Trajectory prediction method, trajectory prediction device and processor of intelligent agent

Abstract

The application provides a trajectory prediction method, a trajectory prediction device and a processor for an agent. The trajectory prediction method comprises: constructing a preset network based on at least a first encoder and a second encoder, the first encoder being used for modeling cross-space-time one-way interaction of a target agent in a first region, and the second encoder being used for modeling cross-space-time two-way interaction of the target agent in a second region; training the preset network by using a target data set to obtain a target network, the target data set comprising a plurality of training data and target results corresponding to the training data; and predicting a trajectory of a real-time target agent by using at least the target network and real-time perception information, the real-time perception information being environment information and agent information in a target region perceived by an autonomous vehicle during driving, so as to solve the problems of large information loss and large calculation amount of interaction modeling between agents in traffic in the existing trajectory prediction technology.

Description

Technical Field

[0001] This application relates to the field of autonomous driving, and more specifically, to a trajectory prediction method, trajectory prediction device, computer-readable storage medium, and processor for an intelligent agent. Background Technology

[0002] For autonomous vehicles, the main components include five core modules: perception, localization, trajectory prediction, decision-making and planning, and control. Trajectory prediction for autonomous vehicles primarily relies on historical state information, contextual semantic information, and high-precision maps input from the perception and localization modules to establish a complex interaction model for highly dynamic traffic scenarios, predicting the target agent's trajectory in the next few seconds. The predicted trajectory is then output to the autonomous vehicle's decision-making and planning module for trajectory planning and decision-making.

[0003] Predicting the future trajectory of a target agent can be divided into target agent selection, agent sequence interaction modeling, and future trajectory decoding. Specifically, agent sequence interaction modeling involves encoding the interaction features of the selected target agents, and future trajectory decoding is based on decoding these encoded interaction features. In highly dynamic scenarios, complex interactions exist between target agents and between target agents and the environment, which affect the accuracy of the target agents' future trajectories and thus the performance of autonomous vehicles in predicting the trajectories of other traffic participants. Summary of the Invention

[0004] The main objective of this application is to provide a trajectory prediction method, trajectory prediction device, computer-readable storage medium, and processor for intelligent agents, so as to at least solve the problems of information loss and large computational load in existing trajectory prediction technologies for modeling interactions between intelligent agents in traffic.

[0005] To achieve the above objectives, according to one aspect of this application, a trajectory prediction method for an intelligent agent is provided, comprising: constructing a preset network based at least on a first encoder and a second encoder, wherein the first encoder is used to perform cross-spatial-space unidirectional interaction modeling of a target intelligent agent in a first region, and the second encoder is used to perform cross-spatial-space bidirectional interaction modeling of the target intelligent agent in a second region, wherein the first region is a portion of the second region, and the preset network is used to predict the trajectory of the target intelligent agent; training the preset network using a target dataset to obtain a target network, wherein the target dataset includes multiple sets of training data and target results corresponding to each set of training data; and predicting the trajectory of a real-time target intelligent agent using at least the target network and real-time perception information, wherein the real-time perception information is environmental information and intelligent agent information in the target region perceived by an autonomous vehicle during driving.

[0006] Optionally, the process of the first encoder performing cross-spatial-space unidirectional interaction modeling of the target agents in the first region includes: an evaluation step, which uses a priority evaluator to evaluate the priorities of multiple target agents in the first region to obtain the priority order between any two target agents; a modeling step, which performs cross-spatial-space unidirectional interaction modeling of the corresponding two target agents based on the priority order between any two target agents; and a repetition step, which repeats the evaluation step and the modeling step at least once in sequence until cross-spatial-space unidirectional interaction modeling is performed for the target agents in the first region at all target times.

[0007] Optionally, after performing cross-spatial one-way interactive modeling on the target agents in the first region at all target times, the trajectory prediction method further includes: performing cross-spatial one-way interactive modeling on the multiple target agents at multiple target times with the high-precision map respectively.

[0008] Optionally, the process of the second encoder performing cross-temporal bidirectional interactive modeling of the target intelligent agents in the second region includes: performing feature information modeling for each target intelligent agent in the second region, wherein the feature information includes at least velocity information, direction information, position information and category information; and performing cross-temporal bidirectional interactive modeling for any two target intelligent agents in the second region at each target time.

[0009] Optionally, the process of determining the priority estimator involves: determining the priority order of the target agents in each predetermined area based on predetermined rules and a high-precision map, thereby obtaining a priority training set; and training a preset priority estimator based on the priority training set to obtain the priority estimator, wherein the preset priority estimator is constructed based on a multilayer perceptron.

[0010] Optionally, the process of determining the target intelligent agent in the second region includes: acquiring target perception information of the autonomous vehicle, wherein the target perception information is environmental information and intelligent agent information of the second region acquired by the target device, wherein the target device includes at least one of the following: image acquisition device, radar device, and the intelligent agent information is information of multiple intelligent agents in the second region; using a target intelligent agent selection algorithm and the target perception information to filter the intelligent agents in the second region to obtain the target intelligent agent, wherein the target intelligent agent selection algorithm is obtained based on neural network training, and the target intelligent agent is a person or vehicle that interacts with the autonomous vehicle within the perception range of the autonomous vehicle.

[0011] Optionally, a preset network is constructed based at least on the first encoder and the second encoder, including: constructing the preset network based on the first encoder, the second encoder and the target decoder, wherein the target decoder is used to predict the trajectory of the target agent based on the target interaction information output by the first encoder and the second encoder.

[0012] According to another aspect of this application, a trajectory prediction device for an intelligent agent is provided, comprising: a construction unit, configured to construct a preset network based at least on a first encoder and a second encoder, wherein the first encoder is configured to perform cross-spatial-space unidirectional interaction modeling of a target intelligent agent in a first region, and the second encoder is configured to perform cross-spatial-space bidirectional interaction modeling of the target intelligent agent in a second region, wherein the first region is a portion of the second region, and the preset network is configured to predict the trajectory of the target intelligent agent; a training unit, configured to train the preset network using a target dataset to obtain a target network, wherein the target dataset includes multiple sets of training data and target results corresponding to each set of training data; and a prediction unit, configured to predict the trajectory of a real-time target intelligent agent using at least the target network and real-time perception information, wherein the real-time perception information is environmental information and intelligent agent information perceived by an autonomous vehicle in the target region during driving.

[0013] According to another aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform any of the described intelligent agent trajectory prediction methods.

[0014] According to another aspect of this application, a processor is provided for running a program, wherein the program executes any of the described trajectory prediction methods for an intelligent agent.

