Ramp merge decision method based on autonomous vehicle

Abstract

The application discloses a ramp merging decision method for an automatic driving vehicle, and comprises the following steps: acquiring state information of the vehicle and vehicles in a preset detection range of the vehicle; constructing an LSTM-Transformer double-branch trajectory prediction model, which is used for predicting a multi-modal prediction trajectory of the surrounding vehicles in a future prediction time domain, and generating three candidate merging strategies, i.e., keeping the current lane, changing the lane to the left and changing the lane to the right, and corresponding planning trajectories for the vehicle; and screening the three candidate merging strategies, i.e., keeping the current lane, changing the lane to the left and changing the lane to the right, respectively, to determine the optimal ramp merging strategy of the vehicle. The application effectively reduces the dependence of the control system on a precise theoretical model, and greatly improves the reliability and robustness of trajectory tracking in a complex environment.

Description

Technical Field

[0001] This invention belongs to the field of driving and relates to a method for decision-making on ramp merging of autonomous vehicles. Background Technology

[0002] The merging area of ​​highway ramps is characterized by traffic flow intersections, forced lane changes, and multi-vehicle competition. If vehicles cannot accurately judge the movement trends of surrounding vehicles during the merging process, it can easily cause congestion, increased collision risk, or unnecessary frequent lane changes.

[0003] Existing rule-based ramp merging or lane-changing decision-making methods typically rely on fixed gaps, fixed thresholds, or preset trajectories, which have limited generalization capabilities in dynamic traffic environments. While existing learning-based methods can extract features from driving data, they still struggle to balance safety, comfort, and traffic efficiency without multi-vehicle interaction modeling and conflict risk constraints.

[0004] In summary, existing ramp merging or lane-changing decision-making methods are not suitable for complex merging behaviors in multi-vehicle interaction scenarios and need to be improved. Summary of the Invention

[0005] To address the aforementioned problems, the technical solution adopted by this invention is: a method for making decisions on ramp merging of autonomous vehicles, comprising the following steps: Acquire the status information of this vehicle and other vehicles within its preset detection range; Based on the state information of the vehicle and the vehicles within the vehicle's preset detection range, an LSTM-Transformer dual-branch trajectory prediction model is constructed to predict the multimodal predicted trajectories of surrounding vehicles in the future prediction time domain, and to generate three candidate merging strategies and corresponding planned trajectories for the vehicle: maintaining the current lane, changing lanes to the left, and changing lanes to the right. The optimal ramp merging strategy for this vehicle is determined by filtering three candidate merging strategies: staying in the current lane, changing lanes to the left, and changing lanes to the right.

[0006] Furthermore, the process of filtering the three candidate merging strategies—holding the current lane, changing lanes to the left, and changing lanes to the right—to determine the optimal ramp merging strategy for this vehicle is as follows: For each of the three candidate merging strategies, perform spatiotemporal collision detection between its planned trajectory and the multimodal predicted trajectories of all surrounding vehicles. If there is a collision risk, exclude the strategy. For candidate merging strategies that pass collision detection, verify whether they meet the constraints of speed, acceleration and rate of change of acceleration, and lane change conflict risk. Strategies that do not meet any of the constraints are eliminated to obtain valid candidate strategies. For the remaining effective candidate strategies, their comprehensive cost is calculated based on the multi-objective cost function; The merging benefit of each candidate lane is calculated based on the set of lane numbers, the current lane number, the driving benefit weight, the lane change cost weight, the average vehicle speed, the maximum speed limit of the corresponding lane, and the time required for the vehicle to change lanes. The optimal ramp merging decision is selected based on the strategy that minimizes overall cost and maximizes merging benefits; if no effective strategy exists, the decision to maintain the current lane is output.

[0007] Furthermore: the LSTM-Transformer dual-branch trajectory prediction model includes: LSTM longitudinal branch: used to encode temporal evolution features of velocity and acceleration; Transformer lateral branch: used to encode spatial interaction features between vehicles; the Transformer lateral branch uses a Transformer encoder; The cross-attention fusion module is used to predict the multimodal prediction trajectories of surrounding vehicles in the future prediction time domain based on the temporal evolution features of velocity and acceleration transmitted by the longitudinal branch of the LSTM and the spatial interaction features transmitted by the lateral branch of the Transformer, and to generate three candidate merging strategies and corresponding planned trajectories for the vehicle: maintaining the current lane, changing lanes to the left, and changing lanes to the right.

