Trajectory prediction method, storage medium, intelligent device and vehicle

Abstract

The application relates to the technical field of intelligent driving of vehicles, in particular to a trajectory prediction method, a storage medium, an intelligent device and a vehicle, and aims to solve the technical problem of how to improve the long-time sequence trajectory understanding capability of a trajectory planning model and improve the prediction trajectory stability. To this end, the application comprises: obtaining a target prediction trajectory through a first trajectory prediction model, wherein the first trajectory prediction model is obtained by cutting a second trajectory prediction model that has been trained, and the number of trajectory points output by the second trajectory prediction model is greater than the number of trajectory points output by the first trajectory prediction model. Through the asymmetric mode of long-time sequence trajectory guiding the training of the trajectory prediction model and short-time sequence trajectory performing real-time inference prediction, the prediction ability of the trajectory prediction model for future long-time sequence environment evolution can be improved without increasing the inference side calculation burden of the vehicle, the global view of the trajectory planning model is improved, and the predicted trajectory is more stable and smooth.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology for vehicles, specifically to a trajectory prediction method, storage medium, intelligent device, and vehicle. Background Technology

[0002] As the construction of intelligent driving technology shifts from "modular" to "end-to-end", end-to-end models, through massive amounts of data, are able to learn the complex decision-making logic of human drivers. They have shown a stronger upper limit when dealing with unstructured roads or complex game scenarios and have been widely applied to vehicle intelligent driving systems.

[0003] Despite the immense potential of end-to-end models, their deployment in real-world vehicles often faces several challenges. For instance, limited by onboard chip computing resources, trajectory planning models often only predict short timeframes (e.g., 2-3 seconds) to meet the real-time requirements of intelligent driving systems, resulting in a lack of global perspective. Furthermore, in complex scenarios, even minute noise fluctuations in the immediate input can lead to unstable driving behavior from the trajectory planning model, impacting driving safety. Therefore, improving the long-term trajectory understanding capabilities of trajectory planning models and enhancing the stability of predicted trajectories has become a pressing issue.

[0004] Accordingly, a new trajectory prediction scheme is needed in this field to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned deficiencies, this application is proposed to solve, or at least partially solve, the technical problem of how to improve the understanding of long-term trajectories of trajectory planning models and improve the stability of predicted trajectories.

[0006] In a first aspect, a trajectory prediction method is provided, the method comprising:

[0007] Acquire environmental perception data;

[0008] Based on the environmental perception data and the target's historical trajectory, the target's predicted trajectory is obtained through the first trajectory prediction model;

[0009] The first trajectory prediction model is obtained by trimming the already trained second trajectory prediction model;

[0010] The target predicted trajectory includes a first number of trajectory points;

[0011] The target prediction training trajectory output by the second trajectory prediction model includes a second number of trajectory points, which is greater than the first number;

[0012] The target historical trajectory includes a third number of trajectory points.

[0013] In one technical solution of the above trajectory prediction method, the method for obtaining the trained second trajectory prediction model includes:

[0014] Based on the first sample set, supervised training is performed on the second trajectory prediction model to obtain a pre-trained second trajectory prediction model.

[0015] Based on the PPO algorithm architecture, using the second sample set, offline reinforcement learning is performed on the pre-trained second trajectory prediction model to obtain the trained second trajectory prediction model.

[0016] In one technical solution of the above trajectory prediction method, the PPO algorithm architecture includes:

[0017] A strategy model, which is constructed based on the pre-trained second trajectory prediction model, is used to generate target prediction training trajectories;

[0018] A value model, which is constructed based on the pre-trained second trajectory prediction model, is used to obtain the value score of the target predicted training trajectory;

[0019] A reward model, used to obtain a reward score for the predicted training trajectory of the target, includes:

[0020] Obtain violation information from the target prediction training trajectory, the violation information including violation points and the temporal positions corresponding to the violation points;

[0021] Based on the violation information, the reward score is determined, wherein all reward scores are less than or equal to 0.

[0022] In one technical solution of the above trajectory prediction method, determining the reward score based on the violation information includes:

[0023] Obtain the acceptable length of the target prediction training trajectory, wherein the acceptable length is the maximum number of consecutive non-violation points in the target prediction training trajectory;

[0024] The reward score is determined based on the length of the target achieved, wherein the larger the value of the target length, the smaller the absolute value of the reward score.

