A trajectory primitive-based multi-modal trajectory prediction system and method for traffic vehicles

Abstract

The application provides a traffic vehicle multi-modal trajectory prediction system and method based on trajectory primitives, the system comprising a similar trajectory screening module, a trajectory primitive library acquisition module, a primitive combination statistical module, a decoding prediction module, a trajectory prediction network training module, the method comprising establishing trajectory primitive libraries of various scenes, building and training trajectory prediction networks corresponding to primitive categories, and multi-modal trajectory prediction.The application first performs multi-modal trajectory prediction on traffic vehicles from the perspective of trajectory primitives, improves the explainability of the trajectory prediction model while ensuring multi-modal prediction results, and only requires the historical trajectory of the traffic vehicle, without requiring the historical trajectory of surrounding vehicles, thereby reducing the requirement of the trajectory prediction algorithm for the perception range of the ego vehicle.

Description

Technical Field

[0001] This invention belongs to the field of intelligent assisted driving for automobiles, and relates to a method for predicting the trajectory of a vehicle, specifically a multimodal trajectory prediction system and method for autonomous vehicles based on trajectory primitives. Background Technology

[0002] To operate safely and efficiently on the road, autonomous vehicles should not only understand the current state of nearby road users, but also proactively predict the expected behavior of other vehicles on the road. By identifying the intentions and behaviors of other vehicles and predicting their trajectories, the risk of traffic collisions can be effectively assessed and reduced, supporting effective decision-making and thus helping to lower the probability of traffic accidents, thereby further improving traffic safety.

[0003] From the perspective of output results, trajectory prediction can be divided into unimodal output and multimodal output. Unimodal output predicts only one trajectory, while multimodal output predicts multiple possible trajectories. Due to the uncertainty of the driver's driving intentions and operations, trajectory uncertainty is inevitable. Therefore, compared to unimodal output, the decision-making and planning module of autonomous driving requires more multimodal trajectory prediction results. However, existing multimodal trajectory prediction methods, whether using data-driven approaches or using intention-based methods, can generate multiple trajectories but lack interpretability. While these methods offer stronger interpretability, they are categorized according to the driver's operational intentions (e.g., lane changing and lane keeping on straight roads, and left / right turns and straight driving at intersections). Within a single intention category, the predicted trajectory result remains unimodal. In reality, even if the driver has a specific intention, multiple possible trajectory results will still occur.

[0004] Since a driver's driving operations consist of different driving behavior patterns, these different patterns generate different trajectory sequences. Considering a trajectory sequence generated under a single driving behavior pattern as a type of trajectory primitive, the vehicle's trajectory is essentially composed of segments of these trajectory primitives. However, current technology lacks a method for multimodal trajectory prediction of vehicles from the perspective of trajectory primitives. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multimodal trajectory prediction system for vehicles based on trajectory primitives. The system includes a similar trajectory filtering module, a trajectory primitive library acquisition module, a primitive combination statistics module, a decoding prediction module, and a trajectory prediction network training module. The input to the multimodal trajectory prediction system is the historical trajectory of the vehicle obtained by the autonomous vehicle's environmental perception module. This trajectory is compared with the trajectory in the trajectory primitive library acquired by the trajectory primitive library acquisition module. Similar trajectories are filtered, primitive combinations are statistically analyzed, and the primitive combination probability and primitive sequence length are obtained. Using the network structure trained by the trajectory prediction network training module, the decoding prediction module decodes and predicts each primitive category sequentially, achieving trajectory splicing and obtaining the predicted trajectory under each primitive combination. Finally, the system output is obtained, namely, multiple possible driving trajectories of the vehicle and the probability of each trajectory occurring. The output trajectories and probabilities serve as input to the autonomous vehicle's decision-making and planning module for decision-making and planning.