[0008] Furthermore, the process of predicting the multimodal predicted trajectories of surrounding vehicles in the future prediction time domain based on the temporal evolution features of velocity and acceleration transmitted by the LSTM longitudinal branch and the spatial interaction features transmitted by the Transformer lateral branch, and generating three candidate merging strategies and corresponding planned trajectories for the vehicle—maintaining the current lane, changing lanes to the left, and changing lanes to the right—is as follows: For the scenario of merging onto a ramp, the vehicle input trajectory is encoded with relative position, as shown in formulas (1) and (2): (1); (2); Where PE is the relative two-dimensional tensor matrix of the vehicle input trajectory, pos is the position index of the data sequence, is the number of features, and i is the feature position; The relative position code PE is fused with the sample tensor X to obtain the input data of the Transformer encoder, as shown in formula (3): (3); The multi-head self-attention mechanism of the Transformer encoder is used to mine the interaction relationship between multiple vehicles in parallel. Its attention calculation and softmax normalization are shown in Equation (4) and Equation (5), respectively: (4); In the formula, the Q vector, K vector, and V vector are specific vectors in the attention mechanism; d k is the dimension of the K vector; softmax is the normalization function of the model; (5); In the formula, e is a natural number; z i It is the i-th element in the input vector; n is the number of elements in the vector. In the Transformer encoder, a feedforward neural network and a ReLU activation function are set up, and their expressions are shown in Equation (6) and Equation (7) respectively: (6); In the formula, x is the network input; W i These are the weights of the corresponding hidden layers; b i It is the corresponding bias vector; The activation function used is the ReLU function, and its calculation formula is as follows: (7).

[0009] Furthermore, the Transformer encoder includes: Multi-head self-attention layer: The multi-head self-attention mechanism can mine vehicle interaction features in different dimensions in parallel, effectively model the dynamic spatiotemporal dependencies of multiple vehicles, and provide key environmental representations for subsequent decisions, extracting the spatiotemporal interaction relationships between different vehicles; Residual Connector Layer: Used to preserve original vehicle feature information, alleviate the gradient data vanishing problem, and ensure cross-layer information data flow; Normalization layer: Used to generate a normalization function from a specific vector in the attention mechanism to standardize the features and reduce the impact of numerical fluctuations; Feedforward neural network layer: used to complete the mapping between corresponding data and feature vectors; The multi-head self-attention layer, residual connection layer, normalization layer, and feedforward neural network layer all use residual connections and layer normalization to stabilize training.

[0010] Furthermore, spatiotemporal collision detection is performed as shown in Equation (8): (8).

[0011] In the formula, d is the distance between the target vehicle and surrounding vehicles; r e and r o These are the maximum outer diameters of the target vehicle and the surrounding vehicles, respectively.

[0012] Furthermore, the speed constraint is as shown in formula (9): (9); In the formula, V x (t) is the longitudinal velocity of the vehicle, V y (t) is the lateral velocity of the vehicle; The acceleration and acceleration rate of change constraints are as shown in formula (10): (10); In the formula a x (t) and a y (t) These are the vehicle's longitudinal acceleration and lateral acceleration, respectively. a xmax and a ymax These are the corresponding maximum accelerations. b x (t) and b y (t) These are the corresponding rates of change of acceleration. b xmax and b ymax These represent the corresponding maximum rate of change of acceleration; μ is the road adhesion coefficient. The lane change conflict risk constraint is shown in formula (11): (11); In the formula, P 0 is the threshold for lane change conflict risk. V lccr This is the lane change conflict risk value. Its calculation is based on historical collision data, relative vehicle speed, rate of change of acceleration, and distance between vehicles, and is output through a pre-trained risk assessment network. Specifically, V lccr = f risk ( v rel , a rel , d , t ttc ), in v rel It is relative velocity. a rel It is relative acceleration. d It is the distance between vehicles. t ttc It refers to the collision time.

[0013] Furthermore, the multi-objective cost function includes a longitudinal comfort term, a lateral comfort term, and a driving efficiency term, and the formula (12) of the multi-objective cost function is as follows: (12).

