Intelligent ramp merging human-like trajectory generation system and method

Abstract

The application provides a kind of intelligent car ramp merging human-like trajectory generation system and method, the establishment process of system includes obtaining and establishing ramp scene expert teaching data set, establishes and trains human-like trajectory generation model, establishes human-like trajectory coding space, establishes and trains traffic situation mapping network and the like steps;The application innovatively introduces generation network, establishes human-like trajectory generation model, utilizes the powerful feature extraction capability of generation network, realizes the human-like trajectory generation of high quality, good human-likeness;By introducing human-like trajectory feature vector, human-like trajectory coding space is established, and traffic situation mapping network is established by using long short-term memory network, the stable and diversified human-like trajectory generation is realized, the ramp merging human-like trajectory of stable, high quality, diversification, good human-likeness is generated, and the social acceptance and trust degree of intelligent car ramp merging trajectory planning are improved.

Description

Technical Field

[0001] This invention relates to a method for generating intelligent vehicle trajectories, and more particularly to a system and method for generating human-like trajectories for intelligent vehicle ramp merging based on a generative network. Background Technology

[0002] As the level of automotive intelligence continues to improve, the production and sales of intelligent vehicles equipped with advanced autonomous driving systems are constantly increasing. Advanced autonomous driving systems should possess the ability to perceive complex traffic situations and make correct behavioral decisions in interactions with surrounding vehicles. Merging trajectory planning in ramp scenarios is a challenging task for intelligent vehicles. As a scenario with frequent interactions with other traffic participants, intelligent vehicle ramp merging trajectory planning must ensure safety, tracking accuracy, and comfort while also considering the acceptability of human drivers. Therefore, intelligent vehicles should possess lane-changing behavior strategies and trajectories similar to those of humans.

[0003] Existing methods for generating human-like trajectories mainly include: (1) methods based on imitation learning, (2) methods based on model predictive control, and (3) methods based on game theory. However, these existing methods still have the following problems and shortcomings:

[0004] Method (1) generates a human-like trajectory by imitating the trajectory of a human in the same environment as the current state of the vehicle's environment. However, the strategy learned by this method cannot be fully optimized during the interaction with the environment, causing the obtained state to gradually deviate from the data distribution of the expert strategy, resulting in compound errors and causing the generated human-like trajectory to gradually deviate from the real human trajectory.

[0005] The trajectory generated by method (2) heavily relies on the optimization function of model predictive control. Generating humanoid trajectories for drivers with different driving styles requires different optimization function parameters. However, the adjustment process of the optimization function parameters set by this method is complex. It is not practical to specify a different set of optimization function parameters for the generation of each humanoid trajectory. Therefore, this method is difficult to generate diverse trajectories. In addition, the computational load of model predictive control is very high, making this method unsuitable for real-time applications.

[0006] Method (3) solves the Nash equilibrium based on the established game payoff matrix, and then determines the optimal action of the current vehicle. In trajectory generation applications, the action is usually set as acceleration and wheel angle. However, since the action of the game payoff matrix is ​​discrete, when applied to continuous action tasks such as motion planning that require high action fine granularity, there are a series of problems such as low simulation accuracy and poor trajectory smoothness, which makes the quality of the humanoid trajectory generated by this method poor.

[0007] In conclusion, how to generate stable, high-quality, diverse, and human-like ramp merging trajectories, and improve the social acceptance and trust in intelligent vehicle ramp merging trajectory planning, has become one of the key technical issues in the field of intelligent vehicles. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides an intelligent vehicle ramp merging human-like trajectory generation system based on a generative network, which is established through the following steps:

[0009] Step 1: Obtain and build expert teaching dataset for ramp scenarios

[0010] (1.1) Acquisition and processing of expert-integrated teaching data:

[0011] Using autonomous driving simulation software, the absolute speed, absolute longitudinal position, absolute lateral position, absolute acceleration, and heading angle data of merging vehicles (EVs) on the ramp merging lanes during the expert merging process, as well as the absolute speed, absolute longitudinal position, absolute lateral position, absolute acceleration, and heading angle data of following vehicles (FVs) on the main road, and the relative speed, relative longitudinal position, and relative lateral position data between EVs and FVs, were recorded. A total of N data points were sampled; the data sampling frequency was 0.1s, and the real time of all sampling points was recorded simultaneously during sampling.

