Modal controllable vehicle trajectory prediction method and device based on dynamic noise mechanism, and medium

Abstract

The invention relates to a modal controllable vehicle trajectory prediction method and device based on a dynamic noise mechanism, and a medium, and the method comprises the following steps: obtaining a traffic scene, carrying out the semantic division of the traffic scene, obtaining different classification scenes, carrying out the vehicle trajectory prediction through a trained multi-modal trajectory prediction model based on a GAN, and obtaining a vehicle trajectory prediction result. Meanwhile, optimal noise is dynamically injected according to the complexity of different classification scenes, and a vehicle track is finally predicted; wherein the multi-modal trajectory prediction model based on the GAN comprises a generator G and a discriminator D, the generator G and the discriminator D are subjected to adversarial training by adopting a nearest neighbor sampling mechanism, and the trained multi-modal trajectory prediction model based on the GAN is obtained. Compared with the prior art, the method has the advantages of being high in trajectory prediction precision, obvious in modal regulation and control mechanism effect, capable of achieving cross-scene optimal balance and the like.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle trajectory prediction technology, and in particular to a modal controllable vehicle trajectory prediction method, device and medium based on dynamic noise mechanism. Background Technology

[0002] With the development and application of autonomous driving technology, safety issues have increasingly attracted public attention. Vehicle behavior prediction, as a key foundation for assessing traffic environment risks, is an essential component for achieving safe autonomous driving. With the advancement of artificial intelligence technology, vehicle trajectory prediction methods have shifted from traditional physics-based models and classical machine learning methods to deep learning models, which have shown significant advantages, especially in long-term prediction tasks. Current research generally agrees that trajectory prediction results exhibit multimodal characteristics, and multimodal trajectory prediction has become a common method for representing the uncertainty of future paths. The multimodality in trajectory prediction stems from two main factors. First, the inherent randomness and uncertainty in human decision-making, and this uncertainty propagates and interacts with traffic objects, forming multiple possible evolutionary patterns. Second, the input to prediction models is often incomplete. Although most models only use historical trajectory coordinates and map information as input, in reality, factors such as vehicle type and function, special traffic signs, surrounding traffic conditions, and weather can also affect trajectory evolution. Therefore, similar or nearly identical inputs may lead to different output trajectories, exhibiting multimodal characteristics.

[0003] Numerous studies have focused on the multimodal nature of trajectories. Some research incorporates external expert knowledge to guide the prediction process, generating physically plausible and interpretable multimodal outputs. Others employ rule-based heuristic sampling methods to generate diverse trajectory patterns, then select the most accurate prediction set through offline optimization. Still others propose anchor-based methods that first predict the target endpoint and then cluster similar trajectories to balance pattern diversity and prediction accuracy. Some studies assume that different patterns are independent and train different models on categorical data to generate specialized trajectory predictions for each pattern. However, these studies seem to learn a modal distribution and generate corresponding modalities based on this static distribution when predicting trajectories. Clearly, the uncertainty of driving behavior varies significantly across different traffic scenarios. For example, at intersections, due to high traffic volume and dense pedestrian traffic, vehicle paths may be straight, left-turn, or right-turn, and may even require yielding to pedestrians. This results in higher uncertainty in driving behavior patterns, thus necessitating the generation of more diverse trajectory predictions. On straight sections of road, due to the simple road conditions and relatively predictable behavior of traffic participants, trajectory patterns should be more concentrated, and the diversity of predicted trajectories is low. That is, different traffic environments should correspond to different modal distributions, but currently the modal distribution of predicted trajectories cannot automatically adapt to changes in the traffic environment. Summary of the Invention

[0004] The purpose of this invention is to provide a modally controllable vehicle trajectory prediction method, device, and medium based on dynamic noise mechanism to improve trajectory prediction accuracy in complex interactive environments.