[0014] In the formula, t fmax Indicates the maximum lane change time; j x (t) It is the rate of change of longitudinal acceleration. j y (t) It is the rate of change of lateral acceleration. j xmax It is the maximum permissible longitudinal jerk value. j ymax This is the maximum permissible lateral acceleration value; ω 1 is the weighting coefficient corresponding to longitudinal comfort; ω 2 is the weighting coefficient corresponding to lateral comfort; ω 3 is a weighting coefficient representing driving efficiency.

[0015] Furthermore, the calculation of the merging benefit for each candidate lane is shown in formula (13): (13).

[0016] In the formula, L It is a set of lane numbers; j It is the current lane number; ω v It is the driving efficiency weight of the current lane; ω c It is the weight of lane change cost; This is the time required for a vehicle to change lanes; It is the average speed of the vehicle; This is the maximum speed limit for the corresponding lane.

[0017] The present invention provides a method for decision-making on ramp merging of autonomous vehicles. Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention enhances the expression of multi-vehicle interaction features through a Transformer encoder, improves trajectory prediction capability through an LSTM-Transformer dual-branch model, and balances safety, comfort and efficiency through risk constraints and multi-objective cost functions, thereby improving the success rate of ramp merging and reducing unnecessary lane changes. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of a method for making decisions on ramp merging of autonomous vehicles; Figure 2 This is an LSTM-Transformer dual-branch trajectory prediction model; Figure 3 For the vehicle lane change decision process; Figure 4 The diagrams show the predicted vehicle trajectories 3 seconds ago, 2 seconds ago, and 1 second ago, where (a) is 3 seconds ago, (b) is 2 seconds ago, and (c) is 1 second ago. Figure 5 This is a schematic diagram illustrating the impact of the prediction time domain on the prediction results.

[0020] Figure 6 This diagram illustrates the impact of data sparsity on prediction results.

[0021] Figure 7 This is a schematic diagram of the changes during the lane merging process.

[0022] Figure 8 This is a schematic diagram comparing the ACC box plots between models.

[0023] Figure 9 This diagram illustrates the success rate of vehicle merging under low traffic density.

[0024] Figure 10 This diagram illustrates the average number of lane changes and speed variations for vehicles under low traffic density. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. 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.

[0027] Figure 1 This is a schematic diagram of a method for making decisions on ramp merging of autonomous vehicles.

[0028] A method for making decisions on merging into an autonomous vehicle ramp includes the following steps: S1: Acquire the status information of the vehicle and vehicles within the preset detection range of the vehicle within a time step t; the status information includes position, speed, acceleration, heading angle and current lane number; S2: Based on the state information of the vehicle and the vehicles within the preset detection range of the vehicle, an LSTM-Transformer dual-branch trajectory prediction model is constructed to predict the multimodal prediction trajectory of surrounding vehicles in the future prediction time domain, and to generate three candidate merging strategies and corresponding planned trajectories for the vehicle: maintaining the current lane, changing lanes to the left, and changing lanes to the right. S3: Select the optimal ramp merging strategy for this vehicle by filtering the three candidate merging strategies: holding the current lane, changing lanes to the left, and changing lanes to the right.

[0029] Furthermore, the process of filtering the three candidate merging strategies—holding the current lane, changing lanes to the left, and changing lanes to the right—to determine the optimal ramp merging strategy for this vehicle is as follows: S31: For each of the three candidate merging strategies, perform spatiotemporal collision detection between its planned trajectory and the multimodal predicted trajectories of all surrounding vehicles. If there is a collision risk, exclude the strategy. S32: For candidate inflow strategies that pass collision detection, verify whether they meet the speed constraints, acceleration and rate of change of acceleration and lane change conflict risk constraints, eliminate strategies that do not meet any of the constraints, and obtain valid candidate strategies. S33: For the remaining effective candidate strategies, calculate their comprehensive cost based on the multi-objective cost function; S34: Calculate the merging benefit of each candidate lane based on the lane number set, current lane number, driving benefit weight, lane change cost weight, average vehicle speed, maximum speed limit of the corresponding lane, and time required for lane change. S35: Select the strategy that minimizes the overall cost and maximizes the merging benefit as the optimal ramp merging decision; if no effective strategy exists, output the decision to maintain the current lane.

