An unknown domain trajectory prediction method combined with a self-supervised mechanism

Abstract

The application relates to a kind of unknown domain trajectory prediction methods combined with self-supervision mechanism, the method comprises the following steps: (1) obtaining target trajectory data and adjacent target trajectory data, scene map data;(2) the maneuver type of trajectory is extracted by clustering algorithm;(3) adjacent target trajectory and scene map information are extracted, (4) the predicted maneuver type label is obtained;(5) after obtaining the predicted maneuver type, the trajectory maneuver type self-supervision subtask loss is calculated, (6), after obtaining the end point, the trajectory end point and data coding cascade input time series neural network decoding obtain the trajectory of the remaining points, (7) model loss calculation, complete vehicle future trajectory prediction.The scheme uses the trajectory maneuver type obtained by clustering algorithm to solve the data domain division problem in the application process of invariance principle, improve the application effect of invariance principle, improve the accuracy of model for unknown domain trajectory prediction.

Description

Technical Field

[0001] This invention belongs to the field of trajectory prediction technology, specifically relating to a method for predicting the trajectory of a target in an unknown domain that incorporates a self-supervised mechanism. Background Technology

[0002] With the rapid development of autonomous driving technology, trajectory prediction has become an indispensable key component of intelligent vehicle systems. Deep learning algorithms have consistently demonstrated outstanding performance in the field of trajectory prediction, but traditional deep learning algorithms generally suffer from insufficient model generalization ability, leading to a significant performance drop when testing unknown domain data.

[0003] Most existing deep learning trajectory prediction models sample the training and testing domains from an approximate domain. Examples include trajectory prediction methods based on BEV bird's-eye view (Publication No.: CN114708723A), trajectory prediction methods fusing multimodal information (Publication No.: CN114648551B), and trajectory prediction methods based on LSTM multi-level motion features (Publication No.: CN115797892A). These methods, which rely on deep learning models to solve trajectory prediction problems, all ignore the generalization problem caused by differences in data distribution. When these models predict trajectories in unknown domains, their prediction accuracy decreases, posing significant safety risks. Currently, some methods are specifically designed to improve the generalization ability of models. Chinese patent CN115797892A proposes a method to improve the generalization ability of vehicle trajectory models by combining an endpoint prediction module based on the invariance principle. However, this method does not consider the data domain partitioning problem during the application of the invariance principle. Invention patent CN116912661A uses pre-training to improve the model's accuracy in unknown domains. However, pre-training methods, accompanied by distributed training, increase the difficulty of training and consume more training resources. Therefore, this invention proposes a target trajectory prediction method in the unknown domain that combines a self-supervised mechanism. Summary of the Invention

[0004] Purpose of the invention: In order to solve the above-mentioned problems, improve the prediction effect of the target trajectory model in the unknown domain, and enhance the generalization ability of the model, this invention discloses an unknown domain trajectory prediction method combined with a self-supervised mechanism. For the data domain partitioning problem in the application of invariant risk minimization, an environment partitioning mechanism integrating self-supervised subtasks is proposed.

[0005] Technical solution: A method for predicting trajectories in unknown domains by combining self-supervised mechanisms, comprising the following steps:

[0006] (1) Acquire target trajectory data, nearby target trajectory data, and scene map data;

[0007] (2) In order to better divide the trajectory data into domains, the maneuver type of the trajectory is extracted by clustering algorithm;

[0008] (3) Extract the trajectory and scene map information of nearby targets. Input the predicted target's historical trajectory, the target's historical trajectory, and the scene map information into a temporal neural network encoder to extract features. Perform convolutional pooling on the map features and nearby vehicle features to extract high-dimensional features;

[0009] (4) Input the historical trajectory features of the neighboring target and the historical trajectory features of the target to the maneuver type prediction module to obtain the predicted maneuver type label;

[0010] (5) After obtaining the predicted maneuver type, use the maneuver type generated in step (2) to jointly calculate the trajectory maneuver type self-supervised sub-task loss.

[0011] (6) The trajectory endpoint is generated by using the data encoding input based on the invariance principle to obtain the predicted endpoint. After obtaining the endpoint, the trajectory endpoint and the data encoding are concatenated and input into the temporal neural network for decoding to obtain the trajectory of the remaining points.

[0012] (7) Model loss calculation: First, the endpoints of the future trajectory are generated through the endpoint module. After generation, the endpoint encoding and feature encoding are concatenated, and then the remaining trajectory points are completed to complete the prediction of the vehicle's future trajectory. The model's loss function consists of five parts: the self-supervised subtask loss in step (5), the invariance principle loss in step (6), the endpoint generation bias loss, the remaining trajectory bias loss, and the Kl divergence loss.

