Lane changing track generation method

Abstract

The invention provides a lane change track generation method, which comprises the following steps of: processing multi-modal dynamic time sequence data through a risk level feature extraction network, and generating a high-dimensional risk feature vector representing a lane change risk level; vehicle kinetic parameter features are extracted and then spliced with the high-dimensional risk feature vectors, the vehicle kinetic parameter features are input into an auto-encoder module for dimension reduction processing, and low-dimensional potential vector representation is generated; and taking the low-dimensional potential vector representation and the high-dimensional risk feature vector as a joint conditional variable, inputting the joint conditional variable into a conditional generative adversarial network module, and taking the high-dimensional risk feature vector as a condition to generate a lane changing track conforming to a specified risk level. And an efficient and reliable test data support is provided for an automatic driving algorithm.

Description

Technical Field

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

[0002] With the rapid development of autonomous driving technology, the importance of safety testing of vehicles in critical scenarios has become increasingly prominent. The lane change operation of autonomous vehicles is one of the important causes of vehicle accidents and has received extensive research attention in the academic community. However, the vehicle trajectories in the real world exhibit a large amount of repetition and low risk. A challenging task is how to generate the parameters of high-risk lane change scenarios, because only these scenarios can expose the design defects of the autonomous driving system and thus promote improvement.

[0003] Existing solutions are mainly divided into two categories: rule-driven methods and generative adversarial networks. Among them, rule-driven methods generate deterministic trajectories based on preset indicators such as time to collision and safety distance, but it is difficult to simulate complex interaction behaviors in real driving, resulting in a single generated scenario; generative adversarial networks, such as DSA-GAN, CSGAN, etc., generate trajectory data through unsupervised learning. However, the existing models have insufficient ability to capture long-term time dependence relationships and lack an explicit embedding mechanism for risk features, resulting in a low matching degree between the generated trajectories and the real risk distribution. In addition, feature extraction mostly uses traditional clustering or LSTM networks, which are sensitive to noise and have limited classification accuracy, making it difficult to support the refined generation of high-risk scenarios. Summary of the Invention

[0004] The present invention proposes a method for generating a lane change trajectory to solve the technical problem that it is difficult for existing technical solutions to generate lane change trajectories with high authenticity and multiple risk levels.

[0005] To solve the above technical problem, the present invention provides a method for generating a lane change trajectory, including the following steps: Step S1: Process multi-modal 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 multi-modal dynamic time series data includes vehicle lateral acceleration, longitudinal speed, steering angle change rate, and the minimum distance from surrounding vehicles; Step S2: Extract vehicle dynamics parameter features and then splice them with the high-dimensional risk feature vector, and input them into an autoencoder module for dimensionality reduction processing to generate a low-dimensional latent vector representation; Step S3: Use the low-dimensional latent vector representation and the high-dimensional risk feature vector as joint conditional variables, input them into a conditional generative adversarial network module, and use the high-dimensional risk feature vector as a condition to generate a lane change trajectory that meets the specified risk level.

[0006] Preferably, in step S1, the risk level feature extraction network performs a feature extraction process of first increasing the dimension and then decreasing the dimension through several layers of the Temporal Convolutional Network (TCN) to obtain a high-dimensional risk feature vector.

[0007] Preferably, the Temporal Convolutional Network (TCN) includes: a causal convolutional layer, a weight normalization layer, a ReLU activation function layer, and a Dropout layer; The causal convolutional layer is used to capture the long-term dependency relationship of the multi-modal dynamic time series data; The weight normalization layer is used to perform normalization constraints on the convolutional kernel parameters; The ReLU activation function layer is used to introduce non-linear feature mapping; The Dropout layer is used to randomly mask the neuron outputs at a preset dropout rate.

[0008] Preferably, the encoder module includes a temporal convolutional layer and a Gated Recurrent Unit (GRU), and its output end is connected to an average pooling layer and a global pooling layer.