[0015] Applying the technical solution of this application, the trajectory prediction method firstly constructs a preset network having at least a first encoder and a second encoder, wherein the first encoder is used to perform cross-temporal unidirectional interaction modeling of the target agent in the first region, and the second encoder is used to perform cross-temporal bidirectional interaction modeling of the target agent in the second target region; then, the preset network is trained using the target dataset to obtain the target network; finally, at least the trained target network and real-time perception information are used to predict the trajectory of the corresponding real-time target agent. Compared with existing technologies that perform cross-spatial bidirectional interactive modeling of target agents in the first region, this application uses a first encoder to perform cross-spatial unidirectional interactive modeling of target agents in the first region (i.e., cross-spatial unidirectional interactive modeling of target agents in a local region) and a second encoder to perform cross-spatial bidirectional interactive modeling of target agents in the second region (i.e., cross-spatial bidirectional interactive modeling of target agents in the entire region), instead of performing interactive modeling of all agents in the first and second regions. This results in a smaller overall computational load and less information loss during the interactive modeling process, ensuring higher accuracy in trajectory prediction of target agents (and real-time target agents in practical applications). This solves the problem of large information loss and computational load in existing trajectory prediction technologies for interactive modeling between agents in traffic, thereby ensuring that the autonomous vehicle's subsequent trajectory planning based on the predicted real-time target agent trajectory is more reasonable and accurate, ensuring higher safety and intelligence of the autonomous vehicle. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 A hardware structure block diagram of a mobile terminal for performing a trajectory prediction method for an autonomous vehicle, provided in an embodiment of this application, is shown.

[0018] Figure 2 A schematic flowchart of a trajectory prediction method for an autonomous vehicle provided by an embodiment of this application is shown.

[0019] Figure 3 This illustration shows a schematic diagram of a cross-spatial one-way interaction modeling of a target agent in a first region, provided by an embodiment of this application.

[0020] Figure 4 This illustration shows a schematic diagram of a spatiotemporal bidirectional interactive modeling of a target agent in a second region, provided by an embodiment of this application.

[0021] Figure 5 A schematic diagram illustrating the determination of a predetermined rule provided by an embodiment of this application is shown;

[0022] Figure 6 A schematic diagram of a training priority evaluator provided by an embodiment of this application is shown;

[0023] Figure 7 A flowchart of another trajectory prediction method for autonomous vehicles provided by an embodiment of this application is shown;

[0024] Figure 8 A schematic diagram of the structure of a trajectory prediction device for an autonomous vehicle provided in an embodiment of this application is shown.

[0025] The above figures include the following reference numerals:

[0026] 2011, First target intelligent agent; 2012, Second target intelligent agent; 2013, Third target intelligent agent; 201, First priority; 202, Second priority; 2021, First other intelligent agent; 2022, Second other intelligent agent; 2023, Third other intelligent agent; 203, Fifth target intelligent agent; 204, Fourth target intelligent agent; 205, Designated point; 102, Processor; 104, Memory; 106, Transmission device; 108, Input / output device. Detailed Implementation

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] As described in the background section, existing trajectory prediction technologies suffer from significant information loss and computational burden in modeling interactions between intelligent agents in traffic. To address these issues, embodiments of this application provide a trajectory prediction method, a trajectory prediction device, a computer-readable storage medium, and a processor for intelligent agents.

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0032] The methods and embodiments provided in this application can be executed on a mobile terminal, a computer terminal, or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 This is a hardware structure block diagram of a mobile terminal for a trajectory prediction method for an autonomous vehicle according to an embodiment of the present invention. Figure 1 As shown, a mobile terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0033] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the device information display method in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the aforementioned networks may include wireless networks provided by the mobile terminal's communication provider. In one example, the transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0034] This embodiment provides a trajectory prediction method for an intelligent agent running on a mobile terminal, computer terminal, or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] Figure 2 This is a flowchart of a trajectory prediction method for an intelligent agent according to an embodiment of this application. Figure 2 As shown, the trajectory prediction method includes the following steps:

[0036] Step S201: Construct a preset network based at least on the first encoder and the second encoder. The first encoder is used to perform cross-temporal unidirectional interaction modeling of the target intelligent agent in the first region, and the second encoder is used to perform cross-temporal bidirectional interaction modeling of the target intelligent agent in the second region. The first region is a part of the second region, and the preset network is used to predict the trajectory of the target intelligent agent.

[0037] Specifically, the second region mentioned above is a region centered on the autonomous vehicle and bounded by a radius equal to the maximum perceptible range of the autonomous vehicle (e.g., 200 meters, a calibrable value). This second region includes multiple agents, and the target agent selection algorithm mentioned later can be used to filter these agents, resulting in multiple target agents within the second region. After determining the multiple target agents in the second region, multiple first regions can be obtained by using each target agent as a center and a radius of 50 meters (calibrable). In other words, the first regions mentioned above are partial regions of the second region. It should also be noted that the first regions corresponding to each target agent can overlap spatially. Of course, if the number of target agents in the filtered second region is very small, the first regions corresponding to each target agent may not overlap spatially. That is, in this application, the intersection of the first regions corresponding to each target agent can form the second region. Furthermore, other target agents (i.e., other agents) within each first region may or may not overlap.

[0038] Specifically, the first encoder is used to perform cross-temporal and spatial unidirectional interaction modeling of target agents in the first region. That is, it models the cross-temporal and spatial unidirectional interactions between target agents in the first region and other agents. In other words, it extracts the cross-temporal and spatial unidirectional interaction influence relationships between target agents in the first region and other agents, where "other agents" refers to target agents in the first region other than the target agent. For example, in a lane with two target vehicles (two target agents), the first target vehicle is in front of the second target vehicle. Since the actions of the first target vehicle will affect the second target vehicle, the first target vehicle has a higher priority than the second target vehicle. Because human and vehicle reactions take time, the impact of the actions of the preceding vehicle on the actions of the following vehicle is often delayed in the displayed representation. Therefore, it is necessary to establish the cross-time and space interaction between the first target vehicle and the second target vehicle, i.e., cross-temporal and spatial interaction (for example, the turning of the first target vehicle at time T-2 affects the turning of the second target vehicle at time T; the acceleration of the first target vehicle at time T-2 affects the acceleration of the second target vehicle at time T; and the starting of the first target vehicle at time T-2 affects the starting of the second target vehicle at time T, etc.). Of course, in practical applications, it is not limited to the time and scenario examples listed above, and the interaction modeling between the first target vehicle and the second target vehicle is not unidirectional across time and space.

[0039] Specifically, the second encoder is used to perform bidirectional interaction modeling of target agents in the second target region across time and space, that is, to perform bidirectional interaction modeling for any two target agents in the second region, which is to extract the mutual influence relationship between any two target agents. For example, there are three target vehicles (i.e., three target agents) in two parallel lanes. The two lanes are the first lane and the second lane. The first target vehicle is in the first lane, while the second target vehicle and the third target vehicle are in the second lane, with the second target vehicle in front of the third target vehicle, and the first and second target vehicles are in a parallel state. Then, the turning of the first target vehicle will affect whether the third target vehicle can overtake and enter the first lane, and the second target vehicle will affect the state of the third target vehicle. Of course, there is also a mutual influence relationship between the first and second target vehicles. Therefore, it is necessary to perform bidirectional interaction modeling across time and space between the first target vehicle, the second target vehicle, and the third target vehicle.