[0005] The objective of this invention can be achieved through the following technical solutions: A modally controllable vehicle trajectory prediction method based on dynamic noise mechanism includes the following steps: The traffic scene is acquired and semantically segmented to obtain different classification scenes. A trained GAN-based multimodal trajectory prediction model is used to predict vehicle trajectories. At the same time, the optimal noise is dynamically injected according to the complexity of different classification scenes to finally predict the vehicle trajectory. The GAN-based multimodal trajectory prediction model includes a generator G and a discriminator D. The generator G and discriminator D are trained adversarially using a nearest neighbor sampling mechanism to obtain the trained GAN-based multimodal trajectory prediction model.

[0006] Furthermore, both the generator G and the discriminator D are composed of a context encoder and a trajectory decoder.

[0007] Furthermore, the execution process of the generator G includes: The context encoder is used to extract historical trajectories and map features to generate a fixed-dimensional latent representation. , where the potential representation Represented as: In the formula, For context encoder, For historical trajectory and map data, For the real number field, For dimensions; During each forward pass, a noise vector is sampled. and add to the potential representation The underlying representation obtained after perturbation , is represented as: The perturbated latent representation The generated trajectory is passed to the trajectory decoder to obtain the generated trajectory. ,in, For trajectory decoders.

[0008] Furthermore, the context encoder is a graph neural network.

[0009] Furthermore, the loss function of the generator G is: in: In the formula, Let G be the total loss function of the generator. To supervise the loss, used to measure the predicted trajectory With the actual trajectory The differences between them For hyperparameters, To combat losses, and to control the reliability of the generated trajectory, The time step length for predicting the trajectory, , The predicted and actual trajectories at time steps are respectively. The position vector, For historical trajectory and map data, The generator samples the trajectory that is closest to the real one. A generated trajectory is called the optimal generated trajectory. As expected, For the actual trajectory, ~pdata This represents a sample taken from the actual trajectory. z Let z~pz represent noise, which is sampled from a Gaussian noise distribution. Indicates in X Under the constraints, discriminator D determines the true trajectory Y The rating, Indicates in X Under the constraints, discriminator D determines the optimal generated trajectory. The rating.

[0010] Furthermore, the steps of adversarially training the generator G and the discriminator D include: In each trajectory prediction process, the generator G is used to sample and generate the trajectory at one time. Generate trajectory ,in, For historical trajectory and map data, For Gaussian noise variance control; In the Select the generated trajectory that matches the real trajectory exist The closest generated trajectory under the distance metric is taken as the best generated trajectory. , is represented as: In the formula, For the first Generate trajectory With the actual trajectory Between distance; Discriminator D undergoes adversarial training under the optimization objective to distinguish the real trajectory. and optimal generated trajectory ; The adversarial training process is completed by iteratively training the generator G and the discriminator D through forward and backward propagation.

[0011] Furthermore, the optimization objective is expressed as: In the formula, To optimize the objective, ~pdata This represents a sample taken from the actual trajectory. For the actual trajectory, As expected, Indicates in X Under the constraints, discriminator D determines the true trajectory Y The rating, The trajectory that most closely resembles the true trajectory generated by the generator G is called the optimal generated trajectory. Indicates in X Under the constraints, discriminator D determines the optimal generated trajectory. The rating, zLet z be noise, and z~pz represent the noise input sampled from a Gaussian noise distribution.

[0012] Furthermore, different semantic segments are performed based on scene semantic features. If the semantic features of the scene are a map structure, then the traffic scene is divided into intersection areas and non-intersection areas; If the semantic feature of the scene is a road attribute, then the traffic scene is divided into motor vehicle lanes and non-motor vehicle lanes; If the semantic feature of the scene is weather information, then the traffic scene will be divided into severe weather scene and non-severe weather scene.

[0013] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs stored in the memory, said one or more programs including instructions for executing the modally controllable vehicle trajectory prediction method based on the dynamic noise mechanism as described above.