[0009] Preferably, the autoencoder module generates a low-dimensional latent vector through the following formula: ; ; ; In the formula, represents the activation function; and respectively represent the max pooling and average pooling operations; FC represents the fully connected layer; represents the feature extraction through the TCN network; represents the vehicle dynamics parameters; represents the high-dimensional risk feature vector.

[0010] Preferably, the training process of the conditional generative adversarial network module includes: The generator takes the high-dimensional risk feature vector as a condition and generates a simulated lane-changing trajectory according to the low-dimensional latent vector; The discriminator calculates the adversarial loss based on the distribution difference between the real trajectory dataset and the generated trajectory to constrain the update of the discriminator parameters until the diversity and authenticity of the generated trajectory meet the preset threshold.

[0011] Preferably, before the multi-modal dynamic time series data is input into the risk level feature extraction network, it is pre-normalized through the following formula: ; In the formula, Input represents the input data;Input min Represents the minimum value of the input data; Input max Represents the maximum value of the input data.

[0012] Preferably, the risk level feature extraction network classifies based on the collision time, and divides the lane - changing risk into three levels: ; In the formula, Represents the time before the collision occurs.

[0013] Preferably, the risk level feature extraction network is trained through a cross - entropy loss function.

[0014] Preferably, the expression of the cross - entropy loss function is: ; In the formula, C represents the number of categories; Represents the probability distribution of the i th true label; Represents the predicted probability of the i th category.

[0015] The beneficial effects of the present invention at least include: The present invention extracts high - dimensional risk feature vectors from multi - modal dynamic time - series data through a risk level feature extraction network, and then reduces the dimension of the high - dimensional features through an auto - encoder module to generate a low - dimensional latent vector representation; The conditional generative adversarial network conditions on the high - dimensional risk feature vectors, and the generator and discriminator will optimize the low - dimensional feature representation of the auto - encoder during the training process; The generator will try to generate trajectories that conform to the real data distribution, while the discriminator will judge the authenticity of the generated trajectories. It provides efficient and reliable test data support for the autonomous driving algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Is a schematic flow chart of an embodiment of the present invention; Figure 2 Is a comparison diagram of the four risk lane - changing trajectory reconstruction signals and the original signals of the auto - encoder in an embodiment of the present invention; Figure 3 Is a schematic diagram of the convergence effect of the auto - encoder in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] The following clearly and completely describes the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the protection scope of the present invention.

[0018] As Figure 1 shown, the embodiment of the present invention provides a lane-changing trajectory generation method, including the following steps: Step S1: Process the multi-modal dynamic time series data through a risk level feature extraction network to generate a high-dimensional risk feature vector representing the lane-changing risk level.

[0019] The multi-modal dynamic time series data includes vehicle lateral acceleration, longitudinal speed, steering angle change rate, and the minimum distance from surrounding vehicles.

[0020] Specifically, the risk level feature extraction network first extracts the multi-modal dynamic time series data during the vehicle's driving process to form a dynamic parameter matrix. Then, these dynamic parameters are associated with the lane-changing risk to form a risk feature vector. Finally, the risk feature vector is input into a neural network model, and after multiple convolutional and pooling operations, a high-dimensional feature vector is obtained. This feature vector can represent the risk degree of the vehicle during lane-changing, thereby helping us better evaluate the safety of lane-changing.

[0021] In this embodiment, the risk level feature extraction network in the entire large network architecture is mainly responsible for inputting multi-modal dynamic time series data, including speed, acceleration, the minimum distance from surrounding vehicles, the minimum collision time with surrounding vehicles, etc., a total of 7×80-dimensional data. Through the processing of the TCN network, a vector that can comprehensively represent the risk characteristics is obtained, and the dimension of the vector does not change, and the effectiveness of the comprehensive risk feature extraction is verified by performing fault classification on this vector.

[0022] The main structure of this part of the network is composed of the TCN network. In order to improve the comprehensive extraction ability of the TCN network for the risks of multi-modal input vectors, in this embodiment, the network is designed as a classification network. By classifying the risk level types of the extracted comprehensive risk feature vectors, the effectiveness of the feature extraction is verified. The following explains some structures and processes of the risk level feature extraction network: 1) Input layer The input signal is , and its shape is 7×80, where 80 is the time step and 7 is the feature dimension.