[0030] Figure 2 This is an LSTM-Transformer dual-branch trajectory prediction model; Figure 3 For the vehicle lane change decision process; Furthermore: the LSTM-Transformer dual-branch trajectory prediction model includes: LSTM longitudinal branch: used to encode temporal evolution features of velocity and acceleration; Transformer Lateral Branch: Used to encode spatial interaction features between vehicles; the Transformer Lateral Branch employs a Transformer encoder; vehicle interaction features containing relative position encoding are constructed based on state information; the vehicle interaction features are input into the Transformer encoder, and the vehicle interaction features include: encoding the two-dimensional relative positions of the main vehicle and surrounding vehicles to obtain a relative position encoding matrix; the relative position encoding matrix is ​​fused with the vehicle motion state tensor to obtain the input tensor of the Transformer encoder.

[0031] The cross-attention fusion module is used to predict the multimodal prediction trajectories of surrounding vehicles in the future prediction time domain based on the temporal evolution features of velocity and acceleration transmitted by the longitudinal branch of the LSTM and the spatial interaction features transmitted by the lateral branch of the Transformer, and to generate three candidate merging strategies and corresponding planned trajectories for the vehicle: maintaining the current lane, changing lanes to the left, and changing lanes to the right.

[0032] Furthermore: the process of predicting the multimodal predicted trajectories of surrounding vehicles in the future prediction time domain based on the temporal evolution features of velocity and acceleration transmitted by the LSTM longitudinal branch and the spatial interaction features transmitted by the Transformer lateral branch, and generating three candidate merging strategies and corresponding planned trajectories for the vehicle—maintaining the current lane, changing lanes to the left, and changing lanes to the right—is as follows: For the scenario of merging onto a ramp, the vehicle input trajectory is encoded with relative position, as shown in formulas (1) and (2): (1); (2); Where PE is the relative two-dimensional tensor matrix of the vehicle input trajectory, pos is the position index of the data sequence, is the number of features, and i is the feature position; The relative position code PE is fused with the sample tensor X to obtain the input data of the Transformer encoder, as shown in formula (3): (3); The multi-head self-attention mechanism of the Transformer encoder is used to mine the interaction relationship between multiple vehicles in parallel. Its attention calculation and softmax normalization are shown in Equation (4) and Equation (5), respectively: (4); In the formula, Q vector Query, K vector Key, and V vector are specific vectors in the attention mechanism; d k is the dimension of the K vector; softmax is the normalization function of the model; (5); In the formula, e is a natural number; z i It is the i-th element in the input vector; n is the number of elements in the vector. In the Transformer encoder, a feedforward neural network and a ReLU activation function are set up, and their expressions are shown in Equation (6) and Equation (7) respectively: (6); In the formula, x is the network input; W i These are the weights of the corresponding hidden layers; b i It is the corresponding bias vector; The activation function used is the ReLU function, and its calculation formula is as follows: (7).

[0033] Furthermore, the Transformer encoder includes: Multi-head self-attention layer: The multi-head self-attention mechanism can mine vehicle interaction features in different dimensions in parallel, effectively model the dynamic spatiotemporal dependencies of multiple vehicles, and provide key environmental representations for subsequent decisions, extracting the spatiotemporal interaction relationships between different vehicles; Residual Connector Layer: Used to preserve original vehicle feature information, alleviate the gradient data vanishing problem, and ensure cross-layer information data flow; Normalization layer: Used to generate a normalization function from a specific vector in the attention mechanism to standardize the features and reduce the impact of numerical fluctuations; Feedforward neural network layer: used to complete the mapping between corresponding data and feature vectors; The multi-head self-attention layer, residual connection layer, normalization layer, and feedforward neural network layer all employ residual connections and layer normalization for stable training. The logical relationship between the four layers consists of stacked multi-head self-attention layers and feedforward neural network layers.

[0034] Furthermore, spatiotemporal collision detection is performed as shown in Equation (8): (8).

[0035] In the formula, d is the distance between the target vehicle and surrounding vehicles; r e and r o These are the maximum outer diameters of the target vehicle and the surrounding vehicles, respectively.

