A lane change trajectory generation method

Abstract

This invention provides a lane change trajectory generation method. It processes multimodal dynamic time series data through a risk level feature extraction network to generate a high-dimensional risk feature vector representing the lane change risk level. Vehicle dynamic parameter features are extracted and concatenated with the high-dimensional risk feature vector. This is then input into an autoencoder module for dimensionality reduction to generate a low-dimensional latent vector representation. The low-dimensional latent vector representation and the high-dimensional risk feature vector are then used as joint conditional variables and input into a conditional generative adversarial network module. Using the high-dimensional risk feature vector as a condition, the method generates lane change trajectories that meet the specified risk level. This method provides efficient and reliable test data support for autonomous driving algorithms.

Description

Technical Field

[0001] The present invention relates to the field of assisted driving technology, and in particular to a method for generating a lane change trajectory. Background Art

[0002] With the rapid development of autonomous driving technology, the importance of vehicle safety testing in critical scenarios has become increasingly prominent. Lane changes in autonomous vehicles are a major cause of accidents and have attracted extensive research attention. However, real-world vehicle trajectories exhibit a high degree of repetitiveness and low risk. A challenging task is to generate parameters for high-risk lane change scenarios, as these scenarios can expose design flaws in autonomous driving systems and drive improvements.

[0003] Existing approaches fall into two main categories: rule-driven approaches and generative adversarial networks. Rule-driven approaches generate deterministic trajectories based on preset metrics such as collision time and safety distance, but they struggle to simulate the complex interactions experienced in real driving, resulting in a single, monotonous generated scenario. Generative adversarial networks, such as DSA-GAN and CSGAN, generate trajectory data through unsupervised learning. However, existing models lack the ability to capture long-term temporal dependencies and an explicit embedding mechanism for risk features, resulting in a poor match between generated trajectories and the actual risk distribution. Furthermore, feature extraction often relies on traditional clustering or LSTM networks, which are sensitive to noise and have limited classification accuracy, making them incapable of supporting the refined generation of high-risk scenarios. Summary of the Invention

[0004] The present invention proposes a lane change trajectory generation method to solve the technical problem that existing technical solutions are difficult to achieve high-fidelity and multi-risk level lane change trajectory generation.

[0005] To solve the above technical problems, the present invention provides a lane change trajectory generation method, comprising the following steps:

[0006] Step S1: Processing multimodal dynamic time series data through a risk level feature extraction network to generate a high-dimensional risk feature vector representing the lane change risk level;

[0007] The multimodal dynamic time series data includes vehicle lateral acceleration, longitudinal speed, steering angle change rate and minimum distance to surrounding vehicles;

[0008] Step S2: extracting vehicle dynamics parameter features, concatenating them with the high-dimensional risk feature vector, and inputting them into the autoencoder module for dimensionality reduction to generate a low-dimensional latent vector representation;

[0009] Step S3: The low-dimensional latent vector representation and the high-dimensional risk feature vector are used as joint conditional variables and input into a conditional generative adversarial network module. The high-dimensional risk feature vector is used as a condition to generate a lane change trajectory that meets the specified risk level.

[0010] Preferably, the risk level feature extraction network in step S1 performs a feature extraction process of first increasing the dimension and then reducing the dimension through several layers of temporal convolutional networks (TCN) to obtain a high-dimensional risk feature vector.

[0011] Preferably, the temporal convolutional network TCN includes: a causal convolution layer, a weight normalization layer, a ReLU activation function layer and a Dropout layer;

[0012] The causal convolution layer is used to capture the long-term dependencies of the multimodal dynamic time series data;

[0013] The weight normalization layer is used to normalize the convolution kernel parameters;

[0014] The ReLU activation function layer is used to introduce nonlinear feature mapping;

[0015] The Dropout layer is used to randomly mask neuron outputs at a preset dropout rate.

[0016] Preferably, the encoder module includes a temporal convolutional layer and a gated recurrent unit (GRU), and the output end is connected to an average pooling layer and a global pooling layer.