[0014] The present invention also provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the modally controllable vehicle trajectory prediction method based on the dynamic noise mechanism as described above.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention models environmental uncertainty by dynamically injecting optimal noise into the multimodal trajectory prediction model according to the complexity of different classification scenarios, thereby achieving fine control of trajectory diversity, improving the adaptability of the model in different scenarios, and enhancing the prediction accuracy of the model in complex interactive environments.

[0016] (2) To ensure that the generated trajectories after noise modulation not only possess diversity but also maintain physical rationality and consistency with the true distribution, this invention introduces a joint training framework combining adversarial learning and supervised learning. An adversarial discriminator is used to discriminate between the generated and true trajectories. This invention employs conditional adversarial learning to align the true conditional distribution with the generated distribution given historical input. During training, this invention considers nearest-neighbor samples as representatives of the generated distribution in the neighborhood of the true trajectory, while simultaneously introducing the log-likelihood objective of the discriminator to measure the distributional difference between the two. Compared to traditional trajectory prediction models that rely solely on point-to-point regression errors, this adversarial term essentially minimizes the divergence between conditional distributions, providing more discriminative gradient signals and calibrating tail probabilities. Therefore, it can improve trajectory distribution consistency and training stability without sacrificing diversity.

[0017] (3) In this invention, among the multiple candidate trajectories sampled by the generator, some samples may deviate significantly from the statistical distribution of the real trajectory. If these samples are directly included in the optimization process of adversarial loss, the discriminator will overemphasize these anomalous samples with large deviations, leading to instability of the training objective and increasing the risk of mode collapse. The nearest neighbor sampling mechanism designed in this invention avoids this problem by selecting only the optimal trajectory to be included in the discriminator. By selecting only the samples closest to the real trajectory to participate in adversarial optimization, the model can ensure that the game process between the generator and the discriminator is concentrated in the distribution neighborhood, thereby enabling the generator to gradually approach the distribution of the real trajectory during the iteration process. At the same time, this mechanism effectively suppresses the high variance gradient introduced by outlier samples, improves the stability of adversarial training, and enhances the physical rationality and distribution consistency of the generated trajectories while maintaining the diversity of generation.

[0018] (4) To better model the uncertainties of the traffic environment, during the training phase, this invention incorporates noise intensity σ as an adjustable control factor into the generator to dynamically adjust the modal coverage of the generated trajectory. The noise follows a Gaussian distribution. During the inference phase, this invention utilizes a dynamic noise selection mechanism to adaptively adjust the noise intensity based on the scene. In complex scenarios, the noise intensity is increased to introduce stronger uncertainty representation capabilities and improve trajectory modal diversity; in simple scenarios, the noise intensity is reduced to avoid over-divergent prediction, thereby achieving adaptive adjustment of trajectory modal diversity. This mechanism enhances the model's ability to represent uncertainties in different scenarios and achieves an adaptive balance between diversity and accuracy in predicted trajectories.

[0019] (5) By regulating the modal distribution of the predicted trajectory, the downstream decision-making module can achieve a balance between robustness and efficiency under different risk scenarios, thereby further improving decision-making performance and overall system security. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This forms the overall structural framework of the present invention; Figure 3 This invention addresses the impact of dynamic noise on prediction accuracy in intersection and non-intersection scenarios. Figure 4 This invention relates to the impact of dynamic noise on the prediction accuracy of motor vehicle lanes and non-motor vehicle lanes. Figure 5 The multimodal trajectory prediction results under different input noise intensities of the present invention are visualized. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0022] Example 1 This embodiment provides a modally controllable vehicle trajectory prediction method based on a dynamic noise mechanism, such as... Figure 1 The overall structural framework shown illustrates how this method models environmental uncertainty by introducing dynamic noise into the multimodal trajectory prediction model, enabling fine-tuning of trajectory diversity, improving the model's adaptability in different scenarios, and enhancing prediction accuracy in complex interactive environments. By adjusting the modal distribution of the predicted trajectory, the downstream decision-making module can achieve a balance between robustness and efficiency under different risk scenarios, further improving decision performance and overall system security. Specifically, as... Figure 1 As shown, the method includes the following steps: Step 1: Build a training framework for a multimodal trajectory prediction model based on the GAN architecture.