[0012] The expert-integrated trajectory data uses the starting point of the EV integration state as the origin of the coordinate system, the actual time of the EV integration state as the zero moment of the integration time, the main road vehicle travel direction as the x-axis, and the y-axis as the direction 90 degrees counterclockwise from the x-axis as the y-axis to record the coordinates of the integration trajectory in the local rectangular coordinate system and the integration time of each coordinate point.

[0013] After normalizing the trajectories in the expert-input trajectory data, the resulting expert-input trajectory data format is as follows:

[0014]

[0015]

[0016] Where ζ1 represents the set of normalized expert-integrated trajectory data, and each trajectory includes K0 sampling points. Let represent the i-th trajectory in the set, and let T be the merging cutoff time of this trajectory calculated from the merging time of zero. i , Represents the time of the t-th sampling point. This represents the vehicle's ordinate at that moment. The x-coordinate of the vehicle at that moment;

[0017] (1.2) Traffic situation representation data acquisition and processing:

[0018] Based on the absolute velocity, absolute longitudinal position, absolute lateral position, absolute acceleration, and heading angle data of the EV and the FV obtained in step one (1.1), and the relative velocity, relative longitudinal position, and relative lateral position data between the EV and FV, the actual time of the starting point of the EV merging is taken as the ending time t. Several sampling points are traversed forward to extract the data within the time window H. After normalizing the data, the traffic situation characterization data R is formed. Traffic :

[0019] R Traffic =[R1,R2,...,R N ]

[0020] R i =[R t-H+0.1 ,R t-H+0.2 ,...,R t ]

[0021]

[0022] Among them, R Traffic R represents the collection of traffic situation characterization data. i Represents R Traffic Traffic situation representation data for the i-th trajectory, where H is the length of the time window. Represents R i The intermediate time is t h Data from the sampling points, This represents the absolute longitudinal position of the EV at that moment. This represents the absolute lateral position of the EV at that moment. This represents the absolute speed of the EV at that moment. This represents the absolute acceleration of the EV at that moment. This represents the heading angle of the EV at that moment. This represents the absolute longitudinal position of FV at that moment. This represents the absolute lateral position of FV at that moment. This represents the absolute velocity of FV at that moment. This represents the absolute acceleration of FV at that moment. This represents the heading angle of FV at that moment. This represents the relative velocity between EV and FV at that moment. The relative longitudinal positions of EV and FV at this moment This represents the relative lateral position of EV and FV at that moment.

[0023] Step 2: Establish and train a human-like trajectory generation model

[0024] (2.1) Network framework setup for human-like trajectory generation model:

[0025] The human-like trajectory generation model network uses a VAE model network structure as the initial model, and sets the input dimension of the training data to [batch_size, K0, feature_size], where batch_size is the number of human-like trajectories in ζ1 input for each training iteration, K0 is the number of sampling points of the trajectory in ζ1, and feature_size is the number of features of the trajectory in ζ1 at each sampling time point; the input layer size of the encoder network structure is set to match the input dimension of the training data, and the output layer size of the decoder network structure is set to match the input dimension of the training data.

[0026] (2.2) Loss function settings:

[0027] The loss function Loss1 is set as follows:

[0028]

[0029] Where L1 is the reconstruction error, which is designed as the trajectory in ζ1. The corresponding features of each sampling point and the trajectory generated by the human-like trajectory generation model network The mean squared error (MSE) between the corresponding features at each sampling point; L2 is the KL divergence loss term, which is designed as a variational posterior distribution N(μ,σ). 2 The KL divergence between the distribution and the standard normal distribution N(0,I), the mean μ and variance σ of the variational posterior distribution. 2 The encoder network structure is given, and the output is a encoded vector using a renormalization parameterization technique. ζ1 represents the normalized expert-injected trajectory data. The expert input trajectory data is generated by the human-like trajectory generation model network, and α is the adjustment parameter;

[0030] (2.3) Gradient descent updates network parameters:

[0031] After completing the above steps (2.1) and (2.2), each time a batch_size of trajectories is randomly selected from ζ1 as training data, input into the human-like trajectory generation model network for training, the loss function Loss1 is calculated, and the network parameters are updated by gradient descent based on the optimal Adam optimizer until the network converges.