[0040] Specifically, the aforementioned target intelligent agent can be a person or vehicle that interacts with the autonomous vehicle within its perception range. The aforementioned cross-temporal one-way interaction modeling and cross-temporal two-way interaction modeling can be used to model interactions between pedestrians, between pedestrians and vehicles, and between vehicles.

[0041] Step S202: Using the target dataset, train the above-mentioned preset network to obtain the target network. The target dataset includes multiple sets of training data and the target results corresponding to each set of training data.

[0042] Specifically, the target dataset mentioned above can be a self-made dataset, or it can be an open-source dataset. This application does not impose any specific restrictions on the target dataset.

[0043] Specifically, the target dataset is used to train the preset network until the loss function converges or the prediction accuracy reaches a predetermined value. In other words, this application does not impose any restrictions on the cutoff conditions for using the target dataset to train the preset network; these conditions can be flexibly adjusted and designed according to actual circumstances.

[0044] Step S203: At least the above-mentioned target network and real-time perception information are used to predict the trajectory of the real-time target intelligent agent. The above-mentioned real-time perception information is the environmental information and intelligent agent information perceived by the autonomous vehicle in the target area during the driving process.

[0045] Specifically, the aforementioned target area can be defined in practical applications as the region centered on the autonomous vehicle and with the maximum perceptible range of the corresponding autonomous vehicle as its radius. After acquiring the environmental and agent information within the target area—that is, after obtaining the real-time perception information of the corresponding autonomous vehicle—the target agent selection algorithm mentioned later can be used to filter the agents within the target area to obtain real-time target agents. Then, a target network is used to predict the trajectory of the real-time target agents.

[0046] Specifically, in practical applications, after predicting the trajectory of a real-time target intelligent agent based on a target network, the corresponding autonomous vehicle can plan and make decisions about its own trajectory based on the predicted trajectory of the real-time target intelligent agent.

[0047] In this embodiment, firstly, a preset network with at least a first encoder and a second encoder is constructed, wherein the first encoder is used to perform cross-temporal unidirectional interaction modeling of the target agent in the first region, and the second encoder is used to perform cross-temporal bidirectional interaction modeling of the target agent in the second target region; then, the preset network is trained using the target dataset to obtain the target network; finally, at least the trained target network and real-time perception information are used to predict the trajectory of the corresponding real-time target agent. Compared with existing technologies that perform cross-spatial bidirectional interactive modeling of target agents in the first region, this application uses a first encoder to perform cross-spatial unidirectional interactive modeling of target agents in the first region (i.e., cross-spatial unidirectional interactive modeling of target agents in a local region) and a second encoder to perform cross-spatial bidirectional interactive modeling of target agents in the second region (i.e., cross-spatial bidirectional interactive modeling of target agents in the entire region), instead of performing interactive modeling of all agents in the first and second regions. This results in a smaller overall computational load and less information loss during the interactive modeling process, ensuring higher accuracy in trajectory prediction of target agents (and real-time target agents in practical applications). This solves the problem of large information loss and computational load in existing trajectory prediction technologies for interactive modeling between agents in traffic, thereby ensuring that the autonomous vehicle's subsequent trajectory planning based on the predicted real-time target agent trajectory is more reasonable and accurate, ensuring higher safety and intelligence of the autonomous vehicle.

[0048] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0049] In the specific implementation process, the above step S201 can be achieved through the following steps: The process of the first encoder performing cross-spatial unidirectional interaction modeling of the target intelligent agents in the first region includes: an evaluation step, using a priority evaluator to evaluate the priority of multiple target intelligent agents in the first region to obtain the priority order between any two target intelligent agents; a modeling step, based on the priority order between any two target intelligent agents, performing cross-spatial unidirectional interaction modeling on the corresponding two target intelligent agents, that is, modeling the influence relationship between the target intelligent agent with higher priority and the intelligent agent with lower priority between the two target intelligent agents, while not performing cross-spatial unidirectional interaction modeling between two target intelligent agents with the same priority; and a repetition step, repeating the above evaluation step and the above modeling step at least once, until cross-spatial unidirectional interaction modeling is performed on the target intelligent agents in the first region at all target times. In this embodiment, a priority order among target agents in the first region is constructed based on a priority evaluator. Then, cross-spatial unidirectional interaction modeling is performed based on this priority order, specifically extracting the influence features of higher-priority target agents on lower-priority target agents. This further ensures that the number of cross-spatial unidirectional modeling operations is reduced, thus minimizing the overall computational load. Subsequently, cross-spatial bidirectional interaction modeling is performed on the target agents in the second region based on a second encoder. This further ensures that the interaction features obtained from the first and second encoders are relatively accurate, minimizing information loss in the encoding of the target agents' interaction features.

[0050] In one specific embodiment of this application, such as Figure 3As shown, at time T-2, there is a first target agent 2011 and a first other agent 2021; at time T-1, there is a second target agent 2012 and a second other agent 2022; and at time T, there is a third target agent 2013 and a third other agent 2023. The first target agent 2011, the second target agent 2012, and the third target agent 2013 have a first priority 201, while the first other agent 2021, the second other agent 2022, and the third other agent 2023 have a second priority 202, with the first priority 201 being higher than the second priority 202. A cross-spatial-temporal one-way interaction model is performed at time T-2, meaning that the first target agent 2011 performs a cross-spatial-temporal one-way interaction model with the first other agent 2021 at time T-2, the second other agent 2022 at time T-1, and the third other agent 2023 at time T. A one-way interaction model is performed between the second target agent 2012 at time T-1 and other agents 2022 and 2023 at time T. Similarly, a one-way interaction model is performed between the second target agent 2012 and other agents 2023 at time T.

[0051] In practical applications, there are two target vehicles (two target agents) on a lane, with the first target vehicle located in front of the second target vehicle. The actions of the first target vehicle affect the actions of the second target vehicle. For example, there is a delay between the actions of the first target vehicle at the moment of its start and those of the second target vehicle. In other words, the state of the first target vehicle at the moment of its start determines the delayed start of the second target vehicle. In summary, the past historical state of a higher-priority target agent influences the current state of a lower-priority target agent. Therefore, by modeling the unidirectional influence interaction between priority target agents across time and space, we can better establish interactions, reduce the loss of interaction information, further ensure the accuracy of trajectory prediction for the target agents, and further ensure the accuracy of subsequent trajectory planning and decision-making for autonomous vehicles.

[0052] Specifically, in the above embodiments, a priority evaluator can be used to evaluate the priority of the target agents in the first region at each target time.