[0036] Furthermore, the speed constraint is as shown in formula (9): (9); In the formula, V x (t) and V y (t) represent the longitudinal and lateral velocities of the vehicle, respectively; The acceleration and acceleration rate of change constraints are as shown in formula (10): (10); In the formula a x (t) and a y (t) These are the vehicle's longitudinal acceleration and lateral acceleration, respectively. a xmax and a ymax These are the corresponding maximum accelerations. b x (t) and b y (t) These are the corresponding rates of change of acceleration. b xmax and b ymax These represent the corresponding maximum rate of change of acceleration; μ is the road adhesion coefficient. The lane change conflict risk constraint is shown in formula (11): (11); In the formula, P 0 is the threshold for lane change conflict risk. V lccr This is the lane change conflict risk value. Its calculation is based on historical collision data, relative vehicle speed, rate of change of acceleration, and distance between vehicles, and is output through a pre-trained risk assessment network. Specifically, V lccr = f risk ( v rel , a rel , d , t ttc ), in v rel It is relative velocity. a rel It is relative acceleration. d It is the distance between vehicles. t ttc It refers to the collision time.

[0037] Furthermore, the multi-objective cost function includes a longitudinal comfort term, a lateral comfort term, and a driving efficiency term. The weight of the longitudinal comfort term is 0.4, the weight of the lateral comfort term is 0.3, and the weight of the driving efficiency term is 0.3. The formula (12) of the multi-objective cost function is as follows: (12).

[0038] In the formula, t fmax Indicates the maximum lane change time; j x (t) It is the rate of change of longitudinal acceleration. j y (t) It is the rate of change of lateral acceleration. j xmax It is the maximum permissible longitudinal jerk value. j ymax This is the maximum permissible lateral acceleration value; ω 1 is the weighting coefficient corresponding to longitudinal comfort; ω 2 is the weighting coefficient corresponding to lateral comfort; ω 3 is a weighting coefficient representing driving efficiency.

[0039] Furthermore, the calculation of the merging benefit for each candidate lane is shown in formula (13): (13).

[0040] In the formula, L It is a set of lane numbers; j It is the current lane number; ω v It is the driving efficiency weight of the current lane; ω cIt is the weight of lane change cost; This is the time required for a vehicle to change lanes; It is the average speed of the vehicle; This is the maximum speed limit for the corresponding lane.

[0041] The optimal lane at any given time is determined as shown in formula (14): (14).

[0042] In the experimental embodiment, road trajectory data from the entrance ramps of the G6 Beijing-Tibet Expressway in China were used for verification. The data acquisition frequency was 10Hz. After smoothing, coordinate normalization, and outlier removal, the original trajectory data yielded 186,342 trajectory points, corresponding to 1,125 valid vehicle trajectories.

[0043] Evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), true positive rate (TPR), true negative rate (TNR), and accuracy (ACC), corresponding to formulas (15) to (18): (15); (16); (17); (18) The experimental parameters are set as shown in Table 1.

[0044] Table 1. Experimental Parameter Settings:

[0045] Trajectory prediction results are as follows Figure 4 As shown, Figure 4 The diagrams show the predicted vehicle trajectories 3 seconds ago, 2 seconds ago, and 1 second ago, where (a) is 3 seconds ago, (b) is 2 seconds ago, and (c) is 1 second ago. The effects of prediction time domain and data sparsity on prediction results are as follows: Figure 5 and Figure 6 As shown.

[0046] The changes in the merging process during the model learning and training process of this invention are as follows: Figure 7 As shown in Table 2, the performance of different decision-making models is compared.

[0047] Table 2. Comparison of the performance of different decision-making models:

[0048] Table 2 shows that the method of this invention performs better in TPR, TNR, and ACC, with TPR at 95.201%, TNR at 93.192%, and ACC at 96.310%. A comparison of ACC box plots between models is shown below. Figure 8 As shown.

[0049] To verify the rationality of the cost function weight settings, ablation experiments were conducted on different weight combinations, and the results are shown in Table 3.

[0050] Table 3. Comparison of ablation experimental performance using cost function weighting coefficients:

[0051] To further verify that the algorithm has the ability to make lane change decisions in a human-like manner, the model's decisions were compared with the actual lane change decisions of human drivers. The statistical results are shown in Table 4.

[0052] Table 4. Statistical results of various indicators for different vehicles:

[0053] The experimental results under different traffic densities are as follows: the vehicle merging success rate under low traffic density is as follows: Figure 9 As shown, the changes in average lane change frequency and speed of vehicles under low traffic density are as follows: Figure 10 As shown in the figure; the comparison of merging effects under three traffic densities is shown in Table 5.