[0023] This embodiment builds a multimodal trajectory prediction model training framework based on the GAN (Generative Adversarial Network) architecture. The model training framework consists of a generator G and a discriminator D. The generator consists of two components: a context encoder and a noise injection decoder.

[0024] Step 1.1 Design of the Generator's Context Encoder Here, the encoder structure for multimodal trajectory prediction is retained, and a GCN (Graph Neural Network) is used as the context encoder to extract historical trajectory and map features, generating a fixed-dimensional latent representation: Where X represents the historical trajectory, map data, etc. of the past 5 seconds, with a dimension of d=128.

[0025] Step 1.2 Design of the generator's noise injection decoder To ensure the multimodal nature of the predicted trajectory, this invention injects a Gaussian noise vector into the decoding process. During each forward pass, a noise vector is sampled. and add it directly to the encoded context vector. In the middle. Then, the perturbed representation. It is passed to the trajectory decoder to generate a possible future trajectory. By sampling different The model can generate diverse trajectory predictions based on the same scene context.

[0026] Step 1.3 Design of the generator's loss function To enable the generator to produce high-accuracy predicted trajectories, the loss function consists of two parts: supervised regression loss of the trajectory. Losses in the fight against GANs : in It is a loss of supervision. Controlling the reliability of the generated trajectory, It is a hyperparameter, set to 0.001, used to balance the two losses. Mean squared error (MSE) is typically used to measure the predicted trajectory. With the actual trajectory Differences between them: in The time step length for predicting the trajectory is 60 (6 seconds / sampling frequency of 10 Hz). and These are the predicted and actual trajectories at time steps. The position vector. And the adversarial loss. The following is used to deceive the discriminator, making the generated trajectory as close as possible to the distribution of the real trajectory: in, It is the actual trajectory. The trajectory generated by the generator that is closest to the real trajectory is called the optimal generated trajectory, where z is noise. As expected, ~pdata Let z represent a sample taken from the real trajectory, and z~pz represent the noise input sampled from a Gaussian noise distribution. Indicates in X Under the constraints, discriminator D determines the true trajectory Y The rating, Indicates in X Under the constraints, discriminator D determines the optimal generated trajectory. The rating.

[0027] Step 1.4 Discriminator Design The network structure can be designed with reference to the generator, except that no noise needs to be injected, and the loss function can also be designed with reference to the traditional loss function of GAN.

[0028] Step 2: Adversarial training of generator G and discriminator D based on nearest neighbor sampling mechanism.

[0029] Step 2.1 Design the nearest neighbor sampling mechanism In each trajectory prediction process, we will use a generator to sample and generate the trajectory once. Trajectory: in, Represents a generator. This indicates input information such as historical trajectories and maps. This represents the predicted trajectory generated by the generator sampling. This represents the Gaussian noise variance control term. Among all sampled trajectories, only those closely resembling the true trajectory are selected. exist The closest sample in terms of distance metric is used in the adversarial loss. The sampling trajectory selected for the loss function calculation is defined as: Step 2.2 Conduct adversarial training between the generator and the discriminator. Discriminator Trained to distinguish real trajectories and the selected generated trajectory Both are based on the same context encoder to obtain features. generator The trainer generates diverse trajectories that can successfully deceive the discriminator. The optimization objectives are as follows: .

[0030] Step 3: When model G is inferred, a dynamic noise mechanism is used to control the modal distribution of the trajectory.