[0006] This invention provides a method for predicting the multimodal trajectory of a vehicle based on trajectory primitives, according to the above system, comprising the following steps:

[0007] (1) Establish trajectory primitive libraries for each scenario;

[0008] (1.1) Trajectory dataset acquisition and smoothing processing:

[0009] We selected open-source trajectory datasets for various scenarios, with the required features being the vehicle's horizontal and vertical coordinates. Due to issues with the trajectory data acquisition and processing methods, the vehicle's horizontal and vertical coordinates obtained from the datasets contain some noise compared to the actual vehicle coordinates. Therefore, we used a moving average method to smooth the trajectory data. The moving average formula is as follows:

[0010]

[0011] In the formula, p t x represents the filtering result at time t. t Let t represent the observed value at time t, and n represent the sliding window radius;

[0012] (1.2) Acquisition of trajectory lateral and longitudinal velocities and accelerations:

[0013] Take the smoothed vehicle lateral and longitudinal coordinates obtained in step (1.1), calculate the first and second derivatives with respect to time, and obtain the lateral and longitudinal velocities and accelerations of the trajectory.

[0014] (1.3) Trajectory sequence segmentation:

[0015] Adjacent samples with the same data performance trend are used as the data composition of the same sequence segment. The input is the lateral and longitudinal velocities and accelerations of the trajectory obtained in step (1.2). The lateral and longitudinal velocities and accelerations are divided into sequence segments by the Bayesian condensation sequence segmentation method to obtain the trajectory sequence segmentation result.

[0016] (1.4) Trajectory sequence clustering:

[0017] The segmented trajectory sequence will be presented in the form of trajectory sequence segments. Each sequence segment contains an indefinite number of trajectory samples, and trajectory primitives are contained in each sequence segment. The characteristics of different manipulation modes are represented by the sample features.

[0018] First, the trajectory sequence segments obtained in step (1.3) are clustered using a Gaussian mixture model. Then, the cluster labels obtained from the clustering are regarded as words in the Hidden Dirichlet Allocation model. They are used to replace the original driving samples as input data for Hidden Dirichlet Allocation to obtain the trajectory sequence segment categories. It is believed that trajectory sequence segments of the same category are generated by a driving mode or driving behavior, which is a type of trajectory primitive.

[0019] By combining the smoothed vehicle horizontal and vertical coordinates obtained in step (1.1) with their corresponding velocity and acceleration sequences through segmentation and clustering to obtain the primitive categories, the trajectory primitive library for each scene can be obtained.

[0020] (2) Construction and training of trajectory prediction networks corresponding to primitive categories;

[0021] (2.1) Filter data by primitive category:

[0022] The trajectory primitive library of each scene obtained in step (1) is filtered according to primitive category; specifically, the trajectory segment with the same primitive category and all trajectories before the trajectory segment are used as training data under the primitive category.

[0023] (2.2) Trajectory prediction network construction and training:

[0024] An encoder-decoder structure based on an LSTM network is adopted. The data filtered by primitive category in step (2.1) is used as training data, the trajectory segments with the same primitive category are used as predicted trajectories, and all trajectories before the trajectory segment are used as historical trajectories. The network weights are obtained through training, and the trajectory prediction neural network structure of the corresponding primitive category is obtained.

[0025] (3) Multimodal trajectory prediction:

[0026] (3.1) Similarity calculation:

[0027] After the autonomous vehicle perception system obtains a segment of the historical trajectory of a vehicle in a certain scenario, the historical trajectory of the vehicle and the trajectory primitive library of each scenario obtained in step (1) are used as input. The historical trajectory of the vehicle is compared with the trajectory data in the trajectory primitive library of each scenario, and the similarity between the historical trajectory to be predicted and the trajectory in the trajectory primitive library of the corresponding scenario is calculated.

[0028] Similarity is calculated using the distance between the multiline positions, as shown in the following formula:

[0029]

[0030]

[0031] Q and S are two trajectories, Area i Let w be the area of ​​the i-th region enclosed by the two trajectories. i As the weight, I i Let be the i-th intersection point of the two trajectories, and Length be the length of the trajectory between the two points.

[0032] (3.2) Similar trajectory filtering:

[0033] By setting a similarity threshold, those less than the threshold are discarded, and those greater than the threshold, i.e., those similar to the historical trajectory to be predicted, are selected. The primitive combination methods after the similar segments are counted, the frequency of primitive combination occurrence is calculated, the time length of a single primitive under a certain primitive combination is counted, and the average value is the predicted time length of a single primitive.