[0032] Step 3: Establish a human-like trajectory coding space

[0033] (3.1) Obtaining human-like trajectory feature vectors:

[0034] Set the above humanoid trajectory generation model network to debug mode, no longer calculate gradients, and then input the trajectories in ζ1 sequentially to obtain the corresponding output humanoid trajectory feature vectors from the encoder; specifically, when the input trajectory in ζ1 is... Let χ be the output humanoid trajectory feature vector. i (where 1≤i≤N), χ i The data dimensions should match the size of the encoder's output layer;

[0035] (3.2) Saving of humanoid trajectory feature vectors:

[0036] The human-like trajectory feature vectors are sequentially saved to the human-like trajectory encoding space χ:

[0037] χ=[χ1,χ2,χ3,...,χ i ,...,χ N ]

[0038] Step 4: Establish and train a traffic situation mapping network

[0039] (4.1) Traffic situation mapping network framework setup:

[0040] The initial model uses the LSTM network structure. The traffic situation mapping network is set to contain a two-layer LSTM structure, with the input layer size set to the same as R. i The dimensions are matched, and the output layer size is set to match the χ² value. i The dimensions match;

[0041] (4.2) Loss function settings:

[0042] The loss function Loss2 is set as the mean squared error (MSE) between the human-like trajectory feature vector output by the traffic situation mapping network and the corresponding human-like trajectory feature vector obtained by the human-like trajectory generation model network in step three:

[0043]

[0044] in, The human-like trajectory encoding space is formed by the human-like trajectory feature vectors output by the traffic situation mapping network.

[0045] (4.3) Gradient descent updates network parameters:

[0046] After completing steps (4.1) and (4.2) above, R is randomly sampled for each training session. Traffic R in i(where 1≤i≤N) Input the traffic situation mapping network, calculate the loss function Loss2, update the network parameters through gradient descent based on the optimal Adam optimizer until the network converges, and complete the establishment and training of the traffic situation mapping network.

[0047] Furthermore, in step (1.1) during the acquisition and processing of expert-integrated teaching data, due to the randomness of the driver's operation and the unavoidable overshoot error and correction when the human driver operates the vehicle, the collected integration trajectory has random fluctuations. A polynomial fitting method is used to regularize the trajectory, and K0 sampling points are uniformly sampled on all trajectories on the time scale to make the collected trajectory smoother and the number of sampling points consistent.

[0048] The format of the expert-input trajectory data after the above processing is as follows:

[0049]

[0050]

[0051] Here, ζ0 represents the set of trajectory data input by experts, and each trajectory includes K0 sampling points. Let represent the i-th trajectory in ζ0, and let T be the merging cutoff time of this trajectory calculated from the merging time of zero. i , Represents the time of the t-th sampling point. This represents the vehicle's ordinate at that moment. This represents the x-coordinate of the vehicle at that moment.

[0052] To facilitate subsequent model training, the trajectory in ζ0 is normalized:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] The normalized format of the expert pool trajectory data is as follows:

[0060]

[0061] Since the data recorded in ζ1 differs from the data recorded in ζ0 only in numerical scale, their internal data formats will not be described in detail.

[0062] The present invention provides a method for generating a human-like trajectory for merging into an intelligent vehicle ramp, which, utilizing the aforementioned intelligent vehicle ramp merging human-like trajectory generation system, further includes the following steps:

[0063] Step Six: Acquisition and Processing of Real Traffic Situation Representation Data:

[0064] When an intelligent vehicle is merging onto a ramp, it utilizes onboard environmental perception devices such as radar and high-definition cameras, as well as onboard status monitoring devices such as wheel speedometers, high-precision positioning systems, and wheel angle sensors, to acquire absolute speed data, absolute longitudinal position data, absolute lateral position data, absolute acceleration data, and EV heading angle data, as well as absolute speed data, absolute longitudinal position data, absolute lateral position data, absolute acceleration data, and EV heading angle data, and relative speed data, relative longitudinal position data, and relative lateral position data between the EV and EV, and saves the latest traffic situation representation data (R) for the latest time window in real time. real-time In the storage medium; R real-time With R Traffic R in i (where 1≤i≤N) have the same data format;

[0065] Step 7: Obtaining Human-like Trajectory Feature Vectors:

[0066] Once the advanced decision-making unit of the intelligent vehicle decides to integrate, it will store the R in the storage medium. real-time The input is a traffic situation mapping network, which generates and outputs a human-like trajectory feature vector mapped to the current traffic situation based on the current traffic situation representation data. With χ in χ i (where 1≤i≤N) have the same data format;

[0067] Step 8: Humanoid Trajectory Generation:

[0068] Will The decoder of the human-like trajectory generation model is input, and the decoder... Decode the data and output the optimal human-like merging trajectory data under the current traffic conditions.