[0013] Furthermore, the target trajectory data obtained in step (1) includes the trajectory coordinate point set of the predicted target, the trajectory coordinate point set of neighboring targets, the center line of the lane in the scene map where the target is located, and the topological relationship between lanes.

[0014] Furthermore, step (2) also includes the main steps of historical trajectory clustering: first, normalize and rotate the vehicle trajectory coordinates, then extract trajectory feature data, and finally set the relevant parameters of the clustering algorithm, perform clustering operations, and obtain the label of trajectory movement type.

[0015] Furthermore, step (2) vehicle trajectory coordinate normalization and rotation also includes: in order to more centralize the coordinate mapping of the motor type, normalization and coordinate rotation operations are first performed on the vehicle trajectory. Assume the historical trajectory of the m-th predicted vehicle in the n-th data domain is... Future trajectory is The vehicle trajectory normalization operation involves translating the trajectory until the endpoint of the historical trajectory is taken as the origin. The coordinate calculation formula is as follows:

[0016]

[0017] Assume the trajectory vector v from the start to the end of the vehicle's historical trajectory, with coordinates (v... x ,v y The coordinate rotation operation for the vehicle trajectory involves rotating the trajectory vector v until it is parallel to the y-axis. The coordinate calculation process is as follows, assuming the coordinate rotation matrix is ​​R. and For the rotated historical trajectory and future trajectory:

[0018]

[0019]

[0020]

[0021]

[0022] Furthermore, step (2) of extracting trajectory feature data also includes: the present invention classifies the maneuver type into ["left turn", "straight", "right turn"] and ["accelerate", "maintain speed", "decelerate"], and therefore extracts two types of trajectory features. For the trajectory direction of travel, the lateral displacement of the trajectory is extracted as the clustering feature, and for the trajectory speed change, the speed magnitude at that moment is extracted as the clustering feature.

[0023] Furthermore, the relevant parameters of the clustering algorithm in step (2) are: the number of cluster categories is 3, the maximum number of clustering iterations is 100, and the clustering convergence tolerance is 0.0001.

[0024] Further, step (2) of the clustering operation includes: first, acquiring the feature data extracted from the trajectory; randomly selecting three data points as initial cluster centers; for each data point in the feature data, calculating the distance from each data point to the cluster center; assigning the data point to the nearest cluster center and updating its label; recalculating the center of each category based on the latest clustering assignment result; and outputting the category label for each trajectory if the moving clustering of the cluster center is less than the convergence tolerance.

[0025] Furthermore, the input to the motor type prediction module in step (4) is the trajectory endpoint code, map code, historical trajectory code, and nearby vehicle trajectory code, and the output is the predicted travel direction category and trajectory speed category. The motor type predictor consists of two predictors: one is the travel category predictor, which predicts whether the trajectory will turn left, go straight, or turn right; the other is the trajectory speed predictor, which predicts whether the vehicle will accelerate, maintain speed, or decelerate.

[0026] Furthermore, in step (5), the self-supervised maneuver type prediction sub-task loss introduced in this invention is assumed to be... For the prior movement category and maneuver type, Indicates the trajectory speed maneuver type. Let e ​​be the predicted maneuver type label, then the loss calculation formula is as follows, where e represents the scene number:

[0027]

[0028]

[0029]

[0030] in Indicates the loss of self-monitoring mission due to maneuver type in direction of travel. This indicates the loss of self-monitored task due to speed change maneuver type. Losses due to self-monitoring subtasks.

[0031] Furthermore, in step (6), the endpoint prediction module based on the invariance principle consists of an encoder, a decoder, and a latent variable Z, and the loss function that minimizes invariant risk is applied to the CVAE module. If the invariant features in historical encoding and map encoding are set as F... s ={H s C s Let the characteristic of the change be F. r ={H r C r}, then encoder E cvae By applying the KL divergence loss, the fitted q θ (Z|y t+α ,F s ,F r The distribution is close to P. θ (Z|y t+α ,F s ,F r The distribution of ) and decoder D cvae Learning P during training θ (y t+α |Z,F s ,F r The distribution of ) . Minimizing loss with invariant risk prompts the encoder of the endpoint prediction module to eventually learn q θ (Z|y t+α ,F s ,F r The distribution will weaken the relationship with F. r The correlation, the fit is only based on F s and y t+αLet Z be the distribution of the latent variable Z under the condition. The loss function for minimizing invariant risk is calculated based on the mobility type of the clustering results, and its formula is as follows:

[0032]

[0033]

[0034] Where e represents the scene number, M e It is a mobile type. T is the endpoint of the trajectory. e H e C e This includes historical trajectory information, nearby vehicle information, and map scene information for the scene; φ represents the destination prediction module; and f... w For a linear classifier, w e R is a classifier specific to this scenario. e It is the product of the deviation between the predicted value and the true value of the endpoint under environment e and the one-hot encoding of the data field.