[0017] Preferably, the autoencoder module generates a low-dimensional latent vector by the following formula:

[0018] ;

[0019] ;

[0020] ;

[0021] Where, represents the activation function; and Represent the maximum pooling and average pooling operations respectively; FC represents a fully connected layer; Indicates feature extraction through TCN network; represents the vehicle dynamics parameters; Represents a high-dimensional risk feature vector.

[0022] Preferably, the training process of the conditional generative adversarial network module includes:

[0023] The generator generates simulated lane change trajectories based on the low-dimensional latent vector, conditioned on the high-dimensional risk feature vector;

[0024] The discriminator calculates the adversarial loss based on the distribution difference between the real trajectory dataset and the generated trajectory to constrain the discriminator parameter update until the diversity and authenticity of the generated trajectory meet the preset threshold.

[0025] Preferably, the multimodal dynamic time series data is normalized and preprocessed using the following formula before being input into the risk level feature extraction network:

[0026] ;

[0027] Where, Input Represents input data; Input min Indicates the minimum value of the input data; Input max Indicates the maximum value of the input data.

[0028] Preferably, the risk level feature extraction network performs classification based on collision time, dividing lane change risks into three levels:

[0029] ;

[0030] Where, Indicates the time until the collision occurs.

[0031] Preferably, the risk level feature extraction network is trained using a cross entropy loss function.

[0032] Preferably, the expression of the cross entropy loss function is:

[0033] ;

[0034] In the formula, C represents the number of categories; Indicates the i The probability distribution of the true labels; Indicates the i The predicted probability of each category.

[0035] The beneficial effects of the present invention include at least the following: The present invention extracts high-dimensional risk feature vectors from multimodal dynamic time series data through a risk level feature extraction network, then reduces the dimensionality of these high-dimensional features through an autoencoder module to generate a low-dimensional latent vector representation; a conditional generative adversarial network uses the high-dimensional risk feature vector as a condition, and the generator and discriminator optimize the low-dimensional feature representation of the autoencoder during training; the generator attempts to generate trajectories that conform to the distribution of real data, while the discriminator determines the authenticity of the generated trajectories. This provides efficient and reliable test data support for autonomous driving algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention;

[0037] Figure 2 This is a comparison diagram of the reconstructed signals and original signals of four risky lane change trajectories of the autoencoder according to an embodiment of the present invention;

[0038] Figure 3 Schematic diagram of the convergence effect of the autoencoder according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts are within the scope of protection of the present invention.

[0040] like Figure 1 As shown, an embodiment of the present invention provides a method for generating a lane change trajectory, comprising the following steps:

[0041] Step S1: Process the multimodal dynamic time series data through the risk level feature extraction network to generate a high-dimensional risk feature vector that represents the lane change risk level.

[0042] The multimodal dynamic time series data includes vehicle lateral acceleration, longitudinal velocity, steering angle change rate and minimum distance to surrounding vehicles.

[0043] Specifically, the risk level feature extraction network first extracts multimodal dynamic time series data from the vehicle during driving, forming a dynamic parameter matrix. These dynamic parameters are then correlated with lane change risk to form a risk feature vector. Finally, the risk feature vector is input into a neural network model. After multiple layers of convolution and pooling operations, a high-dimensional feature vector is generated. This feature vector represents the risk level of the vehicle during lane changes, helping us better assess lane change safety.

[0044] In this embodiment, the risk level feature extraction network, within the larger network architecture, primarily processes multimodal dynamic time series data, including speed, acceleration, minimum distance to surrounding vehicles, and minimum collision time with surrounding vehicles, totaling 7×80 dimensions. Through TCN processing, a vector representing the comprehensive risk characteristics is generated, maintaining the same dimensionality. Fault classification is then performed on this vector to verify the effectiveness of the comprehensive risk feature extraction.