[0025] In one technical solution of the above trajectory prediction method, determining the reward score based on the violation information further includes:

[0026] Obtain the number of violation points in the target prediction training trajectory;

[0027] When the length of compliance is the same, the greater the number of violations, the greater the absolute value of the reward score.

[0028] In one technical solution of the above trajectory prediction method, determining the reward score based on the violation information further includes:

[0029] When the length of the target prediction training trajectory and the number of violation points are the same, the closer the temporal position of the first violation point is to the starting temporal position of the target prediction training trajectory, the larger the absolute value of the reward score.

[0030] In one technical solution of the above trajectory prediction method, the PPO algorithm architecture further includes:

[0031] The GAE calculation module calculates the GAE advantage based on the value score and the reward score to constrain the parameter updates of the strategy model.

[0032] In a second aspect, a storage medium is provided, the storage medium being adapted to store a plurality of program codes, the program codes being adapted to be loaded and run by a processor to implement the trajectory prediction method described in any of the above technical solutions when executed.

[0033] In a third aspect, a smart device is provided, comprising:

[0034] At least one processor;

[0035] And, a memory communicatively connected to the at least one processor;

[0036] The memory stores a computer program, which, when executed by the at least one processor, implements the trajectory prediction method described in any of the above technical solutions.

[0037] In a fourth aspect, a vehicle is provided, the vehicle including the intelligent device described in the above-described technical solution.

[0038] The above-mentioned technical solutions of this application have at least one or more of the following beneficial effects: by guiding the training of the trajectory prediction model through long-term trajectory and performing real-time inference prediction of short-term trajectory in an asymmetric mode, the trajectory prediction model can improve its ability to predict the future long-term environmental evolution without increasing the computational burden on the vehicle inference side, improve the global vision of the trajectory planning model, and make the predicted trajectory more stable and smooth. Attached Figure Description

[0039] The disclosure of this application will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0040] Figure 1This is a schematic flowchart of the main steps of a trajectory prediction method according to an embodiment of this application.

[0041] Figure 2 This is a schematic flowchart of the main steps of a method for obtaining a trained second trajectory prediction model according to an embodiment of this application.

[0042] Figure 3 This is a schematic diagram of the PPO algorithm architecture according to an embodiment of this application.

[0043] Figure 4 This is a schematic diagram of the main structure of a smart device according to an embodiment of this application. Detailed Implementation

[0044] Some embodiments of this application are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of this application and are not intended to limit the scope of protection of this application.

[0045] In the description of this application, "module" and "processor" can include hardware, software, or a combination of both. A module can include hardware circuitry, various suitable sensors, communication ports, memory, and may also include software components, such as program code, or a combination of software and hardware. A processor can be a central processing unit, microprocessor, image processor, digital signal processor, or any other suitable processor. The processor has data and / or signal processing capabilities. The processor can be implemented in software, in hardware, or a combination of both. Computer-readable storage media includes any suitable medium capable of storing program code, such as magnetic disks, hard disks, optical disks, flash memory, read-only memory, random access memory, etc. The term "A and / or B" means all possible combinations of A and B, such as only A, only B, or A and B. The terms "at least one A or B" or "at least one of A and B" have a similar meaning to "A and / or B" and can include only A, only B, or A and B. The singular terms "a" or "this" can also include plural forms.

[0046] First read Figure 1 , Figure 1 This is a schematic flowchart illustrating the main steps of a trajectory prediction method according to an embodiment of this application. The trajectory prediction method in this embodiment includes:

[0047] Step S101: Acquire environmental perception data;

[0048] Step S102: Based on environmental perception data and target historical trajectory, obtain target predicted trajectory through first trajectory prediction model, wherein the first trajectory prediction model is obtained by pruning the trained second trajectory prediction model, the target predicted trajectory includes a first number of trajectory points, and the target predicted training trajectory output by the second trajectory prediction model includes a second number of trajectory points, the second number being greater than the first number.

[0049] In intelligent driving applications for vehicles, the environmental perception data in step 101 can be obtained through one or more of the following methods, such as obtaining information about roads, obstacles, etc. around the vehicle through onboard image sensors and / or lidar sensors, inputting target location information through voice recognition or onboard operation panel, or inputting high-precision map data from the vehicle controller.