[0031] Step 3.1 Automatic Classification of Traffic Scenarios Based on the experience of transportation experts and knowledge of the traffic environment, the scene is semantically segmented according to its semantic features. Specifically, the traffic scene can be automatically divided into intersection areas and non-intersection road environments according to the map structure; it can also be automatically divided into motor vehicle lanes and non-motor vehicle lanes according to road attributes; or it can be divided into severe weather and non-severe weather scenes according to weather information, thereby providing prior semantic information for subsequent dynamic noise generation.

[0032] Step 3.2 Dynamic noise injection for different classification scenarios For complex scenarios (such as intersections, non-motorized vehicle lanes, and inclement weather), a larger noise variance is injected during the model inference process. Conversely, for simple scenarios, a smaller noise variance is injected. That is, the optimal noise intensity is automatically selected based on the scenario complexity, and the noise follows a Gaussian distribution. : in, .

[0033] In this embodiment, a temporary test set is constructed using a subset of scenarios. Different noise variances are applied to scenarios of different categories to obtain the optimal dynamic noise. The results are as follows: Figure 3 (Intersection and non-intersection scenarios) and Figure 4 As shown in the diagram (motor vehicle lanes and non-motor vehicle lanes), the asterisk (*) indicates the optimal noise input for uncertainty modeling. It can be seen that the optimal variance for intersections is 0.02, while the optimal variance for non-intersections is 0.0002, indicating that the optimal noise variance for intersections is much higher than that for non-intersections. Similarly, the optimal variance for non-motor vehicle lanes is 0.005, while the optimal variance for motor vehicle lanes is 0.001, indicating that the optimal variance for non-motor vehicle lanes is much higher than that for motor vehicle lanes.

[0034] Step 3.3 Analysis and Visualization of Prediction Results The optimal noise variance is used to sample and generate predicted trajectories in different scenarios. Experimental results show that adding noise can effectively improve prediction accuracy regardless of the scenario. The dynamic noise control mechanism proposed in this invention plays a significant role and achieves optimal balance across scenarios.

[0035] This embodiment uses the classic trajectory prediction framework LaneGCN as the baseline trajectory prediction model to evaluate the effectiveness of the proposed invention in real-world scenarios. Based on the Argoverse 2 urban road dataset, static and dynamic noise mechanisms (in this invention) are integrated into the framework. Here, "static" refers to Gaussian noise with a fixed mean and variance, and a static distribution. Five seconds of historical trajectory data are used to predict the trajectory for the next six seconds, and minADE and minFDE are used as metrics for prediction accuracy. Experimental results are shown in Table 1. Here, "static noise" refers to Gaussian noise following a fixed mean and variance distribution, while "dynamic noise" refers to Gaussian noise with a fixed mean but variable variance. The dynamic noise method proposed in this invention achieves optimal performance across all evaluation dimensions, verifying the effectiveness of the invention.

[0036] Meanwhile, to explore the modal diversity of trajectory prediction under different input noise intensities, different variances were injected under fixed input conditions. Multiple sets of predicted trajectories were generated using Gaussian noise, and their impact on modal structure was observed. For example... Figure 5 As shown, with As the magnitude increases, the generated trajectories gradually exhibit richer variations in spatial distribution: from concentrated prediction in a single direction to a bifurcation structure covering multiple directions, fully demonstrating the control capability of multimodal generation.

[0037] Table 1. Comparison of the impact of noise input on prediction accuracy in different scenarios (Unit is meters, each cell represents the minADE / minFDE index, the smaller the better) Experimental results on public datasets and classic trajectory prediction models demonstrate that, compared to noise-free benchmark models and static noise methods, the dynamic noise method proposed in this invention achieves the highest trajectory prediction accuracy. The modal regulation mechanism plays a significant role, achieving optimal balance across scenarios, and the method is advanced. Experimental results also show that this invention can dynamically adjust the diversity of predicted trajectory modes according to the complexity of the traffic environment, enabling downstream decision-making modules to achieve a balance between robustness and efficiency under different risk scenarios, thereby supporting the improvement of overall system safety and decision-making performance.