[0069] The beneficial effects of this invention are:

[0070] This invention provides an innovative intelligent vehicle ramp merging human-like trajectory generation system and method. It introduces a generative network to establish a human-like trajectory generation model, leveraging the powerful feature extraction capabilities of the generative network to achieve high-quality, human-like trajectory generation. By introducing human-like trajectory feature vectors and establishing a human-like trajectory encoding space, and utilizing a long short-term memory network to build a traffic situation mapping network, it achieves stable and diverse human-like trajectory generation. This results in stable, high-quality, diverse, and human-like ramp merging trajectories, improving the social acceptance and trust in intelligent vehicle ramp merging trajectory planning. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the training process of the present invention.

[0072] Figure 2 This is a schematic diagram illustrating the positional relationship between EV, FV, and OV in the ramp scenario of this invention.

[0073] Figure 3 This is a schematic diagram of the actual application process of the present invention.

[0074] Figure 4 This is a comparative schematic diagram of the humanoid trajectory generated by the present invention and the actual trajectory of a human driver. Detailed Implementation

[0075] Part 1: Model Training Process

[0076] See Figure 1 As shown, this invention provides an intelligent vehicle ramp merging humanoid trajectory generation system, which is established through the following steps:

[0077] Step 1: Obtain and build expert teaching dataset for ramp scenarios

[0078] (1.1) Acquisition and processing of expert-integrated teaching data:

[0079] An autonomous driving simulation software was used to construct a ramp scenario. A multi-driver in-loop driving simulator platform was established, and experienced drivers were recruited to operate the simulators, driving both following vehicles (FVs) on the main road and merging vehicles (EVs) in the ramp merging lanes. Other vehicles (OVs) were randomly generated using the autonomous driving simulation software to ensure high fidelity in the simulated traffic flow. The simulation scenario and the relative positions of each vehicle were as follows: Figure 2 As shown.

[0080] During the expert integration process, the system records the absolute velocity, absolute longitudinal position, absolute lateral position, absolute acceleration, and EV heading angle data for both the EV and FV, as well as their relative velocity, relative longitudinal position, and relative lateral position data. A total of N data points are sampled. The data sampling frequency is 0.1 s, and the real-time of all sampling points is recorded simultaneously during sampling. In this embodiment, N = 128.

[0081] The expert-integrated trajectory data uses the starting point of the EV merging state as the origin of the coordinate system, the actual time of the EV merging state as the zero moment of the merging time, the main road vehicle travel direction as the x-axis, and the 90-degree counterclockwise direction of the x-axis as the y-axis to establish a local rectangular coordinate system. The coordinates of the merging trajectory in the local rectangular coordinate system and the merging time of each coordinate point are recorded.

[0082] Due to the randomness of driver operation and the unavoidable overshoot error and correction during human vehicle operation, the collected merging trajectory exhibits random fluctuations. A polynomial fitting method is used to regularize the trajectory, and K0 sampling points are uniformly sampled across all trajectories on the time scale to ensure a smoother trajectory and a consistent number of sampling points. In this embodiment, K0 = 784.

[0083] The format of the expert-input trajectory data after the above processing is as follows:

[0084]

[0085]

[0086] Here, ζ0 represents the set of trajectory data input by experts, and each trajectory includes K0 sampling points. Let represent the i-th trajectory in ζ0, and let T be the merging cutoff time of this trajectory calculated from the merging time of zero. i , Represents the time of the t-th sampling point. This represents the vehicle's ordinate at that moment. This represents the x-coordinate of the vehicle at that moment.

[0087] To facilitate subsequent model training, the trajectory in ζ0 is normalized:

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] The normalized format of the expert pool trajectory data is as follows:

[0095]

[0096] Since the data recorded in ζ1 differs from the data recorded in ζ0 only in numerical scale, their internal data formats will not be described in detail.