[0053] To further improve the accuracy of the trajectory prediction method of this application, the trajectory prediction method for autonomous vehicles of this application may further include step S204. After performing cross-temporal one-way interactive modeling on the target intelligent agents in the first region at all target times, the trajectory prediction method further includes: step S204, performing cross-temporal one-way interactive modeling on multiple target intelligent agents at multiple target times with high-precision maps respectively. For example, at time T-2, there is one target intelligent agent and two other intelligent agents in the first region; at time T-1, there is one target intelligent agent and one other intelligent agent in the first region (the other intelligent agent is one of the two other intelligent agents at time T-2); at time T, there is one target intelligent agent and two other intelligent agents in the first region (the two other intelligent agents are both the two other intelligent agents at time T-2). Therefore, the intelligent agents at the intersection of the target intelligent agent and other intelligent agents at time T-2, time T-1, and time T (here, the target intelligent agent and other intelligent agents are collectively referred to as such) are respectively subjected to cross-temporal one-way interactive modeling on high-precision maps, that is, adding the influence relationship of road maps on the behavior of the target intelligent agents.

[0054] Step S201 of this application can also be implemented through the following steps: The process of the second encoder performing cross-spatial bidirectional interaction modeling of the target intelligent agents in the second region includes: modeling the feature information of each target intelligent agent in the second region, wherein the feature information includes at least velocity information, direction information, position information, and category information; and performing cross-spatial bidirectional interaction modeling of any two target intelligent agents in the second region at each target time. Specifically, in this scheme, modeling the feature information of each target intelligent agent in the second region means extracting the features of the feature information of each target intelligent agent in the second region. And performing cross-spatial bidirectional interaction modeling of any two target intelligent agents in the second region at each target time allows for a greater acquisition of the mutual influence information between any two target intelligent agents, further ensuring that the interaction features obtained based on the first encoder and the second encoder are relatively accurate, and further ensuring that the information loss in encoding the interaction features of the target intelligent agents is minimal.

[0055] Specifically, the speed information can be calculated using any feasible method in the prior art, and this application does not limit the calculation method of the speed information.

[0056] Specifically, the aforementioned directional information can be the driving direction information of the target intelligent agent, or it can be the steering information of the target intelligent agent, etc., and is not limited to the directional information listed in this application. Furthermore, this application does not limit the method for obtaining the aforementioned steering information; it can be obtained using any feasible method in the prior art.

[0057] Specifically, the aforementioned location information refers to the location information of the target intelligent agent relative to the corresponding autonomous driving vehicle. This application does not limit the method for obtaining the aforementioned location information; any feasible method in the prior art can be used to obtain it.

[0058] Specifically, the aforementioned category information can be the vehicle category, such as truck, car, electric vehicle, etc. This application does not limit the method for obtaining the aforementioned category information; it can be obtained using any feasible method in the prior art.

[0059] In one specific embodiment of this application, such as Figure 4 As shown, at time T-2, there is a first target agent 2011 and a first other agent 2021; at time T-1, there is a second target agent 2012 and a second other agent 2022; and at time T, there is a third target agent 2013 and a third other agent 2023. Cross-spatial bidirectional interaction modeling is performed at time T-2, meaning that the first target agent 2011 performs cross-spatial bidirectional interaction modeling with the first other agent 2021 at time T-2, the second other agent 2022 at time T-1, and the third other agent 2023 at time T. Cross-spatial bidirectional interaction modeling is also performed with the second target agent 2012 at time T-1, meaning that the second target agent 2012 performs cross-spatial bidirectional interaction modeling with the second other agent 2022 at time T-1 and the third other agent 2023 at time T. A cross-temporal bidirectional interaction model is performed between the third target agent 2013 at time T and the third other agent 2023 at time T.

[0060] In some embodiments, the process of determining the priority estimator involves: determining the priority order of the target agents in each predetermined region based on predetermined rules and a high-precision map, thus obtaining a priority training set; and training a preset priority estimator based on the priority training set to obtain the preset priority estimator, which is constructed based on a multilayer perceptron. In this scheme, using the priority training set to train the preset priority estimator ensures that the priority estimator can be obtained relatively easily, further guaranteeing the accuracy of the subsequent priorities among the target agents in the first region obtained based on the priority evaluation.

[0061] Specifically, such as Figure 5 As shown, the above-mentioned predetermined rule can be set such that the priority of the fourth target agent 204, which passes through the designated point 205 in the first region first, is higher than that of the fifth target agent 203, which passes through the designated point 205 later. In other words, the priority of the fourth target agent 204 is higher than that of the fifth target agent 203.

[0062] In one specific embodiment of this application, such as Figure 6 As shown, after identifying the target agents in the first region, a human can determine the priority of each target agent according to predetermined rules. Then, the priorities of the target agents in the first region are manually labeled, resulting in a priority training set with labeled information. This priority training set is then used to train a pre-defined priority estimator to obtain the priority estimator. Specifically, the input can be the first target agent, the second target agent, and lane information from a high-precision map in the first region. A self-attention mechanism is used to encode features of the first target agent, the second target agent, and their respective surrounding environment. Then, an MLP (Multi-Layer Perceptron) is used to decode the priority relationship between the first and second target agents.

[0063] In the specific implementation process, the above step S201 can also be implemented through the following steps: the process of determining the target intelligent agent in the second region includes: acquiring target perception information of the autonomous vehicle, wherein the target perception information is environmental information and intelligent agent information of the second region acquired by the target device, wherein the target device includes at least one of the following: image acquisition device, radar device, and the intelligent agent information is information of multiple intelligent agents in the second region; using a target intelligent agent selection algorithm and the above target perception information to filter the intelligent agents in the second region to obtain the target intelligent agent, wherein the target intelligent agent selection algorithm is obtained based on neural network training, and the target intelligent agent is a person or vehicle that interacts with the autonomous vehicle within the perception range of the autonomous vehicle. In this embodiment, before using the first encoder to perform cross-spatial one-way interactive modeling of the target agent in the first region, and before using the second encoder to perform cross-spatial two-way interactive modeling of the target agent in the second region, the agents in the second region are screened by a target agent selection algorithm. This ensures that the target agents in the second region can be obtained more accurately, and further reduces the computational load of performing cross-spatial one-way interactive modeling of the first region and cross-spatial two-way interactive modeling of the second region.

[0064] Specifically, the aforementioned intelligent agent can be all traffic participants detected in the target perception information perceived by the autonomous vehicle. To reduce computational load, this application proposes a target intelligent agent selection algorithm to filter the intelligent agents in the target perception information to obtain the target intelligent agents.

[0065] Specifically, the target agent selection algorithm described above can be a classification algorithm. After training, all agents in the target perception information perceived by the autonomous vehicle can be divided into target agents that interact with the autonomous vehicle and non-target agents that do not interact with the autonomous vehicle.

[0066] Of course, in practical applications, the second region is first determined based on the autonomous vehicle. After determining the second region, a target agent selection algorithm is used to filter the agents in the second region to obtain the target agents in the second region. The first region corresponding to each target agent is then determined again. This reduces the computational workload of subsequent cross-spatial unidirectional interaction modeling of the first region and cross-spatial bidirectional interaction modeling of the second region.