[0054] Table 5. Comparison of merging effects under different traffic densities:

[0055] As shown in Table 5, the merging success rates are 100%, 98%, and 96% under low, medium, and high traffic densities, respectively, indicating that the algorithm can facilitate autonomous vehicles to complete ramp merging under different traffic densities.

[0056] To evaluate the feasibility of deploying the algorithm to a real system, the computational efficiency was analyzed, and the results are shown in Table 6.

[0057] Table 6. Computational Efficiency Analysis:

[0058] As shown in Table 6, the total time for a single decision by this algorithm is 98.2ms, which meets the 100ms cycle requirement typically required by decision modules in autonomous driving systems.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for decision-making regarding ramp merging of autonomous vehicles, characterized in that, Includes the following steps: Acquire the status information of this vehicle and other vehicles within its preset detection range; Based on the state information of the vehicle and the vehicles within the vehicle's preset detection range, an LSTM-Transformer dual-branch trajectory prediction model is constructed to predict the multimodal predicted trajectories of surrounding vehicles in the future prediction time domain, and to generate three candidate merging strategies and corresponding planned trajectories for the vehicle: maintaining the current lane, changing lanes to the left, and changing lanes to the right. The optimal ramp merging strategy for this vehicle is determined by filtering three candidate merging strategies: staying in the current lane, changing lanes to the left, and changing lanes to the right.

2. The method for decision-making on ramp merging of autonomous vehicles according to claim 1, characterized in that, The process of filtering the three candidate merging strategies—maintaining the current lane, changing lanes to the left, and changing lanes to the right—to determine the optimal ramp merging strategy for this vehicle is as follows: For each of the three candidate merging strategies, perform spatiotemporal collision detection between its planned trajectory and the multimodal predicted trajectories of all surrounding vehicles. If there is a collision risk, exclude the strategy. For candidate merging strategies that pass collision detection, verify whether they meet the constraints of speed, acceleration and rate of change of acceleration, and lane change conflict risk. Strategies that do not meet any of the constraints are eliminated to obtain valid candidate strategies. For the remaining effective candidate strategies, their comprehensive cost is calculated based on the multi-objective cost function; The merging benefit of each candidate lane is calculated based on the set of lane numbers, the current lane number, the driving benefit weight, the lane change cost weight, the average vehicle speed, the maximum speed limit of the corresponding lane, and the time required for the vehicle to change lanes. The optimal ramp merging decision is selected based on the strategy that minimizes overall cost and maximizes merging benefits; if no effective strategy exists, the decision to maintain the current lane is output.

3. The autonomous vehicle ramp merging decision method according to claim 1, characterized in that: The LSTM-Transformer dual-branch trajectory prediction model includes: LSTM longitudinal branch: used to encode temporal evolution features of velocity and acceleration; Transformer lateral branch: used to encode spatial interaction features between vehicles; the Transformer lateral branch uses a Transformer encoder; The cross-attention fusion module is used to predict the multimodal prediction trajectories of surrounding vehicles in the future prediction time domain based on the temporal evolution features of velocity and acceleration transmitted by the longitudinal branch of the LSTM and the spatial interaction features transmitted by the lateral branch of the Transformer, and to generate three candidate merging strategies and corresponding planned trajectories for the vehicle: maintaining the current lane, changing lanes to the left, and changing lanes to the right.