[0035] Furthermore, in step (7), the model's loss function is as follows, where n represents the number of data domains:

[0036]

[0037] in Generate deviation loss for the endpoint. For KL divergence loss, For the remaining trajectory deviation loss, Loss due to the invariance principle, The loss is for the self-supervised subtask. The correlation coefficient is set as γ = δ = η = ρ = 1, and λ is a hyperparameter for adjusting the weights of each loss during training.

[0038] Beneficial Effects: The unknown domain trajectory prediction method provided by this invention, which combines a self-supervised mechanism, has the following advantages: 1. It proposes an endpoint prediction module based on the invariance principle, and uses the trajectory maneuver type obtained by clustering algorithm to solve the data domain partitioning problem in the application of the invariance principle, thereby improving the application effect of the invariance principle and improving the accuracy of the model in predicting unknown domain trajectories; 2. By designing self-supervised sub-tasks related to maneuver types, it enables the model to improve its understanding of maneuver types, improves the model's efficiency in utilizing limited data, and further enhances the model's domain generalization ability. Attached Figure Description

[0039] Figure 1 This is a flowchart of the unknown domain trajectory prediction method combining a self-supervised mechanism according to the present invention.

[0040] Figure 2 This is a schematic diagram of the maneuver type prediction module of the present invention.

[0041] Figure 3 This is a schematic diagram of the endpoint prediction module of the present invention. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0043] This method can be used in fields such as autonomous driving and driver assistance. Example 1 describes a case where this method is applied to the field of autonomous driving.

[0044] Example 1: Refer to Figures 1 to 3 This embodiment describes a trajectory prediction method for unknown domains that incorporates a self-supervised mechanism, and includes the following steps:

[0045] S1: Preprocessing of training data.

[0046] In the training data preprocessing stage, it is necessary to extract the predicted vehicle trajectory for each interaction scenario in the dataset, as well as the target trajectory data of neighboring vehicles and the lane centerline information of the scene map.

[0047] S2: Extract trajectory maneuver type

[0048] To better divide the trajectory data into domains, a clustering algorithm is used to extract the maneuver type of the trajectory.

[0049] S21: Vehicle coordinate normalization and rotation

[0050] Normalization and coordinate rotation operations are performed on the target's trajectory. Assume the historical trajectory of the m-th predicted vehicle within the n-th data domain is... Future trajectory is The vehicle trajectory normalization operation involves translating the trajectory until the endpoint of the historical trajectory is taken as the origin. The coordinate calculation formula is as follows:

[0051]

[0052] Assume the trajectory vector v from the start to the end of the vehicle's historical trajectory, with coordinates (v... x ,v y The coordinate rotation operation for the vehicle trajectory involves rotating the trajectory vector v until it is parallel to the y-axis. The coordinate calculation process is as follows, assuming the coordinate rotation matrix is ​​R. and For the rotated historical trajectory and future trajectory:

[0053]

[0054]

[0055]

[0056]

[0057] S22: Extracting trajectory features

[0058] The vehicle trajectory movement types are divided into ["left turn", "straight", "right turn"] and ["accelerate", "maintain speed", "decelerate"]. Therefore, two types of trajectory features are extracted. For the vehicle trajectory direction clustering, the lateral displacement of the trajectory is extracted as the clustering feature, and for the trajectory speed change, the speed magnitude at that moment is extracted as the clustering feature.

[0059] S23: Trajectory Clustering

[0060] The relevant parameters of the clustering algorithm are: the number of clusters is 3, the maximum number of clustering iterations is 100, and the clustering convergence tolerance is 0.0001.

[0061] First, relevant features are extracted from the trajectory data. Then, K data points are randomly selected as initial cluster centers C, where K represents the number of clusters. The iteration count is initialized to 0. During a single iteration, for each data point representing a trajectory feature, its distance to each cluster center is calculated, and the data point is assigned to the nearest cluster center, updating the cluster label. Then, based on the new cluster assignments, each cluster center is recalculated, and the cluster centers and iteration count are updated. If the movement distance of a cluster center is less than the tolerance threshold, the iteration ends. Finally, the program outputs the cluster centers and the category label for each trajectory.