[0067] To further improve the accuracy of trajectory prediction for autonomous vehicles, step S201 can be implemented through the following steps: Constructing a preset network based at least on the first encoder and the second encoder, including: constructing the preset network based on the first encoder, the second encoder, and the target decoder. The target decoder is used to predict the trajectory of the target agent based on the target interaction information output by the first encoder and the second encoder. In other words, in this scheme, the first encoder and the second encoder are used to extract interaction information between target agents from different perspectives. Then, the extracted interaction information between target agents is input to the target decoder to achieve trajectory prediction of the target agent. This further ensures the accuracy of trajectory prediction for the target agent and further ensures the accuracy of subsequent trajectory planning and decision-making for autonomous vehicles.

[0068] Specifically, the aforementioned target interaction information can be obtained by fusing feature information from the spatiotemporal unidirectional interaction modeling between target agents in the first region obtained by the first encoder and the spatiotemporal bidirectional interaction modeling between target agents in the second region obtained by the second encoder. Subsequently, trajectory prediction of the target agents can be performed based on the target interaction information and the target decoder.

[0069] Of course, in practical applications, to further reduce the computational load of the trajectory prediction method of this application, the interaction features encoded by the first encoder and the second encoder can be simply combined to obtain the target interaction information. Subsequently, trajectory prediction of the target agent can be performed based on the target interaction information and the target decoder.

[0070] To enable those skilled in the art to better understand the technical solution of this application, the implementation process of the trajectory prediction method for intelligent agents of this application will be described in detail below with reference to specific embodiments.

[0071] This embodiment relates to a specific trajectory prediction method for an intelligent agent, such as... Figure 7 As shown, it includes the following steps:

[0072] Step S1: Using the autonomous vehicle as the center and the maximum range that the autonomous vehicle can perceive as the radius, determine the second area corresponding to the autonomous vehicle.

[0073] Step S2: Using a target agent selection algorithm, all agents in the target perception information of the second region perceived by the autonomous vehicle are filtered to obtain the target agents in the second region;

[0074] Step S3: Determine the first region corresponding to each target agent;

[0075] Step S4: Using the first encoder, perform cross-temporal and spatiotemporal one-way interaction modeling of the target intelligent agent in the first region, and using the second encoder, perform cross-temporal and spatiotemporal two-way interaction modeling of the target intelligent agent in the second region.

[0076] Step S5: The one-way interaction information of the target agent in the first region obtained by the first encoder interaction modeling, and the two-way interaction information of the target agent in the second region obtained by the second encoder interaction modeling, constitute the target interaction information;

[0077] Step S6: The target decoder predicts the trajectory of the target agent based on the target interaction information, thereby obtaining the target network.

[0078] This application also provides a trajectory prediction device for an intelligent agent. It should be noted that the trajectory prediction device for an intelligent agent in this application can be used to execute the trajectory prediction method for autonomous vehicles provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0079] The trajectory prediction device for intelligent agents provided in the embodiments of this application will be described below.

[0080] Figure 8 This is a schematic diagram of the trajectory prediction device for an intelligent agent according to an embodiment of this application. Figure 8 As shown, the trajectory prediction device includes:

[0081] The construction unit 10 is used to construct a preset network based at least on a first encoder and a second encoder. The first encoder is used to perform cross-temporal unidirectional interaction modeling of the target intelligent agent in the first region, and the second encoder is used to perform cross-temporal bidirectional interaction modeling of the target intelligent agent in the second region. The first region is a part of the second region, and the preset network is used to predict the trajectory of the target intelligent agent.

[0082] Specifically, the second region mentioned above is a region centered on the autonomous vehicle and bounded by a radius equal to the maximum perceptible range of the autonomous vehicle (e.g., 200 meters, a calibrable value). This second region includes multiple agents, and the target agent selection algorithm mentioned later can be used to filter these agents, resulting in multiple target agents within the second region. After determining the multiple target agents in the second region, multiple first regions can be obtained by using each target agent as a center and a radius of 50 meters (calibrable). In other words, the first regions mentioned above are partial regions of the second region. It should also be noted that the first regions corresponding to each target agent can overlap spatially. Of course, if the number of target agents in the filtered second region is very small, the first regions corresponding to each target agent may not overlap spatially. That is, in this application, the intersection of the first regions corresponding to each target agent can form the second region. Furthermore, other target agents (i.e., other agents) within each first region may or may not overlap.

[0083] Specifically, the first encoder is used to perform cross-temporal and spatial unidirectional interaction modeling of target agents in the first region. That is, it models the cross-temporal and spatial unidirectional interactions between target agents in the first region and other agents. In other words, it extracts the cross-temporal and spatial unidirectional interaction influence relationships between target agents in the first region and other agents, where "other agents" refers to target agents in the first region other than the target agent. For example, in a lane with two target vehicles (two target agents), the first target vehicle is in front of the second target vehicle. Since the actions of the first target vehicle will affect the second target vehicle, the first target vehicle has a higher priority than the second target vehicle. Because human and vehicle reactions take time, the impact of the actions of the preceding vehicle on the actions of the following vehicle is often delayed in the displayed representation. Therefore, it is necessary to establish the cross-time and space interaction between the first target vehicle and the second target vehicle, i.e., cross-temporal and spatial interaction (for example, the turning of the first target vehicle at time T-2 affects the turning of the second target vehicle at time T; the acceleration of the first target vehicle at time T-2 affects the acceleration of the second target vehicle at time T; and the starting of the first target vehicle at time T-2 affects the starting of the second target vehicle at time T, etc.). Of course, in practical applications, it is not limited to the time and scenario examples listed above, and the interaction modeling between the first target vehicle and the second target vehicle is not unidirectional across time and space.

[0084] Specifically, the second encoder is used to perform bidirectional interaction modeling of target agents in the second target region across time and space, that is, to perform bidirectional interaction modeling for any two target agents in the second region, which is to extract the mutual influence relationship between any two target agents. For example, there are three target vehicles (i.e., three target agents) in two parallel lanes. The two lanes are the first lane and the second lane. The first target vehicle is in the first lane, while the second target vehicle and the third target vehicle are in the second lane, with the second target vehicle in front of the third target vehicle, and the first and second target vehicles are in a parallel state. Then, the turning of the first target vehicle will affect whether the third target vehicle can overtake and enter the first lane, and the second target vehicle will affect the state of the third target vehicle. Of course, there is also a mutual influence relationship between the first and second target vehicles. Therefore, it is necessary to perform bidirectional interaction modeling across time and space between the first target vehicle, the second target vehicle, and the third target vehicle.

[0085] Specifically, the aforementioned target intelligent agent can be a person or vehicle that interacts with the autonomous vehicle within its perception range. The aforementioned cross-temporal one-way interaction modeling and cross-temporal two-way interaction modeling can be used to model interactions between pedestrians, between pedestrians and vehicles, and between vehicles.

[0086] Training unit 20 is used to train the above-mentioned preset network using the target dataset to obtain the target network. The target dataset includes multiple sets of training data and the target results corresponding to each set of training data.

[0087] Specifically, the target dataset mentioned above can be a self-made dataset, or it can be an open-source dataset. This application does not impose any specific restrictions on the target dataset.

