging at high-speed ramps according to claim 1, characterized in that, The import mode signal decision is determined based on the following logic: If the binary classification signal indicates that the vehicle cannot merge, then set decision = 0, control the vehicle to adjust its speed and search for a mergeable area; If the binary classification signal is merging and TTM is greater than 2.5s, then set decision = 1 and control the vehicle to execute a smooth merging trajectory. If the binary classification signal is merging and TTM is less than 2.5s, then set decision = 2 and control the vehicle to execute the emergency merging trajectory.

7. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method as described in any one of claims 1 to 6.

Images