[0062] S3: Feature extraction encoding

[0063] The system extracts the trajectories of nearby targets and scene map information. The predicted target's historical trajectory, the target's historical trajectory, and scene map information are input into a temporal neural network encoder to extract features. High-dimensional features are then extracted using convolutional pooling on the map features and the features of nearby vehicles.

[0064] S31: Extract map topology feature encoding through social convolutional pooling structure. The implementation method consists of two parts. First, before training, a depth-first search algorithm is used to extract relative position matrix information from the map data. Then, during training, road social vectors are established, and LSTM encodings for lane centerlines are filled in according to the relative matrix information. Convolution operations are performed on the road social vectors to finally obtain map information features.

[0065] S4: Maneuver Type Prediction

[0066] The historical trajectory features of nearby targets and the historical trajectory features of the target to be predicted are input into the maneuver type prediction module to obtain the predicted maneuver type label.

[0067] S41: The structure of the maneuver type prediction module takes as input the trajectory endpoint code, map code, historical trajectory code, and neighboring vehicle trajectory code, and outputs the predicted travel direction category and trajectory speed category. The maneuver type predictor consists of two predictors, such as... Figure 2 As shown, one is a travel category predictor that predicts whether the trajectory will turn left, go straight, or turn right; the other is a trajectory speed predictor that predicts whether the vehicle will accelerate, maintain speed, or decelerate.

[0068] S5: Mobility Type Loss

[0069] The self-supervised maneuver type prediction sub-task loss introduced in this invention assumes... For the prior movement category and maneuver type, Indicates the trajectory speed maneuver type. If the predicted maneuver type is denoted as a label, the loss calculation formula is as follows: Indicates the loss of self-monitoring mission due to maneuver type in direction of travel. This indicates the loss of self-monitored task due to speed change maneuver type. This represents the loss for the self-supervised subtask. Where 'e' represents the scenario number.

[0070]

[0071]

[0072]

[0073] S6: Generate the trajectory endpoint and other trajectory points

[0074] The trajectory endpoint encoding is concatenated with map features and nearby vehicle features, and the trajectory endpoint is generated through a conditional autoencoder structure.

[0075] S61: The structure of a conditional autoencoder is as follows Figure 3 As shown, the endpoint prediction module based on the invariance principle consists of an encoder, a decoder, and a latent variable Z. A loss function minimizing invariant risk is applied to the CVAE module. If the invariant features in historical encoding and map encoding are set as F... s ={H s C s Let the characteristic of the change be F. r ={H r C r}, then encoder E cvae By applying the KL divergence loss, the fitted qθ (Z|y t+α ,F s ,F r The distribution is close to P. θ (Z|y t+α ,F s ,F r The distribution of ) and decoder D cvae Learning P during training θ (y t+α |Z,F s ,F r The distribution of ) . Minimizing loss with invariant risk prompts the encoder of the endpoint prediction module to eventually learn q θ (Z|y t+α ,F s ,F r The distribution will weaken the relationship with F. r The correlation, the fit is only based on F s and y t+α Let Z be the distribution of the latent variable Z under the condition. The loss function for minimizing invariant risk is calculated based on the mobility type of the clustering results, and its formula is as follows:

[0076]

[0077]

[0078] Where e represents the scene number, M e It is a mobile type. T is the endpoint of the trajectory. e H e C e This includes historical trajectory information, nearby vehicle information, and map scene information for the scene; φ represents the destination prediction module; and f... w For a linear classifier, w e R is a classifier specific to this scenario. e It is the product of the deviation between the predicted value and the true value of the endpoint under environment e and the one-hot encoding of the data field.

[0079] S62: After obtaining the endpoint of the generated trajectory, it is concatenated with the existing data encoding: map encoding, historical trajectory encoding, and neighboring vehicle feature encoding, and input into the LSTM time-series decoder to supplement the remaining vehicle trajectories.

[0080] S7: Model loss calculation. The model loss function is as follows, where n represents the number of data domains:

[0081]

[0082] in Generate deviation loss for the endpoint. For KL divergence loss, For the remaining trajectory deviation loss, Loss due to the invariance principle, The loss is set to the self-supervised subtask. The correlation coefficient is set to γ ​​= δ = η = ρ = 1, and λ is a hyperparameter used to adjust the weights of each loss during training. During the model usage and testing phases, historical trajectory data is input into the model. First, the distribution of the endpoint prediction module is sampled to generate trajectory points, and then the remaining trajectory points are completed to finish predicting the future trajectory of the vehicle.