[0088] Specifically, the target dataset is used to train the preset network until the loss function converges or the prediction accuracy reaches a predetermined value. In other words, this application does not impose any restrictions on the cutoff conditions for using the target dataset to train the preset network; these conditions can be flexibly adjusted and designed according to actual circumstances.

[0089] The prediction unit 30 is used to predict the trajectory of a real-time target agent by at least using the aforementioned target network and real-time perception information. The aforementioned real-time perception information is environmental information and agent information in the target area perceived by the autonomous vehicle during its operation.

[0090] Specifically, the aforementioned target area can be defined in practical applications as the region centered on the autonomous vehicle and with the maximum perceptible range of the corresponding autonomous vehicle as its radius. After acquiring the environmental and agent information within the target area—that is, after obtaining the real-time perception information of the corresponding autonomous vehicle—the target agent selection algorithm mentioned later can be used to filter the agents within the target area to obtain real-time target agents. Then, a target network is used to predict the trajectory of the real-time target agents.

[0091] Specifically, in practical applications, after predicting the trajectory of a real-time target intelligent agent based on a target network, the corresponding autonomous vehicle can plan and make decisions about its own trajectory based on the predicted trajectory of the real-time target intelligent agent.

[0092] In the aforementioned trajectory prediction device, the construction unit is used to construct a preset network having at least a first encoder and a second encoder, wherein the first encoder is used to perform cross-temporal unidirectional interaction modeling of the target agent in the first region; the second encoder is used to perform cross-temporal bidirectional interaction modeling of the target agent in the second target region; the training unit is used to train the preset network using the target dataset to obtain the target network; and the prediction unit uses at least the trained target network and real-time perception information to perform trajectory prediction of the corresponding real-time target agent. Compared with existing technologies that perform cross-spatial bidirectional interactive modeling of target agents in the first region, this application uses a first encoder to perform cross-spatial unidirectional interactive modeling of target agents in the first region (i.e., cross-spatial unidirectional interactive modeling of target agents in a local region) and a second encoder to perform cross-spatial bidirectional interactive modeling of target agents in the second region (i.e., cross-spatial bidirectional interactive modeling of target agents in the entire region), instead of performing interactive modeling of all agents in the first and second regions. This results in a smaller overall computational load and less information loss during the interactive modeling process, ensuring higher accuracy in trajectory prediction of target agents (and real-time target agents in practical applications). This solves the problem of large information loss and computational load in existing trajectory prediction technologies for interactive modeling between agents in traffic, thereby ensuring that the autonomous vehicle's subsequent trajectory planning based on the predicted real-time target agent trajectory is more reasonable and accurate, ensuring higher safety and intelligence of the autonomous vehicle.

[0093] In the specific implementation process, the aforementioned construction unit includes an evaluation module, a first modeling module, and a repetition module. The evaluation module is used for the evaluation step, employing a priority evaluator to evaluate the priorities of multiple target agents in the first region, obtaining the priority order between any two target agents. The first modeling module is used for the modeling step, performing cross-spatial unidirectional interaction modeling on the corresponding two target agents based on the priority order between any two target agents. That is, it models the influence relationship between the higher-priority target agent and the lower-priority agent between the two target agents, while not performing cross-spatial unidirectional interaction modeling between two target agents of the same priority. The repetition module is used for the repetition step, repeating the evaluation step and the modeling step at least once, until cross-spatial unidirectional interaction modeling is performed on the target agents in the first region at all target times. In this embodiment, a priority order among target agents in the first region is constructed based on a priority evaluator. Then, cross-spatial unidirectional interaction modeling is performed based on this priority order, specifically extracting the influence features of higher-priority target agents on lower-priority target agents. This further ensures that the number of cross-spatial unidirectional modeling operations is reduced, thus minimizing the overall computational load. Subsequently, cross-spatial bidirectional interaction modeling is performed on the target agents in the second region based on a second encoder. This further ensures that the interaction features obtained from the first and second encoders are relatively accurate, minimizing information loss in the encoding of the target agents' interaction features.

[0094] In one specific embodiment of this application, such as Figure 3As shown, at time T-2, there is a first target agent 2011 and a first other agent 2021; at time T-1, there is a second target agent 2012 and a second other agent 2022; and at time T, there is a third target agent 2013 and a third other agent 2023. The first target agent 2011, the second target agent 2012, and the third target agent 2013 have a first priority 201, while the first other agent 2021, the second other agent 2022, and the third other agent 2023 have a second priority 202, with the first priority 201 being higher than the second priority 202. A cross-spatial-temporal one-way interaction model is performed at time T-2, meaning that the first target agent 2011 performs a cross-spatial-temporal one-way interaction model with the first other agent 2021 at time T-2, the second other agent 2022 at time T-1, and the third other agent 2023 at time T. A one-way interaction model is performed between the second target agent 2012 at time T-1 and other agents 2022 and 2023 at time T. Similarly, a one-way interaction model is performed between the second target agent 2012 and other agents 2023 at time T.

[0095] In practical applications, there are two target vehicles (two target agents) on a lane, with the first target vehicle located in front of the second target vehicle. The actions of the first target vehicle affect the actions of the second target vehicle. For example, there is a delay between the actions of the first target vehicle at the moment of its start and those of the second target vehicle; that is, the state of the first target vehicle at the moment of its start determines the delay in the start of the second target vehicle. In summary, the past state of a higher-priority target agent influences the current state of a lower-priority target agent. Therefore, by modeling the unidirectional influence interaction between priority target agents across time and space, we can better establish interactions, reduce the loss of interaction information, further ensure the accuracy of trajectory prediction for the target agents, and further ensure the accuracy of subsequent trajectory planning and decision-making for autonomous vehicles.

[0096] Specifically, in the above embodiments, a priority evaluator can be used to evaluate the priority of the target agents in the first region at each target time.

[0097] To further improve the accuracy of the trajectory prediction method of this application, the trajectory prediction device further includes a modeling unit, which, after performing cross-spatial one-way interactive modeling of the target agents in the first region at all target times, performs cross-spatial one-way interactive modeling of multiple target agents at multiple target times with high-precision maps respectively. For example, at time T-2, there is one target agent and two other agents in the first region; at time T-1, there is one target agent and one other agent in the first region (the other agent is one of the two other agents at time T-2); at time T, there is one target agent and two other agents in the first region (the two other agents are both the two other agents at time T-2). Therefore, the agents at the intersection of the target agent and other agents at time T-2, time T-1, and time T (here, the target agent and other agents are collectively referred to as such) are respectively modeled with high-precision maps to perform cross-spatial one-way interactive modeling, that is, to add the influence relationship of road maps on the behavior of the target agents.

[0098] Specifically, the aforementioned construction unit further includes a second modeling module and a third modeling module. The second modeling module is used to model the feature information of each target agent in the second region, where the feature information includes at least velocity, direction, and category information. The third modeling module is used to model the cross-spatial-temporal bidirectional interaction between any two target agents in the second region at each target time. Specifically, in this scheme, modeling the feature information of each target agent in the second region involves extracting the features of each target agent in the second region. Furthermore, modeling the cross-spatial-temporal bidirectional interaction between any two target agents in the second region at each target time allows for a greater acquisition of the mutual influence information between any two target agents, further ensuring the accuracy of the interaction features obtained based on the first and second encoders, and minimizing information loss in the encoding of the target agents' interaction features.