[0045] The main structure of this part of the network is composed of a TCN network. In order to improve the comprehensive extraction capability of the TCN network for multimodal input vector risks, this embodiment designs the network as a classification network. By classifying the risk level type of the extracted comprehensive risk feature vector, the effectiveness of feature extraction is verified. The following describes part of the structure and process of the risk level feature extraction network:

[0046] 1) Input layer

[0047] The input signal is , whose shape is 7×80, where 80 is the time step and 7 is the feature dimension.

[0048] 2) TCN layer

[0049] Each TCN layer consists of the following components:

[0050] Dilated Causal Convolution:

[0051] The input is x, the output is y, the convolution kernel size is k, and the dilation factor is d, then the output y can be expressed as:

[0052] ;

[0053] in is the weight of the convolution kernel.

[0054] Weight Normalization:

[0055] Weight normalization is to normalize the convolution kernel so that the norm of the weight matrix remains within a certain range. The specific formula is as follows:

[0056] ;

[0057] Where W is the original weight matrix, g is the scaling factor, is the normalized weight matrix.

[0058] Activation function (ReLU Activation):

[0059] .

[0060] Dropout layer:

[0061] Dropout is a regularization technique that randomly discards some of the neuron outputs to prevent overfitting. Assuming the dropout rate is p, the output can be expressed as:

[0062] .

[0063] 3) Final classification layer

[0064] The final feature vector is passed through a fully connected layer for classification:

[0065] ;

[0066] in, and are the weight matrix and bias term of the fully connected layer respectively.

[0067] In the embodiment of the present invention, the multimodal dynamic time series data is normalized and pre-processed using the following formula before being input into the risk level feature extraction network:

[0068] ;

[0069] Where, Input Represents input data; Input min Indicates the minimum value of the input data; Input max Indicates the maximum value of the input data.

[0070] In this embodiment, the risk level feature extraction network classifies lane change risks based on collision time and divides them into three levels:

[0071] ;

[0072] Where, Indicates the time until the collision occurs.

[0073] During the training process, cross entropy is used as the loss function, which is expressed as:

[0074] ;

[0075] In the formula, C represents the number of categories; Indicates the i The probability distribution of the true labels; Indicates the i The predicted probability of each category.

[0076] During operation, multimodal dynamic time series data is input into the network. It then passes through several layers of the TCN network, performing a feature extraction process that involves first increasing the dimensionality and then reducing it, ultimately generating a high-dimensional risk feature vector. These features are then converted into probabilities for three lane change risk levels through a linear layer. To prevent overfitting, a dropout layer with a dropout rate of 0.2 is used in the network. The network is trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32 for 1000 epochs.

[0077] Step S2: Input the high-dimensional risk feature vector into the autoencoder module for dimensionality reduction, and generate a low-dimensional latent vector representation through the encoder.

[0078] Specifically, in this embodiment, the autoencoder network compresses and then decompresses the training data, ensuring that the final decompressed data closely matches the original data. We chose an autoencoder because the original dataset contained significant noise and mutations. By training the autoencoder with a gated recurrent unit (GRU), the lane change data exhibits significantly improved smoothness while retaining its inherent properties.

[0079] An autoencoder consists of two main parts: an encoder and a decoder. The encoder compresses the time series into a latent vector that captures the essential features of the input; the decoder reconstructs the original time series from the latent vector.

[0080] The core component of the autoencoder is the gated recurrent unit (GRU). Compared to the long short-term memory (LSTM), the GRU only contains update gates and reset gates, and has fewer parameters, but is equally good at handling long-term dependencies in time series.

[0081] When selecting vehicle dynamic parameters such as speed and acceleration as input, we first use a three-layer TCN network to extract features. These features are then concatenated with the high-dimensional risk feature vector obtained by the risk feature extraction network and used as the input to the GRU network:

[0082] ;

[0083] Where, represents the vehicle dynamics parameters, Represents a high-dimensional risk feature vector.