[0038] The advantages of the method in this embodiment are as follows: (1) To ensure that the generated trajectories after noise modulation not only possess diversity but also maintain physical rationality and consistency with the real distribution, this invention introduces a joint training framework combining adversarial learning and supervised learning. An adversarial discriminator is used to discriminate between the generated and real trajectories. This invention employs conditional adversarial learning to align the real conditional distribution with the generated distribution given historical input. During training, this invention considers nearest-neighbor samples as representatives of the generated distribution in the neighborhood of the real trajectory, while simultaneously introducing the log-likelihood objective of the discriminator to measure the distributional difference between the two. Compared to traditional trajectory prediction models that rely solely on point-to-point regression errors, this adversarial term essentially minimizes the divergence between conditional distributions, providing more discriminative gradient signals and calibrating tail probabilities. Therefore, it can improve trajectory distribution consistency and training stability without sacrificing diversity.

[0039] (2) In this invention, among the multiple candidate trajectories sampled by the generator, some samples may deviate significantly from the statistical distribution of the real trajectory. If these samples are directly included in the optimization process of adversarial loss, the discriminator will overemphasize these anomalous samples with large deviations, leading to instability of the training objective and increasing the risk of mode collapse. The nearest neighbor sampling mechanism designed in this invention avoids this problem by selecting only the optimal trajectory to be included in the discriminator. By selecting only the samples closest to the real trajectory to participate in adversarial optimization, the model can ensure that the game process between the generator and the discriminator is concentrated in the distribution neighborhood, thereby enabling the generator to gradually approach the distribution of the real trajectory during the iteration process. At the same time, this mechanism effectively suppresses the high variance gradient introduced by outlier samples, improves the stability of adversarial training, and enhances the physical rationality and distribution consistency of the generated trajectories while maintaining the diversity of generation.

[0040] (3) To better model the uncertainties of the traffic environment, during the training phase, this invention incorporates noise intensity σ as an adjustable control factor into the generator to dynamically adjust the modal coverage of the generated trajectory. The noise follows a Gaussian distribution. During the inference phase, this invention utilizes a dynamic noise selection mechanism to adaptively adjust the noise intensity based on the scene. In complex scenarios, the noise intensity is increased to introduce stronger uncertainty representation capabilities and improve trajectory modal diversity; in simple scenarios, the noise intensity is reduced to avoid over-divergent prediction, thereby achieving adaptive adjustment of trajectory modal diversity. This mechanism enhances the model's ability to represent uncertainties in different scenarios and achieves an adaptive balance between diversity and accuracy in predicted trajectories.

[0041] Example 2 This embodiment provides an electronic device, including: one or more processors; a memory; and one or more programs stored in the memory, the one or more programs including instructions for executing the modally controllable vehicle trajectory prediction method based on dynamic noise mechanism as described in Embodiment 1.

[0042] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0043] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0044] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0046] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0047] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0048] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A modally controllable vehicle trajectory prediction method based on dynamic noise mechanism, characterized in that, Includes the following steps: The system acquires traffic scenes and performs semantic segmentation to obtain different classification scenes. It then uses a trained GAN-based multimodal trajectory prediction model to predict vehicle trajectories. Simultaneously, it dynamically injects optimal noise based on the complexity of different classification scenes to ultimately predict vehicle trajectories. The GAN-based multimodal trajectory prediction model includes a generator G and a discriminator D. The generator G and discriminator D are trained adversarially using a nearest neighbor sampling mechanism to obtain the trained GAN-based multimodal trajectory prediction model.

2. The modally controllable vehicle trajectory prediction method based on dynamic noise mechanism according to claim 1, characterized in that, Both the generator G and the discriminator D are composed of a context encoder and a trajectory decoder.