[0099] Specifically, the speed information can be calculated using any feasible method in the prior art, and this application does not limit the calculation method of the speed information.

[0100] Specifically, the aforementioned directional information can be the driving direction information of the target intelligent agent, or it can be the steering information of the target intelligent agent, etc., and is not limited to the directional information listed in this application. Furthermore, this application does not limit the method for obtaining the aforementioned steering information; it can be obtained using any feasible method in the prior art.

[0101] Specifically, the aforementioned location information refers to the location information of the target intelligent agent relative to the corresponding autonomous driving vehicle. This application does not limit the method for obtaining the aforementioned location information; any feasible method in the prior art can be used to obtain it.

[0102] Specifically, the aforementioned category information can be the vehicle category, such as truck, car, electric vehicle, etc. This application does not limit the method for obtaining the aforementioned category information; it can be obtained using any feasible method in the prior art.

[0103] In one specific embodiment of this application, such as Figure 4 As shown, at time T-2, there is a first target agent 2011 and a first other agent 2021; at time T-1, there is a second target agent 2012 and a second other agent 2022; and at time T, there is a third target agent 2013 and a third other agent 2023. Cross-spatial bidirectional interaction modeling is performed at time T-2, meaning that the first target agent 2011 performs cross-spatial bidirectional interaction modeling with the first other agent 2021 at time T-2, the second other agent 2022 at time T-1, and the third other agent 2023 at time T. Cross-spatial bidirectional interaction modeling is also performed with the second target agent 2012 at time T-1, meaning that the second target agent 2012 performs cross-spatial bidirectional interaction modeling with the second other agent 2022 at time T-1 and the third other agent 2023 at time T. A cross-temporal bidirectional interaction model is performed between the third target agent 2013 at time T and the third other agent 2023 at time T.

[0104] To further ensure the accuracy of the trajectory prediction method for autonomous vehicles in this application, step S201 of this application may further include: performing feature information fusion on the cross-temporal one-way interaction modeling between target intelligent agents in the first region obtained by the first encoder and the cross-temporal two-way interaction modeling between target intelligent agents in the second region obtained by the second encoder to obtain target feature information, and subsequently performing trajectory prediction for autonomous vehicles based on the target feature information.

[0105] In some embodiments, the evaluation module includes a determination submodule and a training submodule. The determination submodule determines the priority order of the target agents in each predetermined region based on predetermined rules and a high-precision map, obtaining a priority training set. The training submodule trains a preset priority estimator based on the priority training set, obtaining the preset priority estimator, which is constructed based on a multilayer perceptron. In this scheme, using the priority training set to train the preset priority estimator ensures that the priority estimator can be obtained relatively easily, further guaranteeing the accuracy of the priorities among the target agents in the first region obtained subsequently based on the priority evaluation.

[0106] Specifically, such as Figure 5 As shown, the above-mentioned predetermined rule can be set such that the priority of the fourth target agent 204, which passes through the designated point 205 in the first region first, is higher than that of the fifth target agent 203, which passes through the designated point 205 later. In other words, the priority of the fourth target agent 204 is higher than that of the fifth target agent 203.

[0107] In one specific embodiment of this application, such as Figure 6 As shown, after identifying the target agents in the first region, a human can determine the priority of each target agent according to predetermined rules. Then, the priorities of the target agents in the first region are manually labeled, resulting in a priority training set with labeled information. This priority training set is then used to train a pre-defined priority evaluator to obtain a priority evaluation period. Specifically, the input can be the first target agent, the second target agent, and lane information of the high-precision area in the first region. A Self-Attention mechanism is used to encode features of the first target agent, the second target agent, and their respective surrounding environment. Then, an MLP (Multi-Layer Perceptron) is used to decode the priority relationship between the first and second target agents.

[0108] In the specific implementation process, the above-mentioned construction unit also includes an acquisition module and a filtering module. The acquisition module is used to acquire target perception information of the autonomous vehicle. The target perception information includes environmental information and agent information of the second region acquired by the target device. The target device includes at least one of the following: image acquisition equipment, radar equipment. The agent information includes information of multiple agents in the second region. The filtering module is used to filter the agents in the second region using a target agent selection algorithm and the target perception information to obtain the target agents. The target agent selection algorithm is based on neural network training. The target agents are people or vehicles that interact with the autonomous vehicle within its perception range. In this embodiment, before using the first encoder to perform cross-spatial-space unidirectional interaction modeling of the target agents in the first region, and before using the second encoder to perform cross-spatial-space bidirectional interaction modeling of the target agents in the second region, the target agent selection algorithm filters the agents in the second region. This ensures that the target agents in the second region can be obtained more accurately, further reducing the computational load of cross-spatial-space unidirectional interaction modeling of the first region and cross-spatial-space bidirectional interaction modeling of the second region.

[0109] Specifically, the aforementioned intelligent agent can be any traffic participant in the target perception information perceived by the autonomous vehicle. To reduce computational load, this application proposes a target intelligent agent selection algorithm to filter the intelligent agents in the target perception information to obtain the target intelligent agent.

[0110] Specifically, the target agent selection algorithm described above can be a classification algorithm. After training, all agents in the target perception information perceived by the autonomous vehicle can be divided into target agents that interact with the autonomous vehicle and non-target agents that do not interact with the autonomous vehicle.

[0111] Of course, in practical applications, the second region is first determined based on the autonomous vehicle. After determining the second region, a target agent selection algorithm is used to filter the agents in the second region to obtain the target agents in the second region. The first region corresponding to each target agent is then determined again. This reduces the computational workload of subsequent cross-spatial unidirectional interaction modeling of the first region and cross-spatial bidirectional interaction modeling of the second region.

[0112] To further improve the accuracy of trajectory prediction for autonomous vehicles, the aforementioned construction unit also includes a construction module for building the preset network based on the first encoder, the second encoder, and the target decoder. The target decoder is used to predict the trajectory of the autonomous vehicle based on the target interaction information output by the first and second encoders. In other words, in this scheme, the first and second encoders are used to extract interaction information between target agents from different perspectives. Then, the extracted interaction information between target agents is input into the target decoder to predict the trajectory of the target agents. This further ensures the accuracy of trajectory prediction for the target agents and, consequently, the accuracy of subsequent trajectory planning and decision-making for the autonomous vehicle.

[0113] Specifically, the aforementioned target interaction information can be obtained by fusing feature information from the spatiotemporal unidirectional interaction modeling between target agents in the first region obtained by the first encoder and the spatiotemporal bidirectional interaction modeling between target agents in the second region obtained by the second encoder. Subsequently, trajectory prediction of the target agents can be performed based on the target interaction information and the target decoder.