[0023] 2) TCN layer Each TCN layer includes the following components: Dilated Causal Convolution: Given an input \(x\), an output \(y\), a convolution kernel size \(k\), and a dilation factor \(d\), the output \(y\) can be expressed as: ; where are the weights of the convolution kernel.

[0024] Weight Normalization: Weight normalization normalizes the convolution kernel to keep the norm of the weight matrix within a certain range. The specific formula is as follows: ; where \(W\) is the original weight matrix, \(g\) is the scaling factor, is the normalized weight matrix.

[0025] ReLU Activation: .

[0026] Dropout layer: Dropout is a regularization technique that randomly discards the outputs of some neurons to prevent overfitting. Assuming a dropout rate of \(p\), the output can be expressed as: .

[0027] 3) Final classification layer The final feature vector is classified through a fully connected layer: ; where, and are the weight matrix and bias term of the fully connected layer, respectively.

[0028] In the embodiments of the present invention, before the multi-modal dynamic time series data is input into the risk level feature extraction network, it is pre-normalized through the following formula: ; In the formula, Input represents the input data; Input min represents the minimum value of the input data; Input max represents the maximum value of the input data.

[0029] In this embodiment, the risk level feature extraction network classifies based on the collision time and divides the lane-changing risk into three levels: ; In the formula, represents the time before the collision occurs.

[0030] During the training process, cross-entropy is used as the loss function, and its expression is: ; In the formula, C represents the number of categories; represents the i th probability distribution of the true label; represents the i th predicted probability of the category.

[0031] During the operation process, after the multi-modal dynamic time series data is input into the network, the feature extraction process of first dimension elevation and then dimension reduction is completed through several layers of TCN networks, and finally a high-dimensional risk feature vector is obtained. These features are converted into probabilities of three lane-changing risk levels through a linear layer. To prevent overfitting, a Dropout layer with a dropout rate of 0.2 is adopted in the network. The network is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and a total of 1000 training rounds.

[0032] Step S2: Input the high-dimensional risk feature vector into the autoencoder module for dimension reduction processing, and generate a low-dimensional latent vector representation through the encoder.

[0033] Specifically, in this embodiment, the autoencoder network compresses and then decompresses the training data to ensure that the finally decompressed data is highly close to the original data. The reason for our choice of the autoencoder is that the original data set contains a large amount of noise and mutations. Through the training of the autoencoder based on the gated recurrent unit (GRU), the lane-changing data significantly improves its smoothness while retaining its inherent attributes.

[0034] The autoencoder mainly consists of two parts: an encoder and a decoder. The encoder compresses the time series into a latent vector to capture the essential features of the input; the decoder reconstructs the original time series from the latent vector.

[0035] The core component of the autoencoder is the gated recurrent unit (GRU). Compared with the long short-term memory unit (LSTM), the GRU only contains an update gate and a reset gate, with fewer parameters, but is equally good at handling long-term dependencies in time series.

[0036] When selecting vehicle dynamics parameters, such as speed and acceleration, as inputs, first use a three-layer TCN network to extract features, and then concatenate these features with the high-dimensional risk feature vector obtained from the risk feature extraction network as the input to the GRU network: ; In the formula, represents the vehicle dynamics parameters, Represents a high-dimensional risk feature vector.

[0037] 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: ; ; ; In the formula, 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.

[0038] The decoding process is the inverse process of encoding, and the original time series is reconstructed through the fully connected layer and GRU network.

[0039] In summary, in order to generate a lane-changing trajectory signal that is as consistent as possible with the characteristics of the original signal, 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 respectively 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.

[0040] 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.

[0041] 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, and the high-dimensional risk feature vector is used as a condition to generate a lane change trajectory that meets the specified risk level.