[0083] To verify the correctness and rationality of the above implementation method, tests were conducted using two road conditions: roundabouts and intersections, from the publicly available dataset INTERACTON. The test data was not visible to the model during training. This implementation method segments the trajectory into 5-second segments, including a 2-second historical trajectory and a 3-second future trajectory. To smooth the real trajectory and reduce the number of model parameters, the trajectory was downsampled with a downsampling coefficient of 2, meaning the step size for the historical trajectory was 10 and the step size for the future trajectory was 15.

[0084] Regarding the settings of the experimental network parameters, the initial learning rate for the experiments in this chapter was set to 0.001. The learning rate decreased with each training epoch, decreasing by 20% every two epochs. The Adam optimizer was used, and all experiments were trained for a uniform 25 epochs.

[0085] The evaluation metric uses the minimum mean displacement error (mDE) to measure the generalization ability of the vehicle trajectory model. mADE is defined as the average L2 norm of the predicted trajectory with the smallest interpolation between the generated k endpoints and the ground truth trajectory points, as shown in the formula below:

[0086]

[0087] The test results are shown in Tables 1 and 2:

[0088] Table 1 Test Results for the Roundabout Scene

[0089] time ERM MMD DANN IRM Rex TNT SSL-lane This invention 1s 031 0.17 0.30 0.23 0.21 0.47 125 018 2s 064 0.49 066 0.60 0.62 166 2.68 0.34 3s 136 116 139 129 135 2.35 4.91 0.53

[0090] Table 2 Test Results for Intersection Scenarios

[0091] time ERM MMD DANN IRM Rex TNT SSL-lane This invention 1s 029 018 027 031 020 044 135 012 2s 0.69 0.40 055 0.62 048 169 279 023 3s 123 0.81 130 123 107 228 501 035

[0092] The comparison methods are described in Table 3:

[0093] Table 3. Comparison Methods

[0094]

[0095] Based on the experimental results, this invention can play a role in improving the generalization ability of trajectory prediction models and explores an optimization approach for minimizing the invariant risk of trajectory prediction methods in unknown domains.

[0096] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for predicting trajectories in an unknown domain by incorporating a self-supervised mechanism, characterized in that, The method includes the following steps: (1) Acquire target trajectory data, adjacent target trajectory data, and scene map data; (2) Extract the maneuver type of the trajectory using a clustering algorithm; (3) Extract the trajectory and scene map information of the nearby target. Input the predicted target historical trajectory, target historical trajectory and scene map information into the temporal neural network encoder to extract features. Perform convolutional pooling to extract high-dimensional features for map features and nearby vehicle features respectively. (4) Input the historical trajectory features of the neighboring target and the historical trajectory features of the target to the maneuver type prediction module to obtain the predicted maneuver type label; (5) After obtaining the predicted maneuver type, use the maneuver types generated in step (2) to jointly calculate the trajectory maneuver type self-supervised sub-task loss. (6) The trajectory endpoint is generated by using a conditional autoencoder based on the invariance principle to obtain the predicted endpoint. After obtaining the endpoint, the trajectory endpoint and the data encoder are concatenated and input into a temporal neural network for decoding to obtain the trajectory of the remaining points. (7) Model loss calculation: First, the endpoints of the future trajectory are generated through the endpoint module. After generation, the endpoint encoding and feature encoding are concatenated, and the remaining trajectory points are filled in to complete the prediction of the vehicle's future trajectory. The model's loss function consists of five parts: the self-supervised subtask loss in step (5), the invariance principle loss in step (6), the endpoint generation bias loss, the remaining trajectory bias loss, and the kl divergence loss. In step (5), the self-supervised maneuver type prediction sub-task loss is assumed to be... For the prior movement category and maneuver type, Indicates the trajectory, speed, and maneuver type. , Let the predicted maneuver type labels be denoted as follows, then the loss calculation formula is as follows, where Indicates scene number: in Indicates the loss of self-monitoring mission due to maneuver type in direction of travel. Indicates the loss of self-monitored mission due to speed change maneuver type. Losses due to self-supervised subtasks; In step (6), the endpoint prediction module based on the invariance principle consists of an encoder, a decoder, and a latent variable Z. The loss function that minimizes invariant risk is applied to the CVAE module. If the invariant features in the historical encoding and map encoding are set as The characteristics of the change are set as Then the encoder By applying the KL divergence loss, the fitted... Distribution close Distribution of decoders Learning during training The distribution of invariant risk minimization loss prompts the encoder of the endpoint prediction module to eventually learn... Distribution will weaken and The correlation, the fit is only based on and implicit variables for conditions The distribution of invariant risk is calculated using the loss function that minimizes invariant risk based on the mobility type of the clustering results, and its formula is as follows: in Indicates the scene number. It is a mobile type. The endpoint of the trajectory, This includes historical trajectory information, nearby vehicle information, and map scene information. For the endpoint prediction module, For linear classifiers, A scene-specific classifier It is the environment The product of the deviation between the predicted value and the true value of the next endpoint and the one-hot encoding of the data field.