3. The modally controllable vehicle trajectory prediction method based on dynamic noise mechanism according to claim 2, characterized in that, The execution process of the generator G includes: The context encoder is used to extract historical trajectories and map features to generate a fixed-dimensional latent representation. , where the potential representation Represented as: In the formula, For context encoder, For historical trajectory and map data, For the real number field, For dimensions; During each forward pass, a noise vector is sampled. and add to the potential representation The underlying representation obtained after perturbation , represented as: The perturbated latent representation The generated trajectory is passed to the trajectory decoder to obtain the generated trajectory. ,in, For trajectory decoders.

4. The modally controllable vehicle trajectory prediction method based on dynamic noise mechanism according to claim 3, characterized in that, The context encoder is a graph neural network.

5. The modally controllable vehicle trajectory prediction method based on dynamic noise mechanism according to claim 1, characterized in that, The loss function of the generator G is: in: In the formula, Let G be the total loss function of the generator. To supervise the loss, used to measure the predicted trajectory With the actual trajectory The differences between them For hyperparameters, To combat losses, and to control the reliability of the generated trajectory, The time step length for predicting the trajectory, , The predicted and actual trajectories at time steps are respectively. The position vector, For historical trajectory and map data, The closest trajectory to the real trajectory generated by sampling the generator G. A generated trajectory is called the optimal generated trajectory. As expected, For the actual trajectory, ~pdata This represents a sample taken from the actual trajectory. z Let z~pz represent noise, which is sampled from a Gaussian noise distribution. Indicates in X Under the constraints, discriminator D determines the true trajectory Y The rating, Indicates in X Under the constraints, discriminator D determines the optimal generated trajectory. The rating.

6. The modally controllable vehicle trajectory prediction method based on dynamic noise mechanism according to claim 1, characterized in that, The steps of adversarial training of the generator G and discriminator D include: In each trajectory prediction process, the generator G is used to sample and generate the trajectory at one time. Generate trajectory ,in, For historical trajectory and map data, For Gaussian noise variance control; In the Select the generated trajectory that matches the real trajectory exist The closest generated trajectory under the distance metric is taken as the best generated trajectory. , represented as: In the formula, For the first Generate trajectory With the actual trajectory Between distance; Discriminator D undergoes adversarial training under the optimization objective to distinguish the real trajectory. and optimal generated trajectory ; The adversarial training process is completed by iteratively training the generator G and the discriminator D through forward and backward propagation.

7. The modally controllable vehicle trajectory prediction method based on dynamic noise mechanism according to claim 6, characterized in that, The optimization objective is expressed as: In the formula, To optimize the objective, ~pdata This represents a sample taken from the actual trajectory. For the actual trajectory, As expected, Indicates in X Under the constraints, discriminator D determines the true trajectory Y The rating, The trajectory that most closely resembles the true trajectory generated by the generator G is called the optimal generated trajectory. Indicates in X Under the constraints, discriminator D determines the optimal generated trajectory. The rating, z Let z be noise, and z~pz represent the noise input sampled from a Gaussian noise distribution.

8. The modally controllable vehicle trajectory prediction method based on dynamic noise mechanism according to claim 1, characterized in that, Different semantic segments are made based on scene semantic features. If the semantic features of the scene are a map structure, then the traffic scene is divided into intersection areas and non-intersection areas; If the semantic feature of the scene is a road attribute, then the traffic scene is divided into motor vehicle lanes and non-motor vehicle lanes; If the semantic feature of the scene is weather information, then the traffic scene will be divided into severe weather scene and non-severe weather scene.

9. An electronic device, characterized in that, include: One or more processors; Memory; and One or more programs stored in memory, the one or more programs including instructions for executing the modally controllable vehicle trajectory prediction method based on dynamic noise mechanism as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It includes one or more programs that are executed by one or more processors of an electronic device, the one or more programs including instructions for executing the modally controllable vehicle trajectory prediction method based on dynamic noise mechanism as described in any one of claims 1-8.

Images