4. The autonomous vehicle ramp merging decision method according to claim 1, characterized in that: The process of predicting the multimodal predicted trajectories of surrounding vehicles in the future prediction time domain based on the temporal evolution features of velocity and acceleration transmitted by the LSTM longitudinal branch and the spatial interaction features transmitted by the Transformer lateral branch, and generating three candidate merging strategies and corresponding planned trajectories for the vehicle—maintaining the current lane, changing lanes to the left, and changing lanes to the right—is as follows: For the scenario of merging onto a ramp, the vehicle input trajectory is encoded with relative position, as shown in formulas (1) and (2): (1)； (2)； Where PE is the relative two-dimensional tensor matrix of the vehicle input trajectory, pos is the position index of the data sequence, is the number of features, and i is the feature position; The relative position code PE is fused with the sample tensor X to obtain the input data of the Transformer encoder, as shown in formula (3): (3)； The multi-head self-attention mechanism of the Transformer encoder is used to mine the interaction relationship between multiple vehicles in parallel. Its attention calculation and softmax normalization are shown in Equation (4) and Equation (5), respectively: (4)； In the formula, the Q vector, K vector, and V vector are specific vectors in the attention mechanism; d k is the dimension of the K vector; softmax is the normalization function of the model; (5)； In the formula, e is a natural number; z i It is the i-th element in the input vector; n is the number of elements in the vector. In the Transformer encoder, a feedforward neural network and a ReLU activation function are set up, and their expressions are shown in Equation (6) and Equation (7) respectively: (6)； In the formula, x is the network input; W i These are the weights of the corresponding hidden layers; b i It is the corresponding bias vector; The activation function used is the ReLU function, and its calculation formula is as follows: (7)。 5. The autonomous vehicle ramp merging decision method according to claim 3, characterized in that, Transformer encoders include: Multi-head self-attention layer: The multi-head self-attention mechanism can mine vehicle interaction features in different dimensions in parallel, effectively model the dynamic spatiotemporal dependencies of multiple vehicles, and provide key environmental representations for subsequent decisions, extracting the spatiotemporal interaction relationships between different vehicles; Residual Connector Layer: Used to preserve original vehicle feature information, alleviate the gradient data vanishing problem, and ensure cross-layer information data flow; Normalization layer: Used to generate a normalization function from a specific vector in the attention mechanism to standardize the features and reduce the impact of numerical fluctuations; Feedforward neural network layer: used to complete the mapping between corresponding data and feature vectors; The multi-head self-attention layer, residual connection layer, normalization layer, and feedforward neural network layer all use residual connections and layer normalization to stabilize training.

6. The autonomous vehicle ramp merging decision method according to claim 1, characterized in that, The spatiotemporal collision detection uses the formula shown in formula (8): (8) In the formula, d is the distance between the target vehicle and surrounding vehicles; r e and r o These are the maximum outer diameters of the target vehicle and the surrounding vehicles, respectively.

7. The autonomous vehicle ramp merging decision method according to claim 1, characterized in that, The velocity constraint is expressed by formula (9): (9)； In the formula, V x (t) is the longitudinal velocity of the vehicle, V y (t) is the lateral velocity of the vehicle; The acceleration and acceleration rate of change constraints are as shown in formula (10): (10)； In the formula a x (t) and a y (t) These are the vehicle's longitudinal acceleration and lateral acceleration, respectively. a xmax and a ymax These are the corresponding maximum accelerations. b x (t) and b y (t) These are the corresponding rates of change of acceleration. b xmax and b ymax These represent the corresponding maximum rate of change of acceleration; μ is the road adhesion coefficient. The lane change conflict risk constraint is shown in formula (11): (11)； In the formula, P 0 is the threshold for lane change conflict risk. V lccr This is the lane change conflict risk value. Its calculation is based on historical collision data, relative vehicle speed, rate of change of acceleration, and distance between vehicles, and is output through a pre-trained risk assessment network. Specifically, V lccr = f risk ( v rel , a rel , d , t ttc ), in v rel It is relative velocity. a rel It is relative acceleration. d It is the distance between vehicles. t ttc It refers to the collision time.

8. The autonomous vehicle ramp merging decision method according to claim 1, characterized in that, The multi-objective cost function includes a longitudinal comfort term, a lateral comfort term, and a driving efficiency term. The formula (12) for the multi-objective cost function is as follows: (12)。 In the formula, t fmax Indicates the maximum lane change time; j x (t) It is the rate of change of longitudinal acceleration. j y (t) It is the rate of change of lateral acceleration. j xmax It is the maximum permissible longitudinal jerk value. j ymax This is the maximum permissible lateral acceleration value; ω 1 is the weighting coefficient corresponding to longitudinal comfort; ω 2 is the weighting coefficient corresponding to lateral comfort; ω 3 is a weighting coefficient representing driving efficiency.

9. The autonomous vehicle ramp merging decision method according to claim 1, characterized in that, The merging benefit for each candidate lane is calculated as shown in formula (13): (13) In the formula, L It is a set of lane numbers; j It is the current lane number; ω v It is the driving efficiency weight of the current lane; ω c It is the weight of lane change cost; This is the time required for a vehicle to change lanes; It is the average speed of the vehicle; This is the maximum speed limit for the corresponding lane.

Images