[0097] (1.2) Traffic situation representation data acquisition and processing:

[0098] Based on the absolute velocity, absolute longitudinal position, absolute lateral position, absolute acceleration, and heading angle data of the EV and the FV obtained in step one (1.1), and the relative velocity, relative longitudinal position, and relative lateral position data between the EV and FV, the actual time of the starting point of the EV merging is taken as the ending time t. The data is extracted from 30 sampling points forward within the time window H, and after normalization, the traffic situation characterization data R is formed. Traffic :

[0099] R Traffic =[R1,R2,...,R N ]

[0100] R i =[R t-H+0.1 ,R t-H+0.2 ,...,R t ] H＝3s

[0101]

[0102] Among them, R Traffic R represents the collection of traffic situation characterization data. i Represents R Traffic The traffic situation representation data for the i-th trajectory, where H is the length of the time window, and in this embodiment H = 3s. Represents R i The intermediate time is t h Data from the sampling points, This represents the absolute longitudinal position of the EV at that moment. This represents the absolute lateral position of the EV at that moment. This represents the absolute speed of the EV at that moment. This represents the absolute acceleration of the EV at that moment. This represents the heading angle of the EV at that moment. This represents the absolute longitudinal position of FV at that moment. This represents the absolute lateral position of FV at that moment. This represents the absolute velocity of FV at that moment. This represents the absolute acceleration of FV at that moment. This represents the heading angle of FV at that moment. This represents the relative velocity between EV and FV at that moment. The relative longitudinal positions of EV and FV at this moment This represents the relative lateral position of EV and FV at that moment.

[0103] Step 2: Establish and train a human-like trajectory generation model

[0104] (2.1) Network framework setup for human-like trajectory generation model:

[0105] The human-like trajectory generation model network uses a VAE model network structure as the initial model. The VAE model is a variational autoencoder model, a generative neural network model, including two network structures: an encoder and a decoder. The encoder encodes the input information and outputs an encoded vector. The decoder receives the encoded vector output by the encoder and outputs the reconstructed input information of the encoder. The input dimension of the training data is set to [batch_size, K0, feature_size], where batch_size is the number of human-like trajectories in ζ1 input for each training iteration (in this embodiment, batch_size = 32), K0 is the number of sampling points in ζ1, and feature_size is the number of features in ζ1 at each sampling time point (based on the data format of ζ1, feature_size = 3). The input layer of the encoder network structure is set to match the input dimension of the training data, the hidden layer size is set to 300, the number of layers is two, and the output layer size is set to 30. The input layer size of the decoder network structure is set to 30, the hidden layer size is set to 300, the number of layers is two, and the output layer size is set to match the input dimension of the training data.

[0106] (2.2) Loss function settings:

[0107] The loss function Loss1 is set as follows:

[0108]

[0109] Where L1 is the reconstruction error, which is designed as the trajectory in ζ1. The corresponding features of each sampling point and the trajectory generated by the human-like trajectory generation model network The mean squared error (MSE) between the corresponding features at each sampling point, and L2 is the KL divergence loss term, which is designed as a variational posterior distribution N(μ,σ). 2 The KL divergence between the distribution and the standard normal distribution N(0,I), the mean μ and variance σ of the variational posterior distribution. 2 The encoder network structure is given, and the output is a encoded vector using a renormalization parameterization technique. ζ1 represents the normalized expert-injected trajectory data. The expert input trajectory data generated by the human-like trajectory generation model network is α, which is an adjustment parameter. In this embodiment, α = 0.1.

[0110] (2.3) Gradient descent updates network parameters:

[0111] After completing the above steps (2.1) and (2.2), each time a batch_size of trajectories is randomly selected from ζ1 as training data, input into the human-like trajectory generation model network for training, the loss function Loss1 is calculated, and the network parameters are updated by gradient descent based on the optimal Adam optimizer until the network converges.