[0050] In step S102, the target historical trajectory refers to the vehicle time sequence trajectory (trajectory points arranged in chronological order) before the current prediction time. The target historical trajectory includes a third number of trajectory points. The third number can be greater than or equal to the second number, or less than or equal to the first number, or greater than the first number and less than the second number.

[0051] As an example, the first quantity is set to 2; the second quantity is greater than the first quantity and is set to 5; the third quantity is set to 3. The temporal position of the trajectory point corresponding to the current prediction time is 10. For the first trajectory prediction model, the target predicted trajectory includes 2 trajectory points, and the temporal positions of these two trajectory points are 11 and 12, respectively. For the second trajectory prediction model, the target predicted training trajectory includes 5 trajectory points, and the temporal positions of these five trajectory points are 11, 12, 13, 14 and 15, respectively. The target historical trajectory includes 3 historical trajectory points adjacent to the current trajectory point (temporal position 10), and the temporal positions of these 3 historical trajectory points are 7, 8 and 9, respectively.

[0052] By training the trajectory prediction model using the aforementioned long-time trajectory guidance, the asymmetric mode of real-time inference prediction of short-time trajectory can improve the trajectory prediction model's ability to predict future long-time environmental evolution without increasing the computational burden on the vehicle's inference side, enhance the global perspective of the trajectory planning model, and make the predicted time-series trajectory more stable and smooth.

[0053] It should be noted that this application does not limit the specific structure of the first trajectory prediction model (second trajectory prediction model). As an example, a structure of Transformer model + diffusion model can be adopted. The Transformer model obtains the high-dimensional feature vector of environmental perception data and target historical trajectory, and the diffusion model outputs the predicted vehicle trajectory based on the high-dimensional feature vector. Alternatively, a structure of GNN model + diffusion model can be adopted, etc.

[0054] Next, read the appendix. Figure 2 , Figure 2 This is a schematic flowchart illustrating the main steps of a method for obtaining a trained second trajectory prediction model according to an embodiment of this application. The training method for the second trajectory prediction model in this embodiment includes:

[0055] Step S201: Based on the first sample set, perform supervised training on the second trajectory prediction model to obtain a pre-trained second trajectory prediction model;

[0056] Step S202: Based on the PPO (Proximal Policy Optimization) algorithm architecture, use the second sample set to perform offline reinforcement learning on the pre-trained second trajectory prediction model to obtain the trained second trajectory prediction model.

[0057] In step S201, the first sample in the first sample set can be either a time-series trajectory of a human-driven vehicle or a time-series trajectory of a vehicle driven intelligently. Taking automatic parking as an example, each first sample includes a complete parking time-series trajectory, including target location information (parking space location), trajectory point sequence, and environmental perception data sequence corresponding to each trajectory point (sample trajectory point). The sample trajectory points and target location information serve as labels for supervised training.

[0058] During supervised training, the difference between the trajectory points predicted by the second trajectory prediction model and the sample trajectory points is compared. The loss function (such as mean squared error, cross entropy, etc.) is used to guide the model training, adjust the model parameters, and gradually optimize the trajectory prediction performance to obtain the pre-trained second trajectory prediction model.

[0059] Continue reading the appendix Figure 3 Combined with the appendix Figure 3 Explain the specific implementation method of step S202. Figure 3 This is a schematic diagram of the PPO algorithm architecture according to an embodiment of this application.

[0060] like Figure 3 As shown, the PPO algorithm architecture mainly includes: a policy model (Actor Model), a value model (Critic Model), a reward model (Reward Model), and a GAE (Generalized Advantage Estimation) calculation module.

[0061] Both the strategy model and the value model are constructed based on the pre-trained second trajectory prediction model obtained in step S201. The strategy model is used to output the target predicted training trajectory (the second number of trajectory points), and the value model is used to calculate the value score V (i.e., expected cumulative reward) of the target predicted training trajectory and its corresponding environmental perception data (from the second sample set). The reward model is used to calculate the reward score r of the target predicted training trajectory output at each time step. The GAE calculation module temporarily stores the value score V and reward score r obtained from the trajectory prediction at each time step and calculates the GAE advantage according to the preset model parameter update cycle.