[0034] (3.3) Decode the prediction using the corresponding primitive prediction network:

[0035] The historical trajectory of the vehicle to be predicted, the various trajectory primitive combinations obtained in step (3.2), and the time series length of each primitive under the corresponding combination are input into the trajectory prediction neural network structure of the corresponding primitive category obtained in step (2.2). The primitive categories are decoded and predicted one by one to realize the splicing of trajectories and obtain the predicted trajectory under each primitive combination. The frequency of occurrence of the primitive combination obtained in step (3.2) is used as the confidence of the predicted trajectory under that combination, and finally the multimodal trajectory prediction result is obtained.

[0036] Furthermore, in step (1.3) trajectory sequence segmentation, for the time series of lateral and longitudinal velocities and accelerations obtained in step (1.2), a piecewise linear model is used to describe the relationship between input and output. The sequence is segmented according to its changing trend, and the time series function of each segment can be represented by a multiple linear regression model. For the parameter matrix of the multiple linear regression model, the conjugate distribution is used for parameter fitting. The posterior parameters of the conjugate distribution can be learned by continuously updating the distribution parameters through a recursive formula based on the prior hyperparameters and input data. Non-information priors are used in the setting of the prior hyperparameters. When judging the fitting effect of the multiple linear regression model on the sample for each segment, the characteristics of the regression model are used to evaluate the degree of fitting of the regression model to the sample by calculating the marginal likelihood. A greedy Bayesian cohesive sequence segmentation algorithm is used to optimize the exponential explosion problem in the traversal. During the iterative calculation process, a local optimum is selected each time, and local sequence segments are gradually merged from bottom to top through cohesion to obtain the sequence segmentation result.

[0037] Furthermore, in step (1.4) trajectory sequence clustering, the Hidden Dirichlet Allocation Model establishes three abstract types: document, topic, and word. The relationship between the three is described using a two-layer Dirichlet-polynomial conjugate. Different sequence segments are regarded as documents in the Hidden Dirichlet Allocation Model, trajectory samples are regarded as words in the Hidden Dirichlet Allocation Model, and trajectory primitive categories are regarded as topics in the Hidden Dirichlet Allocation Model. Thus, the Hidden Dirichlet Allocation Model is used to mine the trajectory primitive categories to which the trajectory sequence belongs.

[0038] The beneficial effects of this invention are:

[0039] This invention provides a multimodal trajectory prediction system and method for vehicles based on trajectory primitives. For the first time, it performs multimodal trajectory prediction for vehicles from the perspective of trajectory primitives. While ensuring the multimodal nature of the prediction results, this invention improves the interpretability of the trajectory prediction model. At the same time, this method only requires the historical trajectory of the vehicle and does not require the historical trajectories of surrounding vehicles, thus reducing the requirements of the trajectory prediction algorithm on the vehicle's perception range. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0041] Figure 2 A flowchart illustrating the process of establishing trajectory primitive libraries for various scenarios in this invention;

[0042] Figure 3 This is a schematic diagram of the trajectory prediction neural network structure of the present invention;

[0043] Figure 4 This is a schematic diagram of the prediction method of the present invention. Detailed Implementation

[0044] like Figure 1 As shown, this invention provides a multimodal trajectory prediction system for vehicles based on trajectory primitives, including a similar trajectory filtering module, a trajectory primitive library acquisition module, a primitive combination statistics module, a decoding prediction module, and a trajectory prediction network training module. The input of the multimodal trajectory prediction system for vehicles is the historical trajectory of the vehicle obtained by the environmental perception module of the autonomous vehicle. This trajectory is compared with the trajectory in the trajectory primitive library in the trajectory primitive library acquisition module to filter similar trajectories, count primitive combinations, obtain primitive combination probabilities and primitive sequence lengths. Using the network structure trained by the trajectory prediction network training module, the decoding prediction module decodes and predicts each primitive category one by one to achieve trajectory splicing, obtain the predicted trajectory under each primitive combination, and finally obtain the system output, namely, multiple possible driving trajectories of the vehicle and the probability of each trajectory. The output trajectory and probability are used as input to the decision planning module of the autonomous vehicle for decision planning.