[0112] Step 3: Establish a human-like trajectory coding space

[0113] (3.1) Obtaining human-like trajectory feature vectors:

[0114] Set the above humanoid trajectory generation model network to debug mode, no longer calculate gradients, and then input the trajectories in ζ1 sequentially to obtain the corresponding output humanoid trajectory feature vectors from the encoder; specifically, when the input trajectory in ζ1 is... Let χ be the output humanoid trajectory feature vector. i (where 1≤i≤N), χ i The data dimensions should match the size of the encoder's output layer;

[0115] (3.2) Saving of humanoid trajectory feature vectors:

[0116] The human-like trajectory feature vectors are sequentially saved to the human-like trajectory encoding space χ:

[0117] χ=[χ1,χ2,χ3,...,χ i ,...,χ N ]

[0118] Step 4: Establish and train a traffic situation mapping network

[0119] (4.1) Traffic situation mapping network framework setup:

[0120] The initial model is an LSTM network structure, which is a recurrent neural network model. The traffic situation mapping network is configured with a two-layer LSTM structure, and the input layer size is set to R. i The dimensions are matched, the hidden layer size is set to 200, the number of layers is three, and the output layer size is set to match the x-axis. i The dimensions match;

[0121] (4.2) Loss function settings:

[0122] The loss function Loss2 is set as the mean squared error (MSE) between the human-like trajectory feature vector output by the traffic situation mapping network and the corresponding human-like trajectory feature vector obtained by the human-like trajectory generation model network in step three:

[0123]

[0124] in, The human-like trajectory encoding space is formed by the human-like trajectory feature vectors output by the traffic situation mapping network.

[0125] (4.3) Gradient descent updates network parameters:

[0126] After completing steps (4.1) and (4.2) above, R is randomly sampled for each training session. Traffic R in i (where 1≤i≤N) Input the traffic situation mapping network, calculate the loss function Loss2, update the network parameters through gradient descent based on the optimal Adam optimizer until the network converges, and complete the establishment and training of the traffic situation mapping network.

[0127] Part Two: Application Steps

[0128] See Figure 3 As shown, the present invention provides a method for generating a human-like trajectory for merging into an intelligent vehicle ramp, which, utilizing the aforementioned intelligent vehicle ramp merging human-like trajectory generation system, further includes the following steps:

[0129] Step 1: Traffic Situation Data Acquisition and Processing

[0130] When an intelligent vehicle is merging onto a ramp, it utilizes onboard environmental perception devices such as radar and high-definition cameras, as well as onboard status monitoring devices such as wheel speedometers, high-precision positioning systems, and wheel angle sensors, to acquire absolute speed data, absolute longitudinal position data, absolute lateral position data, absolute acceleration data, and EV heading angle data, as well as absolute speed data, absolute longitudinal position data, absolute lateral position data, absolute acceleration data, and EV heading angle data, and relative speed data, relative longitudinal position data, and relative lateral position data between the EV and EV. It also saves the latest traffic situation representation data (R) with a time window of 3 seconds in real time. real-time In the storage medium. R real-time With R Traffic R in i (where 1≤i≤N) have the same data format, which will not be elaborated here.

[0131] Step 2: Obtaining Human-like Trajectory Feature Vectors:

[0132] Once the advanced decision-making unit of the intelligent vehicle decides to integrate, it will store the R in the storage medium. real-time The input is a traffic situation mapping network, which generates and outputs a human-like trajectory feature vector mapped to the current traffic situation based on the current traffic situation representation data. With χ in χ i (where 1≤i≤N) have the same data format, which will not be elaborated here.

[0133] Step 3: Human-like trajectory generation:

[0134] Will The decoder of the human-like trajectory generation model is input, and the decoder... Decode the data and output the optimal human-like merging trajectory data under the current traffic conditions.

[0135] See appendix Figure 4 The comparative examples show a comparison between the humanoid trajectory generated by the method described in this invention and the actual trajectory of a human driver. It is evident that the trajectory generated by the method described in this invention has a high degree of human-likeness and is smooth. The longitudinal error at the merging termination point is 0.1624 meters, the lateral error at the merging termination point is 0.0221 meters, and the merging process time error is 0.0095 seconds, all within acceptable ranges, demonstrating the effectiveness of the humanoid trajectory generation method described in this invention.