[0084] After being processed by the N-layer GRU network, the output The GRU output is subjected to maximum pooling and average pooling. After the two are aggregated through the fully connected layer, the LeakyReLU function is applied to obtain the encoder's latent vector , whose expression is:

[0085] ;

[0086] ;

[0087] ;

[0088] Where, represents the activation function; and Represent the maximum pooling and average pooling operations respectively; FC represents a fully connected layer; Indicates feature extraction through TCN network; represents the vehicle dynamics parameters; Represents a high-dimensional risk feature vector.

[0089] The decoding process is the inverse process of encoding, and the original time series is reconstructed through the fully connected layer and GRU network.

[0090] In summary, in order to generate lane-changing trajectory signals that are as consistent as possible with the original signal characteristics, the autoencoder not only extracts the global features of the signal through the TCN network when the signal is input, but also uses average pooling and global pooling after sending the signal to the GRU for processing, so as to further retain the overall information and local features of the signal, improve the sparsity of the feature map, and improve the learning ability of the subsequent layers. The comparison between the lane-changing trajectories of different risk levels reconstructed by the autoencoder in this embodiment and the original lane-changing trajectories is shown in Figure 2. Figure 2 As shown, it can be seen that the smoothness of the effect generated by the autoencoder using this embodiment in the case of feature input is significantly improved.

[0091] The improved network ensures the characteristics of the lane-changing trajectory itself on the basis of improved smoothness. After the autoencoder is trained, the convergence effect is as follows: Figure 3 shown.

[0092] Step S3: The low-dimensional latent vector representation and the high-dimensional risk feature vector are used as joint conditional variables and input into the conditional generative adversarial network module. The high-dimensional risk feature vector is used as a condition to generate a lane change trajectory that meets the specified risk level.

[0093] In this embodiment of the present invention, the risk level feature extraction network first extracts a high-dimensional risk feature vector from the vehicle's dynamic parameters, containing complex information about lane change risk. The autoencoder network then performs dimensionality reduction on these high-dimensional features, compressing them into a low-dimensional vector representation through an encoder. This process not only reduces the data dimensionality but also removes redundant information and noise, preserving key risk features.

[0094] The decoder of the autoencoder reconstructs the low-dimensional vectors back to their original data form. This reconstruction process helps optimize the feature representation of the risk feature extraction network. Through the reconstruction error feedback from the autoencoder, the risk level feature extraction network can further adjust its parameters, making the extracted risk features more accurate and robust.

[0095] In a conditional generative adversarial network (cGAN), the risk feature vector generated by the risk level feature extraction network is used as a conditioning variable and input into the generator and discriminator. This conditioning variable guides the generator to generate lane change trajectories that meet specific risk characteristics, making the generated trajectories more consistent with the risk distribution in real-world driving scenarios.

[0096] The cGAN's discriminator evaluates the generated lane change trajectories to determine whether they conform to the distribution of real-world data. This feedback is then backpropagated to the risk level feature extraction network, helping it further optimize its risk feature extraction process. Through this feedback mechanism, the risk level feature extraction network continuously adjusts its parameters, making the extracted risk features more accurate, thereby improving the quality of the generator.

[0097] The low-dimensional feature vectors generated by the autoencoder network serve as input to the cGAN, which is then used to generate lane change trajectories. Because the autoencoder has already performed dimensionality reduction and denoising on the data, the generator can focus more on key features when generating trajectories, reducing noise interference and generating more realistic and reasonable lane change trajectories.

[0098] During training, the cGAN's generator and discriminator optimize the autoencoder's low-dimensional feature representation. The generator attempts to generate trajectories that conform to the true data distribution, while the discriminator determines the authenticity of the generated trajectories. This process encourages the autoencoder to generate more discriminative low-dimensional features, thereby improving the autoencoder's feature extraction capabilities.

[0099] Overall, the risk level feature extraction network, the autoencoder network, and the cGAN form a closed-loop system. The high-dimensional risk features extracted by the risk level feature extraction network are reduced in dimension by the autoencoder and then fed into the cGAN to generate lane change trajectories. The authenticity of the generated trajectories is fed back to the risk level feature extraction network and the autoencoder via a discriminator, further optimizing the feature extraction and dimensionality reduction process. The risk level feature extraction network extracts risk features from the raw data, the autoencoder performs dimensionality reduction and denoising on these features, and the cGAN generates lane change trajectories that match the risk features. Through continuous feedback and optimization, the entire process gradually improves the overall performance of the system.