[0045] See Figure 2-4 As shown, this invention provides a method for predicting the multimodal trajectory of a vehicle based on trajectory primitives, comprising the following steps:

[0046] (1) Establish trajectory primitive libraries for each scenario. This step is performed in the trajectory primitive library acquisition module, such as... Figure 2 As shown;

[0047] (1.1) Trajectory dataset acquisition and smoothing processing:

[0048] We selected open-source trajectory datasets for various scenarios, with the required features being the vehicle's horizontal and vertical coordinates. Due to issues with the trajectory data acquisition and processing methods, the vehicle's horizontal and vertical coordinates obtained from the datasets contain some noise compared to the actual vehicle coordinates. Therefore, we used a moving average method to smooth the trajectory data. The moving average formula is as follows:

[0049]

[0050] In the formula, p t x represents the filtering result at time t. t Let t represent the observed value at time t, and n represent the sliding window radius;

[0051] (1.2) Acquisition of trajectory lateral and longitudinal velocities and accelerations:

[0052] Take the smoothed vehicle lateral and longitudinal coordinates obtained in step (1.1), calculate the first and second derivatives with respect to time, and obtain the lateral and longitudinal velocities and accelerations of the trajectory.

[0053] (1.3) Trajectory sequence segmentation:

[0054] Adjacent samples with the same data performance trend are used as the data composition of the same sequence segment. The input is the lateral and longitudinal velocities and accelerations of the trajectory obtained in step (1.2). The lateral and longitudinal velocities and accelerations are divided into sequence segments by the Bayesian condensation sequence segmentation method to obtain the trajectory sequence segmentation result.

[0055] For the time series of lateral and longitudinal velocities and accelerations obtained in step (1.2), a piecewise linear model is used to describe the relationship between input and output. The sequence is segmented according to its changing trend, and the time series function of each segment can be represented by a multiple linear regression model. For the parameter matrix of the multiple linear regression model, the conjugate distribution is used for parameter fitting. The posterior parameters of the conjugate distribution are learned by continuously updating the distribution parameters recursively based on the prior hyperparameters and input data. Non-informational priors are used in the setting of the prior hyperparameters. When judging the fitting effect of the multiple linear regression model on the samples for each segment, the marginal likelihood is calculated to evaluate the degree of fit of the regression model to the samples, utilizing the characteristics of the regression model. A greedy Bayesian cohesive sequence segmentation algorithm is used to optimize the exponential explosion problem during traversal. During the iterative calculation process, a local optimum is selected each time, and local sequence segments are gradually merged from bottom to top through cohesion to obtain the sequence segmentation result.

[0056] (1.4) Trajectory sequence clustering:

[0057] The trajectory sequence clustering uses Gaussian mixture model-Hidden Dirichlet Allocation model. First, the trajectory sequence segments obtained in step (1.3) are clustered using Gaussian mixture model. Then, the cluster labels obtained by clustering are regarded as words in the Hidden Dirichlet Allocation model and used to replace the original driving samples as input data for Hidden Dirichlet Allocation to obtain the trajectory sequence segment categories. It is believed that trajectory sequence segments of the same category are generated by a driving mode or driving behavior, which is a type of trajectory primitive.

[0058] The Hidden Dirichlet Allocation Model establishes three abstract types: document, topic, and word. The relationship between the three is described using a two-layer Dirichlet-polynomial conjugate. Different sequence segments are regarded as documents in the Hidden Dirichlet Allocation Model, trajectory samples are regarded as words in the Hidden Dirichlet Allocation Model, and trajectory primitive categories are regarded as topics in the Hidden Dirichlet Allocation Model. Thus, the Hidden Dirichlet Allocation Model is used to mine the trajectory primitive categories to which the trajectory sequence belongs.

[0059] By combining the smoothed vehicle horizontal and vertical coordinates obtained in step (1.1) with their corresponding velocity and acceleration sequences through segmentation and clustering to obtain the primitive categories, the trajectory primitive library for each scene can be obtained.