[0062] Continuing with the example of automatic parking, we will explain how to obtain the reward score r and GAE. The second sample in the second sample set can come from the first sample set or be a sample different from the first sample set. Unlike the first sample set, the second sample set does not include the trajectory point sequence representing the parking time-series trajectory, but it does include the environmental perception data sequence corresponding to the time series of the parking process; that is, the time-series position of the trajectory point is the sequence position of the environmental perception data.

[0063] The environmental perception data, target location information, and historical target prediction training trajectories from the second sample are input into the policy model and the value model, respectively. The target prediction training trajectory is obtained through the policy model, and the value score V corresponding to the target prediction training trajectory at each time step is obtained through the value network. The target prediction training trajectory is then input into the reward model to obtain the reward score r of the target prediction training trajectory.

[0064] In the reward model, the violation information in the target prediction training trajectory is first judged, that is, whether there are violation points in the target prediction training trajectory (i.e., violation points). If there are violation points, the temporal position of all violation points is recorded, and then the reward score is determined based on the violation information.

[0065] Specifically, the following methods can be used to determine if a trajectory point is a violation point, including:

[0066] The trajectory point violates traffic rules, such as the trajectory point exceeding the road boundary, the trajectory point crossing the line, or the trajectory point entering a restricted area;

[0067] The distance between the trajectory point and the obstacle is less than the preset safe distance threshold. The obstacle includes other vehicles, pedestrians, guardrails, etc.

[0068] The driving behavior of the trajectory point exhibits unstable characteristics, such as sudden changes in the steering wheel angle or drastic fluctuations in vehicle speed.

[0069] The motion state of the trajectory point conflicts with environmental constraints, such as the appearance of a velocity component towards the rear of the vehicle in a stationary parking space.

[0070] In this application, the reward score is designed to be less than or equal to 0. That is, when there is no violation point in the target prediction training trajectory, the reward score is 0, and as long as there is a violation point in the target prediction training trajectory, the reward score is negative (a penalty is imposed).

[0071] This design, where all reward scores are less than or equal to 0, aligns with the logic of the real world (no violations are expected, violations are unacceptable and should be punished) and enables the pre-trained second trajectory prediction model to more quickly distinguish between superior and inferior trajectories, reduce ineffective exploration, and accelerate the convergence speed of the model in reinforcement learning in complex or high-dimensional state spaces.

[0072] In one embodiment, the reward score can be determined based on the qualifying length of the target predicted training trajectory, where the qualifying length is the maximum number of consecutive non-violation points in the target predicted training trajectory. The larger the qualifying length, the smaller the absolute value of the reward score (a negative reward score, closer to 0, results in a smaller penalty).

[0073] As an example, the temporal positions corresponding to the 5 trajectory points in the target prediction training trajectory A are 11, 12, 13, 14 and 15. When only the trajectory point corresponding to the temporal position 14 is a violation point, the qualified length of the target prediction training trajectory A is 3.

[0074] The temporal positions corresponding to the 5 trajectory points in the target prediction training trajectory B are: 16, 17, 18, 19 and 20. When only the trajectory point corresponding to the temporal position 20 is a violation point, the qualified length of the target prediction training trajectory B is 4.

[0075] At this point, the target length of the predicted training trajectory B is greater than that of the predicted training trajectory A. Therefore, the absolute value of the reward score of the predicted training trajectory B is less than that of the predicted training trajectory A, meaning that the penalty of the predicted training trajectory B will be smaller.

[0076] In one embodiment, the reward score can be determined based on both the acceptable length of the target predicted training trajectory and the number of violation points. When the acceptable length is the same, the more violation points there are, the greater the absolute value of the reward score (the higher the penalty).

[0077] As an example, the temporal positions corresponding to the 5 trajectory points in the target prediction training trajectory A are 11, 12, 13, 14 and 15. When only the trajectory point corresponding to the temporal position 14 is a violation point, the qualified length of the target prediction training trajectory A is 3 and the number of violation points is 1.

[0078] The temporal positions corresponding to the 5 trajectory points in the target prediction training trajectory B are: 16, 17, 18, 19 and 20. When the trajectory points corresponding to temporal positions 19 and 20 are both violation points, the qualified length of the target prediction training trajectory B is 3 and the number of violation points is 2.