Claims

1. A human-like trajectory generation system for intelligent vehicle ramp merging, characterized in that: Established through the following steps: Step 1: Obtain and build expert teaching dataset for ramp scenarios (1.1) Acquisition and processing of expert-integrated teaching data: Using autonomous driving simulation software, the system records the absolute speed, absolute longitudinal position, absolute lateral position, absolute acceleration, and EV heading angle of the merging vehicle (EV) on the ramp merging lane during the expert merging process, as well as the absolute speed, absolute longitudinal position, absolute lateral position, absolute acceleration, FV heading angle, and relative speed, relative longitudinal position, and relative lateral position of the following vehicle (FV) on the main road. A total of N data points are sampled. The data sampling frequency is 0.1s, and the real time of all sampling points is recorded simultaneously during sampling. The expert-integrated trajectory data uses the starting point of the EV integration state as the origin of the coordinate system, the actual time of the EV integration state as the zero moment of the integration time, the main road vehicle travel direction as the x-axis, and the y-axis as the direction 90 degrees counterclockwise from the x-axis as the y-axis to record the coordinates of the integration trajectory in the local rectangular coordinate system and the integration time of each coordinate point. After normalizing the trajectories in the expert-input trajectory data, the resulting expert-input trajectory data format is as follows: Where ζ1 represents the set of normalized expert-integrated trajectory data, and each trajectory includes K0 sampling points. Let represent the i-th trajectory in the set, and let T be the merging cutoff time of this trajectory calculated from the merging time of zero. i , Represents the time of the t-th sampling point. This represents the vehicle's ordinate at that moment. The x-coordinate of the vehicle at that moment; (1.2) Traffic situation representation data acquisition and processing: Based on the absolute velocity, absolute longitudinal position, absolute lateral position, absolute acceleration, and heading angle data of the EV and the FV obtained in step one (1.1), and the relative velocity, relative longitudinal position, and relative lateral position data between the EV and FV, the actual time of the starting point of the EV merging is taken as the ending time t. Several sampling points are traversed forward to extract the data within the time window H. After normalizing the data, the traffic situation characterization data R is formed. Traffic : R Traffic =[R1,R2,...,R N ] R i =[R t-H+0.1 ,R t-H+0.2 ,...,R t ] Among them, R Traffic R represents the collection of traffic situation characterization data. i Represents R Traffic Traffic situation representation data for the i-th trajectory, where H is the length of the time window. Represents R i The intermediate time is t h Data from the sampling points, This represents the absolute longitudinal position of the EV at that moment. This represents the absolute lateral position of the EV at that moment. This represents the absolute speed of the EV at that moment. This represents the absolute acceleration of the EV at that moment. This represents the heading angle of the EV at that moment. This represents the absolute longitudinal position of FV at that moment. This represents the absolute lateral position of FV at that moment. This represents the absolute velocity of FV at that moment. This represents the absolute acceleration of FV at that moment. This represents the heading angle of FV at that moment. This represents the relative velocity between EV and FV at that moment. The relative longitudinal positions of EV and FV at this moment This represents the relative lateral position of EV and FV at that moment; Step 2: Establish and train a human-like trajectory generation model (2.1) Network framework setup for human-like trajectory generation model: The human-like trajectory generation model network uses a VAE model network structure as the initial model, and sets the input dimension of the training data to [batch_size, K0, feature_size], where batch_size is the number of human-like trajectories in ζ1 input for each training iteration, K0 is the number of sampling points in ζ1, and feature_size is the number of features in ζ1 at each sampling time point; the input layer size of the encoder network structure is set to match the input dimension of the training data, and the output layer size of the decoder network structure is set to match the input dimension of the training data. (2.2) Loss function settings: The loss function Loss1 is set as follows: Where L1 is the reconstruction error, L2 is the KL divergence loss term, and ζ1 is the normalized expert-integrated trajectory data. The expert input trajectory data is generated by the human-like trajectory generation model network, and α is the adjustment parameter; (2.3) Gradient descent updates network parameters: After completing the above steps (2.1) and (2.2), each time a batch_size of trajectories are randomly selected from ζ1 as training data, input into the human-like trajectory generation model network for training, calculate the loss function Loss1, and update the network parameters through gradient descent based on the optimal Adam optimizer until the network converges; Step 3: Establish a human-like trajectory coding space (3.1) Obtaining human-like trajectory feature vectors: Set the above humanoid trajectory generation model network to debug mode, no longer calculate gradients, and then input the trajectories in ζ1 sequentially to obtain the corresponding output humanoid trajectory feature vectors from the encoder; when the input trajectory in ζ1 is... Let χ be the output humanoid trajectory feature vector. i , χ i The data dimensions should match the size of the encoder's output layer; (3.2) Saving of humanoid trajectory feature vectors: The human-like trajectory feature vectors are sequentially saved to the human-like trajectory encoding space χ: x=[x1,x2,x3,...,x i ,...,x N ] Step 4: Establish and train a traffic situation mapping network (4.1) Traffic situation mapping network framework setup: The initial model uses the LSTM network structure. The traffic situation mapping network is set to contain a two-layer LSTM structure, with the input layer size set to the same as R. i The dimensions are matched, and the output layer size is set to match the χ² value. i The dimensions match; (4.2) Loss function settings: The loss function Loss2 is set as the mean square error between the human-like trajectory feature vector output by the traffic situation mapping network and the corresponding human-like trajectory feature vector obtained by the human-like trajectory generation model network in step three. in, The human-like trajectory encoding space is formed by the human-like trajectory feature vectors output by the traffic situation mapping network. (4.3) Gradient descent updates network parameters: After completing steps (4.1) and (4.2) above, R is randomly sampled for each training session. Traffic R in i Input the traffic situation mapping network, calculate the loss function Loss2, and update the network parameters through gradient descent based on the optimal Adam optimizer until the network converges, thus completing the establishment and training of the traffic situation mapping network.