[0042] In the embodiments of the present invention, the risk level feature extraction network first extracts a risk feature vector from the dynamic parameters of vehicle driving. This vector is high-dimensional and contains complex information about lane-changing risks. The autoencoder network further performs dimensionality reduction on these high-dimensional features, compressing the high-dimensional features into a low-dimensional vector representation through the encoder. This process not only reduces the dimensionality of the data but also removes redundant information and noise, retaining the key risk features.

[0043] The decoder part of the autoencoder reconstructs the low-dimensional vector into the original data form. This reconstruction process can help optimize the feature representation of the risk level feature extraction network. Through the reconstruction error feedback of the autoencoder, the risk level feature extraction network can further adjust its parameters to make the extracted risk features more accurate and robust.

[0044] In the conditional generative adversarial network (cGAN), the risk feature vector generated by the risk level feature extraction network is used as a conditional variable and input into the generator and discriminator. This conditional variable guides the generator to generate lane-changing trajectories that conform to specific risk features, making the generated trajectories more in line with the risk distribution in the actual driving scenario.

[0045] The discriminator of the cGAN discriminates the generated lane-changing trajectories to determine whether they conform to the distribution of real data. The feedback information of the discriminator can be backpropagated to the risk level feature extraction network to help it further optimize the risk feature extraction process. Through this feedback mechanism, the risk level feature extraction network can continuously adjust its parameters to make the extracted risk features more accurate, thereby improving the generation quality of the generator.

[0046] The low-dimensional feature vector generated by the autoencoder network is used as the input of the cGAN and further used to generate lane-changing trajectories. Since the autoencoder has already performed dimensionality reduction and denoising on the data, the generator can focus more on the key features when generating trajectories, reducing the interference of noise, and thus generating more realistic and reasonable lane-changing trajectories.

[0047] The generator and discriminator of the cGAN will optimize the low-dimensional feature representation of the autoencoder during the training process. The generator will try to generate trajectories that conform to the distribution of real data, while the discriminator will judge the authenticity of the generated trajectories. This process will prompt the autoencoder to generate more discriminative low-dimensional features, thereby improving the feature extraction ability of the autoencoder.

[0048] Overall, the risk level feature extraction network, the autoencoder network, and cGAN form a closed-loop system. The high-dimensional risk features extracted by the risk level feature extraction network are input into cGAN to generate lane change trajectories after being dimensionally reduced by the autoencoder. The authenticity of the generated trajectories is fed back to the risk level feature extraction network and the autoencoder by the discriminator to further optimize the feature extraction and dimensional reduction processes. The risk level feature extraction network is responsible for extracting risk features from the original data, the autoencoder is responsible for dimensional reduction and denoising of these features, and cGAN is responsible for generating lane change trajectories that conform to the risk features. Through continuous feedback and optimization, the overall performance of the system is gradually improved.

[0049] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. Only the preferred embodiments of the present invention are expressed. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present invention. As long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0050] It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to 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 the vehicle dynamics parameter features, concatenating them with the high-dimensional risk feature vector, and inputting them into the autoencoder module for dimensionality reduction processing 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, input into a conditional generative adversarial network module, and the high-dimensional risk feature vector is used as a condition to generate a lane change trajectory that meets the specified risk level.

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 the neuron output at a preset dropout rate.

4. The lane change trajectory generation method according to claim 1, characterized in that: The encoder module includes a temporal convolutional layer and a gated recurrent unit GRU, and an output end is connected to an average pooling layer and a global pooling layer.

5. The lane change trajectory generation method according to claim 4, characterized in that: The autoencoder module generates a low-dimensional latent vector through the following formula: ； ； ； In the formula, 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.

6. 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 low-dimensional latent vectors, conditioned on high-dimensional risk feature vectors; 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.

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

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

9. The lane change trajectory generation method according to claim 8, characterized in that: The risk level feature extraction network is trained using a cross entropy loss function.

10. The lane change trajectory generation method according to claim 9, characterized in that: The expression of the cross entropy loss function is: ； In the formula, C represents the number of categories; Indicates i The probability distribution of the true labels; Indicates i The predicted probability of each class.

Images