[0079] At this point, the target predicted training trajectory A and the target predicted training trajectory B have the same length of compliance. The number of violation points in the target predicted training trajectory B is greater than the number of violation points in the target predicted training trajectory A. Therefore, the absolute value of the reward score of the target predicted training trajectory B will be greater than the absolute value of the reward score of the target predicted training trajectory A, that is, the penalty of the target predicted training trajectory B will be greater.

[0080] In one embodiment, the reward score can be determined based on the length of the target prediction training trajectory, the number of violations, and the temporal position of the smallest violation point. When the length of the target prediction training trajectory and the number of violations are the same, the closer the temporal position of the first violation point is to the starting position of the target prediction training trajectory, the greater the absolute value of the reward score.

[0081] It should be noted that for any two time positions, the trajectory point corresponding to the time position with the smaller value is the trajectory point with the earlier time. The starting position of the target prediction training trajectory is the trajectory point corresponding to the time position with the smallest value among the 5 trajectory points it includes.

[0082] As an example, the temporal positions corresponding to the 5 trajectory points in the target prediction training trajectory A are 11, 12, 13, 14 and 15. When only the trajectory point corresponding to the temporal position 14 is a violation point, the qualified length of the target prediction training trajectory A is 3 and the number of violation points is 1.

[0083] The temporal positions corresponding to the 5 trajectory points in the target prediction training trajectory B are: 16, 17, 18, 19 and 20. When only the trajectory point corresponding to the temporal position 17 is a violation point, the qualified length of the target prediction training trajectory B is also 3, and the number of violation points is also 1.

[0084] At this point, the target predicted training trajectory A and target predicted training trajectory B have the same passing length and the same number of violation points. The first violation point of target predicted training trajectory B is closer to the starting position of the trajectory. To identify and resolve problems as early as possible, the absolute value of the reward score for target predicted training trajectory B will be greater than the absolute value of the reward score for target predicted training trajectory A; that is, the penalty for target predicted training trajectory B will be greater.

[0085] This application designs a reward system based on the target prediction training trajectory's qualifying length, the number of violation points, and the negative score of the violation point's temporal location. This system enables the model to distinguish between superior and inferior temporal trajectories more quickly, reduces ineffective exploration, accelerates the convergence speed of the trajectory prediction model in reinforcement learning in complex or high-dimensional state spaces, and improves the efficiency and effectiveness of model training.

[0086] After obtaining the value score V and the reward score r, the GAE advantage can be calculated. After obtaining the GAE advantage, the target reward and PPO loss (including policy loss and value loss) are calculated. The policy model and value model are updated through backpropagation until the policy loss and value loss converge, and the offline reinforcement learning ends, resulting in the trained second trajectory prediction model.

[0087] The output layer of the trained second trajectory prediction model is pruned to reduce its dimension, so that the number of trajectory points in the output predicted trajectory is reduced from 5 (second number) to 2 (first number), thus obtaining a first trajectory prediction model that can be deployed on vehicles.

[0088] It should be noted that the number of trajectory points included in the historical target prediction training trajectory during trajectory prediction model training should be the same as the number of trajectory points included in the target historical trajectory during actual vehicle deployment, i.e., both should be the same number.

[0089] The method described in this application predicts and outputs five trajectory points at once (the second quantity) during the training of the second trajectory prediction model. This not only improves the utilization rate of the samples but also helps to mine the contextual information of long-term trajectories, thereby better guiding the prediction of individual trajectory points.

[0090] When the vehicle is deployed, the first trajectory prediction model only predicts and outputs 2 trajectory points (the first quantity), and then iteratively completes the prediction of all trajectories. In this way, the prediction and inference time per time step is not increased. Moreover, since the first trajectory prediction model inherits the training constraint of 5 trajectory points (the second quantity), the predicted time-series trajectory will be more stable and smooth, and less susceptible to the influence of instantaneous noise.

[0091] Another aspect of this application provides a storage medium that can be configured to store a program for executing the trajectory prediction method of the above-described method embodiments. This program can be loaded and run by a processor to implement the trajectory prediction method of the above-described embodiments. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The storage medium can be a storage device comprising various computer devices. Optionally, in the embodiments of this application, the storage medium is a non-transitory read / write storage medium.