2. The unknown domain trajectory prediction method combined with a self-supervised mechanism according to claim 1, characterized in that, The target trajectory data obtained in step (1) includes the set of trajectory coordinate points of the predicted target, the set of trajectory coordinate points of neighboring targets, the center line of the lane in the scene map where the target is located, and the topological relationship between lanes.

3. The unknown domain trajectory prediction method combined with a self-supervised mechanism according to claim 1, characterized in that, Step (2) also includes a step of clustering historical trajectories: first, the vehicle trajectory coordinates are normalized and rotated; then, trajectory feature data is extracted; finally, the relevant parameters of the clustering algorithm are set, clustering operations are performed, and labels for trajectory movement types are obtained. Step (2) Vehicle trajectory coordinate normalization and rotation also includes: In order to make the coordinate mapping of the motor type more concentrated, the vehicle trajectory is first normalized and rotated. Assume that the historical trajectory of the m-th predicted vehicle in the n-th data domain is The future trajectory is The vehicle trajectory normalization operation is a translation of the trajectory until the endpoint of the historical trajectory is taken as the origin. The coordinate calculation formula is as follows: Assume the historical trajectory of a vehicle, from its starting point to its ending point, is a trajectory vector. Its coordinates are The coordinate rotation operation for the vehicle trajectory involves rotating the trajectory vector v until it is parallel to the y-axis. The coordinate calculation process is as follows, assuming the coordinate rotation matrix is ​​R. and For the rotated historical trajectory and future trajectory: The step (2) of extracting trajectory feature data also includes: dividing the maneuver type into ["left turn", "straight", "right turn"] and ["accelerate", "maintain speed", "decelerate"], extracting two types of trajectory features, clustering the trajectory direction to extract the lateral displacement of the trajectory as the clustering feature, and extracting the speed magnitude at any moment as the clustering feature based on the change in trajectory speed.

4. The unknown domain trajectory prediction method combining a self-supervised mechanism according to claim 1, characterized in that, The relevant parameters of the clustering algorithm in step (2) are: the number of cluster categories is 3, the maximum number of clustering iterations is 100, and the clustering convergence tolerance is 0.0001.

5. The unknown domain trajectory prediction method combined with a self-supervised mechanism according to claim 1, characterized in that, The clustering operation in step (2) includes: first, obtaining the feature data of the trajectory extraction, randomly selecting 3 data points as the initial cluster centers, calculating the distance from each data point to the cluster center for each data point of the feature data, assigning the data point to the nearest cluster center, updating its label, recalculating the center of each category based on the latest clustering assignment result, and outputting the category label of each trajectory if the moving cluster of the cluster center is less than the convergence tolerance.

6. The unknown domain trajectory prediction method combining self-supervised mechanism according to claim 1, characterized in that, The input of the motor type prediction module in step (4) is the trajectory endpoint code, map code, historical trajectory code and adjacent vehicle trajectory code. The output is the predicted travel direction category and trajectory speed category. The motor type predictor consists of two predictors: one is the travel category predictor, which predicts whether the trajectory will turn left, go straight or turn right; the other is the trajectory speed predictor, which predicts whether the vehicle will accelerate, maintain speed or decelerate.

7. The unknown domain trajectory prediction method combining self-supervised mechanism according to claim 1, characterized in that, In step (7), the loss function of the model is as follows, where n represents the number of data domains: in Generate deviation loss for the endpoint. For KL divergence loss, For the remaining trajectory deviation loss, Loss due to the invariance principle, The correlation coefficient is set to account for the loss of self-supervised subtasks. , The hyperparameters are used to adjust the various loss weights during training.

Images