[0060] (2) Construction and training of trajectory prediction networks corresponding to primitive categories. This step is performed in the trajectory prediction network training module, such as... Figure 3As shown;

[0061] (2.1) Filter data by primitive category:

[0062] The trajectory primitive library of each scene obtained in step (1) is filtered according to primitive category; specifically, the trajectory segment with the same primitive category and all trajectories before the trajectory segment are used as training data under the primitive category.

[0063] (2.2) Trajectory prediction network construction and training:

[0064] An encoder-decoder structure based on an LSTM network is adopted. The data filtered by primitive category in step (2.1) is used as training data, the trajectory segments with the same primitive category are used as predicted trajectories, and all trajectories before the trajectory segment are used as historical trajectories. The network weights are obtained through training.

[0065] (3) Multimodal trajectory prediction, such as Figure 4 As shown:

[0066] (3.1) Similarity calculation:

[0067] After the autonomous vehicle perception system acquires a segment of the historical trajectory of a vehicle in a certain scenario, it uses the historical trajectory of the vehicle and the trajectory primitive library of each scenario obtained in step (1) as input. The historical trajectory of the vehicle is compared with the trajectory data in the trajectory primitive library of each scenario to calculate the similarity between the historical trajectory to be predicted and the trajectory in the trajectory primitive library of the corresponding scenario. The similarity is calculated using multi-line position distance, as shown in the following formula:

[0068]

[0069]

[0070] Q and S are two trajectories, Area i Let w be the area of ​​the i-th region enclosed by the two trajectories. i As the weight, I i Let be the i-th intersection point of the two trajectories, and Length be the length of the trajectory between the two points.

[0071] (3.2) Similar trajectory filtering:

[0072] By setting a similarity threshold, trajectories below the threshold are discarded, and those above the threshold are selected, i.e., trajectories in the trajectory library that are similar to the historical trajectory to be predicted.

[0073] Steps (3.1) and (3.2) are performed in the similar trajectory filtering module;

[0074] The primitive combination statistics module counts the primitive combination methods after similar segments, calculates the frequency of primitive combination occurrences, counts the time length of a single primitive under a certain primitive combination, and calculates the average value, which is the predicted time length of a single primitive.

[0075] (3.3) Decode the prediction using the corresponding primitive prediction network:

[0076] The historical trajectory of the vehicle to be predicted, the various trajectory primitive combinations obtained in step (3.2), and the time series length of each primitive under the corresponding combination are input into the trajectory prediction neural network structure of the corresponding primitive category obtained in step (2.2). The primitive categories are decoded and predicted one by one to realize the splicing of trajectories and obtain the predicted trajectory under each primitive combination. The frequency of occurrence of the primitive combination obtained in step (3.2) is used as the confidence of the predicted trajectory under that combination, and finally the multimodal trajectory prediction result is obtained.

[0077] Step (3.3) is performed in the decoding prediction module.

Claims

1. A multimodal trajectory prediction system for vehicles based on trajectory primitives, characterized in that: The system includes a similar trajectory filtering module, a trajectory primitive library acquisition module, a primitive combination statistics module, a decoding and prediction module, and a trajectory prediction network training module. The input to the multimodal trajectory prediction system for the vehicle is the historical trajectory of the vehicle obtained by the autonomous vehicle's environmental perception module. This trajectory is compared with the trajectory in the trajectory primitive library acquired by the trajectory primitive library acquisition module to filter similar trajectories, statistically analyze primitive combinations, obtain primitive combination probabilities and primitive sequence lengths, and then, using the network structure trained by the trajectory prediction network training module, decode and predict each primitive category in the decoding and prediction module to achieve trajectory splicing and obtain the predicted trajectory under each primitive combination. Finally, the system output is obtained, namely, multiple possible driving trajectories of the vehicle and the probability of each trajectory occurring. The output trajectories and probabilities serve as input to the autonomous vehicle's decision-making and planning module for decision-making and planning. The trajectory primitive library acquisition module establishes trajectory primitive libraries for various scenarios, including the following steps: (1.1) Trajectory dataset acquisition and smoothing processing: We selected open-source trajectory datasets for various scenarios, with the required features being the vehicle's horizontal and vertical coordinates. Due to issues with the trajectory data acquisition and processing methods, the vehicle's horizontal and vertical coordinates obtained from the datasets contain some noise compared to the actual vehicle coordinates. Therefore, we used a moving average method to smooth the trajectory data. The moving average formula is as follows: In the formula, express The filtering result at time 10:

00. express The observed value at time, Represents the radius of the sliding window; (1.2) Acquisition of trajectory lateral and longitudinal velocities and accelerations: The smoothed vehicle lateral and longitudinal coordinates obtained in step (1.1) are used to calculate the first and second derivatives with respect to time to obtain the lateral and longitudinal velocities and accelerations of the trajectory. (1.3) Trajectory sequence segmentation: Adjacent samples with the same data performance trend are used as the data composition of the same sequence segment. The input is the lateral and longitudinal velocities and accelerations of the trajectory obtained in step (1.2). The lateral and longitudinal velocities and accelerations are divided into sequence segments by the Bayesian condensation sequence segmentation method to obtain the trajectory sequence segmentation result. For the time series of lateral and longitudinal velocities and accelerations obtained in step (1.2), a piecewise linear model is used to describe the relationship between input and output. The sequence is segmented according to the trend of change. The time series function of each sequence segment is represented by a multiple linear regression model. For the parameter matrix of the multiple linear regression model, the conjugate distribution is used for parameter fitting. The posterior parameters of the conjugate distribution are continuously updated and learned based on the prior hyperparameters and input data through a recursive method. Non-information priors are used in the setting of the prior hyperparameters. When judging the fitting effect of the multiple linear regression model of each segment on the sample, the characteristics of the regression model are used to evaluate the degree of fitting of the regression model to the sample by calculating the marginal likelihood. The greedy Bayesian cohesive sequence segmentation algorithm is used to optimize the problem of exponential explosion in traversal. In the iterative calculation process, a local optimum is selected each time, and local sequence segments are gradually merged from bottom to top through cohesion to obtain the sequence segmentation result. (1.4) Trajectory sequence clustering: The trajectory sequence clustering uses Gaussian mixture model-Hidden Dirichlet Allocation model. First, the trajectory sequence segments obtained in step (1.3) are clustered using Gaussian mixture model. Then, the cluster labels obtained by clustering are regarded as words in the Hidden Dirichlet Allocation model and used to replace the original driving samples as input data for Hidden Dirichlet Allocation to obtain the trajectory sequence segment categories. It is believed that trajectory sequence segments of the same category are generated by a driving mode or driving behavior, which is a type of trajectory primitive. The Hidden Dirichlet Allocation Model establishes three abstract types: document, topic, and word. The relationship between the three is described using a two-layer Dirichlet-polynomial conjugate. Different sequence segments are regarded as documents in the Hidden Dirichlet Allocation Model, trajectory samples are regarded as words in the Hidden Dirichlet Allocation Model, and trajectory primitive categories are regarded as topics in the Hidden Dirichlet Allocation Model. Thus, the Hidden Dirichlet Allocation Model is used to mine the trajectory primitive categories to which the trajectory sequence belongs. By combining the smoothed vehicle horizontal and vertical coordinates obtained in step (1.1) with their corresponding velocity and acceleration sequences through segmentation and clustering to obtain the primitive categories, a trajectory primitive library for each scene is obtained.

2. The multimodal trajectory prediction system for vehicles based on trajectory primitives according to claim 1, characterized in that: The trajectory prediction network training module for the aforementioned primitive category includes the following steps for building and training the trajectory prediction network: (2.1) Filtering data by primitive category: The trajectory primitive library of each scene obtained in step (1) is filtered according to primitive category, and the trajectory segment with the same primitive category and all trajectories before the trajectory segment are used as training data under the primitive category. (2.2) Trajectory prediction network construction and training: An encoder-decoder structure based on an LSTM network is adopted. The data filtered by primitive category in step (2.1) is used as training data, the trajectory segments with the same primitive category are used as predicted trajectories, and all trajectories before the trajectory segment are used as historical trajectories. The network weights are obtained by training.