[0092] Another aspect of this application provides a smart device. In one embodiment of a smart device according to this application, the smart device may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, which, when executed by the at least one processor, implements the trajectory prediction method described in any of the above embodiments. The smart device described in this application may include driving equipment, smart vehicles, robots, and other devices. See Appendix Figure 4 , Figure 4 The example illustrates a smart device 4 including a memory 41 and a processor 42 connected via a bus.

[0093] In some embodiments of this application, the smart device 4 may further include at least one sensor for sensing information. For example, the sensor may include an image sensor, a lidar sensor, etc. The sensor is communicatively connected to any type of processor mentioned in this application. Optionally, the trajectory prediction method may further include an intelligent driving system for guiding the smart device to drive autonomously or providing driving assistance. The processor communicates with the sensor and / or the intelligent driving system to perform the trajectory prediction method described in any of the above embodiments.

[0094] Another aspect of this application provides a vehicle. The vehicle includes the intelligent device described in the above embodiments. As an example, the vehicle is a new energy vehicle.

[0095] It should be noted that although the steps in the above embodiments are described in a specific order, those skilled in the art will understand that in order to achieve the effect of this application, different steps do not necessarily have to be executed in such an order. They can be executed simultaneously (in parallel) or in other orders. These adjusted solutions are equivalent to the technical solutions described in this application and therefore will also fall within the protection scope of this application.

[0096] Those skilled in the art will understand that all or part of the processes in the method of the above-described embodiment can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0097] The technical solution of this application has been described above with reference to one embodiment shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A trajectory prediction method, characterized in that, The method includes: Acquire environmental perception data; Based on the environmental perception data and the target's historical trajectory, the target's predicted trajectory is obtained through the first trajectory prediction model; The first trajectory prediction model is obtained by pruning the trained second trajectory prediction model. The target predicted trajectory includes a first number of trajectory points. The target predicted training trajectory output by the second trajectory prediction model includes a second number of trajectory points, which is greater than the first number. The target historical trajectory includes a third number of trajectory points. The method for obtaining the trained second trajectory prediction model includes: Based on the first sample set, supervised training is performed on the second trajectory prediction model to obtain a pre-trained second trajectory prediction model. Based on the PPO algorithm architecture, offline reinforcement learning is performed on the pre-trained second trajectory prediction model using a second sample set to obtain the trained second trajectory prediction model. The PPO algorithm architecture includes: A strategy model, which is constructed based on the pre-trained second trajectory prediction model, is used to generate target prediction training trajectories; A value model, which is constructed based on the pre-trained second trajectory prediction model, is used to obtain the value score of the target predicted training trajectory; A reward model, used to obtain a reward score for the predicted training trajectory of the target, includes: Obtain violation information from the target prediction training trajectory, the violation information including violation points and the temporal positions corresponding to the violation points; Based on the violation information, the reward score is determined, wherein all reward scores are less than or equal to 0.

2. The trajectory prediction method according to claim 1, characterized in that, The process of determining the reward score based on the violation information includes: Obtain the acceptable length of the target prediction training trajectory, wherein the acceptable length is the maximum number of consecutive non-violation points in the target prediction training trajectory; The reward score is determined based on the length of the target achieved, wherein the larger the value of the target length, the smaller the absolute value of the reward score.

3. The trajectory prediction method according to claim 2, characterized in that, The step of determining the reward score based on the violation information also includes: Obtain the number of violation points in the target prediction training trajectory; When the length of compliance is the same, the greater the number of violations, the greater the absolute value of the reward score.

4. The trajectory prediction method according to claim 3, characterized in that, The step of determining the reward score based on the violation information also includes: When the number of qualifying lengths and the number of violation points in the target prediction training trajectory are the same, the closer the temporal position of the first violation point is to the starting temporal position of the target prediction training trajectory, the larger the absolute value of the reward score.

5. The trajectory prediction method according to claim 1, characterized in that, The PPO algorithm architecture also includes: The GAE calculation module calculates the GAE advantage based on the value score and the reward score to constrain the parameter updates of the strategy model.

6. A storage medium suitable for storing multiple lines of program code, characterized in that, The program code is adapted to be loaded and run by a processor to perform the trajectory prediction method according to any one of claims 1 to 5.

7. A smart device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores a computer program, which, when executed by the at least one processor, implements the trajectory prediction method according to any one of claims 1 to 5.

8. A vehicle, characterized in that, The vehicle includes the intelligent device as described in claim 7.

Images