2. The intelligent vehicle ramp merging humanoid trajectory generation system according to claim 1, characterized in that: In step (1.1) during the acquisition and processing of expert-integrated teaching data, due to the randomness of the driver's operation and the unavoidable overshoot error and correction when the human driver operates the vehicle, the collected integration trajectory has random fluctuations. The multinomial fitting method is used to regularize the trajectory, and K0 sampling points are uniformly sampled on the time scale for all trajectories to make the collected trajectory smoother and the number of sampling points consistent. The format of the expert-input trajectory data after the above processing is as follows: Here, ζ0 represents the set of trajectory data input by experts, and each trajectory includes K0 sampling points. Let represent the i-th trajectory in ζ0, and let T be the merging cutoff time of this trajectory calculated from the merging time of zero. i , Represents the time of the t-th sampling point. This represents the vehicle's ordinate at that moment. The x-coordinate of the vehicle at that moment; To facilitate subsequent model training, the trajectory in ζ0 is normalized: The normalized format of the expert pool trajectory data is as follows:

3. The intelligent vehicle ramp merging humanoid trajectory generation system according to claim 1, characterized in that: In step (2.2) of setting the loss function, the reconstruction error L1 of the loss function Loss1 is designed as the trajectory in ζ1. The corresponding features of each sampling point and the trajectory generated by the human-like trajectory generation model network The mean square error between the corresponding features at each sampling point; the KL divergence loss term L2 is designed as a variational posterior distribution N(μ,σ) 2 The KL divergence between the distribution and the standard normal distribution N(0,I), the mean μ and variance σ of the variational posterior distribution. 2 The encoder network structure is given, and the output is a coded vector through renormalization parameterization techniques.

4. A method for generating human-like trajectories for intelligent vehicle ramp merging, characterized in that: Using the intelligent vehicle ramp merging humanoid trajectory generation system according to any one of claims 1-3, the following operation steps are performed: Step 1: Acquisition and Processing of Real Traffic Situation Representation Data: When an intelligent vehicle is merging onto a ramp, it uses its onboard environmental perception and status monitoring devices to acquire absolute speed, absolute longitudinal position, absolute lateral position, absolute acceleration, and EV heading angle data, as well as absolute speed, absolute longitudinal position, absolute lateral position, absolute acceleration, and EV heading angle data, and relative speed, relative longitudinal position, and relative lateral position data between the EV and EV. It also saves the latest traffic situation representation data (R) for the latest time window in real time. real-time In the storage medium; R real-time With R Traffic R in i They have the same data format; Step 2: Obtaining Human-like Trajectory Feature Vectors: Once the advanced decision-making unit of the intelligent vehicle decides to integrate, it will store the R in the storage medium. real-time The input is a traffic situation mapping network, which generates and outputs a human-like trajectory feature vector mapped to the current traffic situation based on the current traffic situation representation data. With χ in χ i They have the same data format; Step 3: Humanoid trajectory generation: Will The decoder of the human-like trajectory generation model is input, and the decoder... Decode the data and output the optimal human-like merging trajectory data under the current traffic conditions.

Images