3. The multimodal trajectory prediction system for vehicles based on trajectory primitives according to claim 1, characterized in that: The similar trajectory filtering module performs the following steps: (3.1) Similarity calculation: After the autonomous vehicle perception system acquires a segment of the historical trajectory of a vehicle in a certain scenario, the historical trajectory of the vehicle and the trajectory primitive library of each scenario obtained in step (1) are used as input. The historical trajectory of the vehicle is compared with the trajectory data in the trajectory primitive library of each scenario, and the similarity between the historical trajectory to be predicted and the trajectory in the trajectory primitive library of the corresponding scenario is calculated. The similarity is calculated using multi-line position distance, and the formula is as follows: Q and S are two trajectories. The first trajectories enclosed by the two trajectories Area of ​​each region As weight, For the first of the two trajectories One intersection point, The length of the trajectory between the two points; (3.2) Similar trajectory filtering: By setting a similarity threshold, trajectories below the threshold are discarded, and those above the threshold are selected, i.e., trajectories in the trajectory library that are similar to the historical trajectory to be predicted. In the aforementioned primitive combination statistics module, the primitive combination methods after similar segments are counted, the frequency of primitive combination occurrence is calculated, the time length of a single primitive under a certain primitive combination is counted, and the average value is the predicted time length of a single primitive.

4. The multimodal trajectory prediction system for vehicles based on trajectory primitives according to claim 1, characterized in that: The decoding and prediction module uses the corresponding primitive prediction network for decoding and prediction: The historical trajectory of the vehicle to be predicted, the various trajectory primitive combinations obtained in step (3.2), and the time series length of each primitive under the corresponding combination are input into the trajectory prediction neural network structure of the corresponding primitive category obtained in step (2.2). The primitive categories are decoded and predicted one by one to realize the splicing of trajectories and obtain the predicted trajectory under each primitive combination. The frequency of occurrence of the primitive combination obtained in step (3.2) is used as the confidence of the predicted trajectory under that combination, and finally the multimodal trajectory prediction result is obtained.

5. A method for predicting the multimodal trajectory of a vehicle based on trajectory primitives, characterized in that: Includes the following steps: (1) Establish trajectory primitive libraries for each scenario; including the following steps: (1.1) Trajectory dataset acquisition and smoothing processing: We selected open-source trajectory datasets for various scenarios, with the required features being the vehicle's horizontal and vertical coordinates. Due to issues with the trajectory data acquisition and processing methods, the vehicle's horizontal and vertical coordinates obtained from the datasets contain some noise compared to the actual vehicle coordinates. Therefore, we used a moving average method to smooth the trajectory data. The moving average formula is as follows: In the formula, express The filtering result at time 10:

00. express The observed value at time, Represents the radius of the sliding window; (1.2) Acquisition of trajectory lateral and longitudinal velocities and accelerations: The smoothed vehicle lateral and longitudinal coordinates obtained in step (1.1) are used to calculate the first and second derivatives with respect to time to obtain the lateral and longitudinal velocities and accelerations of the trajectory. (1.3) Trajectory sequence segmentation: Adjacent samples with the same data performance trend are used as the data composition of the same sequence segment. The input is the lateral and longitudinal velocities and accelerations of the trajectory obtained in step (1.2). The lateral and longitudinal velocities and accelerations are divided into sequence segments by the Bayesian condensation sequence segmentation method to obtain the trajectory sequence segmentation result. (1.4) Trajectory sequence clustering: First, the trajectory sequence segments obtained in step (1.3) are clustered using a Gaussian mixture model. Then, the cluster labels obtained from the clustering are regarded as words in the Hidden Dirichlet Allocation model. They are used to replace the original driving samples as input data for Hidden Dirichlet Allocation to obtain the trajectory sequence segment categories. It is believed that trajectory sequence segments of the same category are generated by a driving mode or driving behavior, which is a type of trajectory primitive. The smoothed vehicle horizontal and vertical coordinates obtained in step (1.1) are combined with their corresponding velocity and acceleration sequences through segmentation and clustering to obtain the primitive categories for each scene, thus obtaining the trajectory primitive library for each scene. (2) Construction and training of trajectory prediction networks corresponding to primitive categories; (2.1) Filtering data by primitive category: The trajectory primitive library of each scene obtained in step (1) is filtered according to primitive category, and the trajectory segment with the same primitive category and all trajectories before the trajectory segment are used as training data under the primitive category. (2.2) Trajectory prediction network construction and training: An encoder-decoder structure based on an LSTM network is adopted. The data filtered by primitive category in step (2.1) is used as training data, the trajectory segments with the same primitive category are used as predicted trajectories, and all trajectories before the trajectory segment are used as historical trajectories. The network weights are obtained through training, and the trajectory prediction neural network structure of the corresponding primitive category is obtained. (3) Multimodal trajectory prediction: (3.1) Similarity calculation: After the autonomous vehicle perception system obtains a segment of the historical trajectory of a vehicle in a certain scenario, the historical trajectory of the vehicle and the trajectory primitive library of each scenario obtained in step (1) are used as input. The historical trajectory of the vehicle is compared with the trajectory data in the trajectory primitive library of each scenario, and the similarity between the historical trajectory to be predicted and the trajectory in the trajectory primitive library of the corresponding scenario is calculated. (3.2) Similar trajectory filtering: By setting a similarity threshold, those less than the threshold are discarded, and those greater than the threshold, i.e., those similar to the historical trajectory to be predicted, are selected. The primitive combination methods after the similar segments are counted, the frequency of primitive combination occurrence is calculated, the time length of a single primitive under a certain primitive combination is counted, and the average value is the predicted time length of a single primitive. (3.3) Decode the prediction using the corresponding primitive prediction network: The historical trajectory of the vehicle to be predicted, the various trajectory primitive combinations obtained in step (3.2), and the time series length of each primitive under the corresponding combination are input into the trajectory prediction neural network structure of the corresponding primitive category obtained in step (2.2). The primitive categories are decoded and predicted one by one to realize the splicing of trajectories and obtain the predicted trajectory under each primitive combination. The frequency of occurrence of the primitive combination obtained in step (3.2) is used as the confidence of the predicted trajectory under that combination, and finally the multimodal trajectory prediction result is obtained.

6. The method for predicting multimodal trajectories of vehicles based on trajectory primitives according to claim 5, characterized in that: In step (1.3) trajectory sequence segmentation, for the time series of lateral and longitudinal velocities and accelerations obtained in step (1.2), a piecewise linear model is used to describe the relationship between input and output. The sequence is segmented according to the changing trend, and the time series function of each sequence segment is represented by a multiple linear regression model. For the parameter matrix of the multiple linear regression model, the conjugate distribution is used for parameter fitting. The posterior parameters of the conjugate distribution are continuously updated and learned based on the prior hyperparameters and input data through a recursive method. Non-information priors are used in the setting of the prior hyperparameters. When judging the fitting effect of the multiple linear regression model of each segment on the sample, the characteristics of the regression model are used to evaluate the degree of fitting of the regression model to the sample by calculating the marginal likelihood. The greedy Bayesian cohesive sequence segmentation algorithm is used to optimize the problem of exponential explosion in traversal. In the iterative calculation process, a local optimal solution is selected each time, and local sequence segments are gradually merged from bottom to top through cohesion to obtain the trajectory sequence segmentation result.

7. The method for predicting multimodal trajectories of vehicles based on trajectory primitives according to claim 5, characterized in that: In step (1.4) trajectory sequence clustering, the Hidden Dirichlet Allocation Model establishes three abstract types: document, topic, and word. The relationship between the three is described by two-layer Dirichlet-polynomial conjugation. Different sequence segments are regarded as documents in the Hidden Dirichlet Allocation Model, trajectory samples are regarded as words in the Hidden Dirichlet Allocation Model, and trajectory primitive categories are regarded as topics in the Hidden Dirichlet Allocation Model. Thus, the Hidden Dirichlet Allocation Model is used to mine the trajectory primitive categories to which the trajectory sequence belongs.

8. The method for predicting multimodal trajectories of vehicles based on trajectory primitives according to claim 5, characterized in that: Step (3.1) calculates the similarity using the multiline positional distance formula, as follows: Q and S are two trajectories. The first trajectories enclosed by the two trajectories Area of ​​each region As weight, For the first of the two trajectories One intersection point, Let be the length of the trajectory between the two points.

Images