[0100] The technical features of the above embodiments may be combined in any manner. To simplify the description, not all possible combinations of the technical features in the above embodiments are described. Only preferred embodiments of the present invention are presented. While the description is relatively specific and detailed, it should not be construed as limiting the scope of the present invention. As long as there are no conflicts in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] It should be noted that, for those skilled in the art, various modifications and improvements can be made without departing from the scope of the present invention, and these modifications and improvements fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A lane change trajectory generation method, characterized by: The following steps are involved: Step S1: Processing multimodal dynamic time series data through a risk level feature extraction network to generate a high-dimensional risk feature vector representing the lane change risk level; The multimodal dynamic time series data includes vehicle lateral acceleration, longitudinal speed, steering angle change rate and minimum distance to surrounding vehicles; Step S2: extracting vehicle dynamics parameter features, concatenating them with the high-dimensional risk feature vector, and inputting them into the autoencoder module for dimensionality reduction to generate a low-dimensional latent vector representation; Step S3: The low-dimensional latent vector representation and the high-dimensional risk feature vector are used as joint conditional variables and input into a conditional generative adversarial network module, which generates a lane change trajectory that meets the specified risk level using the high-dimensional risk feature vector as a condition. The encoder module includes a temporal convolutional layer and a gated recurrent unit (GRU), and the output end is connected to an average pooling layer and a global pooling layer; The autoencoder module generates a low-dimensional latent vector using the following formula: ； ； ； Where, represents the activation function; and Represent the maximum pooling and average pooling operations respectively; FC represents a fully connected layer; Indicates feature extraction through TCN network; represents the vehicle dynamics parameters; Represents a high-dimensional risk feature vector.

2. The lane change trajectory generation method according to claim 1, characterized in that: The risk level feature extraction network in step S1 performs a feature extraction process of first increasing the dimension and then reducing the dimension through several layers of temporal convolutional networks (TCN) to obtain a high-dimensional risk feature vector.

3. The lane change trajectory generation method according to claim 2, characterized in that: The temporal convolutional network TCN includes: a causal convolution layer, a weight normalization layer, a ReLU activation function layer and a Dropout layer; The causal convolution layer is used to capture the long-term dependencies of the multimodal dynamic time series data; The weight normalization layer is used to normalize the convolution kernel parameters; The ReLU activation function layer is used to introduce nonlinear feature mapping; The Dropout layer is used to randomly mask neuron outputs at a preset dropout rate.

4. The lane change trajectory generation method according to claim 1, characterized in that: The training process of the conditional generative adversarial network module includes: The generator generates simulated lane change trajectories based on the low-dimensional latent vector, conditioned on the high-dimensional risk feature vector; The discriminator calculates the adversarial loss based on the distribution difference between the real trajectory dataset and the generated trajectory to constrain the discriminator parameter update until the diversity and authenticity of the generated trajectory meet the preset threshold.

5. The lane change trajectory generation method according to claim 1, characterized in that: The multimodal dynamic time series data is normalized and preprocessed using the following formula before being input into the risk level feature extraction network: ； Where, Input Represents input data; Input min Indicates the minimum value of the input data; Input max Indicates the maximum value of the input data.

6. The lane change trajectory generation method according to claim 1, characterized in that: The risk level feature extraction network classifies lane change risks into three levels based on collision time: ； Where, Indicates the time until the collision occurs.

7. The lane change trajectory generation method according to claim 6, characterized in that: The risk level feature extraction network is trained using a cross entropy loss function.

8. The lane change trajectory generation method according to claim 7, characterized in that: The expression of the cross entropy loss function is: ； In the formula, C represents the number of categories; Indicates the i The probability distribution of the true labels; Indicates the i The predicted probability of each category.

Images