[0114] Of course, in practical applications, to further reduce the computational load of the trajectory prediction method of this application, the interaction features encoded by the first encoder and the second encoder can be simply combined to obtain the target interaction information. Subsequently, trajectory prediction of the target agent can be performed based on the target interaction information and the target decoder.

[0115] The trajectory prediction device for the aforementioned intelligent agent includes a processor and a memory. The aforementioned construction units, training units, and prediction units are all stored as program units in the memory, and the processor executes these program units to achieve the corresponding functions. All of the aforementioned modules reside in the same processor; alternatively, the modules may be located in different processors in any combination.

[0116] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters can address the problems of information loss and high computational cost in existing trajectory prediction technologies for modeling interactions between agents in traffic.

[0117] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0118] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the trajectory prediction method of the intelligent agent.

[0119] Specifically, trajectory prediction methods for intelligent agents include:

[0120] Step S201: Based on at least the first encoder and the second encoder, a preset network is constructed. The first encoder is used to perform cross-temporal and spatiotemporal one-way interaction modeling of the target intelligent agent in the first region; the second encoder is used to perform cross-temporal and spatiotemporal two-way interaction modeling of the target intelligent agent in the second region. The first region is a part of the second region. The preset network is used to predict the trajectory of the target intelligent agent.

[0121] Step S202: Using the target dataset, train the above-mentioned preset network to obtain the target network. The target dataset includes multiple sets of training data and the target results corresponding to each set of training data.

[0122] Step S203: At least the above-mentioned target network and real-time perception information are used to predict the trajectory of the real-time target intelligent agent. The above-mentioned real-time perception information is the environmental information and intelligent agent information perceived by the autonomous vehicle in the target area during the driving process.

[0123] Optionally, the process of the first encoder performing cross-spatial-space unidirectional interaction modeling of the target agents in the first region includes: an evaluation step, which uses a priority evaluator to evaluate the priorities of multiple target agents in the first region to obtain the priority order between any two target agents; a modeling step, which performs cross-spatial-space unidirectional interaction modeling of the corresponding two target agents based on the priority order between any two target agents; and a repetition step, which repeats the evaluation step and the modeling step at least once in sequence until cross-spatial-space unidirectional interaction modeling is performed for the target agents in the first region at all target times.

[0124] Optionally, after performing cross-spatial one-way interactive modeling of the target agents in the first region at all target times, the trajectory prediction method further includes: performing cross-spatial one-way interactive modeling of the target agents at multiple target times with the high-precision map respectively.

[0125] Optionally, the process of the second encoder performing cross-temporal bidirectional interactive modeling of the target intelligent agents in the second region includes: performing feature information modeling for each of the target intelligent agents in the second region, wherein the feature information includes at least velocity information, direction information and category information; and performing cross-temporal bidirectional interactive modeling for any two of the target intelligent agents in the second region at each target time.

[0126] Optionally, the process of determining the priority estimator is as follows: based on predetermined rules and high-precision maps, the priority order of the target agents in each predetermined area is determined to obtain a priority training set; based on the priority training set, a preset priority estimator is trained to obtain the priority estimator, which is constructed based on a multilayer perceptron.

[0127] Optionally, the process of determining the target intelligent agent in the second region includes: acquiring target perception information of the autonomous vehicle, wherein the target perception information is environmental information and intelligent agent information of the second region acquired by the target device, wherein the target device includes at least one of the following: image acquisition device, radar device, and the intelligent agent information is information of multiple intelligent agents in the second region; using a target intelligent agent selection algorithm and the target perception information to filter the intelligent agents in the second region to obtain the target intelligent agent, wherein the target intelligent agent selection algorithm is obtained based on neural network training, and the target intelligent agent is a person or vehicle that interacts with the autonomous vehicle within the perception range of the autonomous vehicle.

[0128] Optionally, a preset network is constructed based at least on the first encoder and the second encoder, including: constructing the preset network based on the first encoder, the second encoder and the target decoder, wherein the target decoder is used to predict the trajectory of the target agent based on the target interaction information output by the first encoder and the second encoder.

[0129] This invention provides a processor for running a program, wherein the program executes the trajectory prediction method of the intelligent agent.

[0130] Step S201: Based on at least the first encoder and the second encoder, a preset network is constructed. The first encoder is used to perform cross-temporal and spatiotemporal one-way interaction modeling of the target intelligent agent in the first region; the second encoder is used to perform cross-temporal and spatiotemporal two-way interaction modeling of the target intelligent agent in the second region. The first region is a part of the second region. The preset network is used to predict the trajectory of the target intelligent agent.

[0131] Step S202: Using the target dataset, train the above-mentioned preset network to obtain the target network. The target dataset includes multiple sets of training data and the target results corresponding to each set of training data.

[0132] Step S203: At least the above-mentioned target network and real-time perception information are used to predict the trajectory of the real-time target intelligent agent. The above-mentioned real-time perception information is the environmental information and intelligent agent information perceived by the autonomous vehicle in the target area during the driving process.

[0133] Optionally, the process of the first encoder performing cross-spatial-space unidirectional interaction modeling of the target agents in the first region includes: an evaluation step, which uses a priority evaluator to evaluate the priorities of multiple target agents in the first region to obtain the priority order between any two target agents; a modeling step, which performs cross-spatial-space unidirectional interaction modeling of the corresponding two target agents based on the priority order between any two target agents; and a repetition step, which repeats the evaluation step and the modeling step at least once in sequence until cross-spatial-space unidirectional interaction modeling is performed for the target agents in the first region at all target times.

[0134] Optionally, after performing cross-spatial one-way interactive modeling of the target agents in the first region at all target times, the trajectory prediction method further includes: performing cross-spatial one-way interactive modeling of the target agents at multiple target times with the high-precision map respectively.

[0135] Optionally, the process of the second encoder performing cross-temporal bidirectional interactive modeling of the target intelligent agents in the second region includes: performing feature information modeling for each of the target intelligent agents in the second region, wherein the feature information includes at least velocity information, direction information and category information; and performing cross-temporal bidirectional interactive modeling for any two of the target intelligent agents in the second region at each target time.

[0136] Optionally, the process of determining the priority estimator is as follows: based on predetermined rules and high-precision maps, the priority order of the target agents in each predetermined area is determined to obtain a priority training set; based on the priority training set, a preset priority estimator is trained to obtain the priority estimator, which is constructed based on a multilayer perceptron.

[0137] Optionally, the process of determining the target intelligent agent in the second region includes: acquiring target perception information of the autonomous vehicle, wherein the target perception information is environmental information and intelligent agent information of the second region acquired by the target device, wherein the target device includes at least one of the following: image acquisition device, radar device, and the intelligent agent information is information of multiple intelligent agents in the second region; using a target intelligent agent selection algorithm and the target perception information to filter the intelligent agents in the second region to obtain the target intelligent agent, wherein the target intelligent agent selection algorithm is obtained based on neural network training, and the target intelligent agent is a person or vehicle that interacts with the autonomous vehicle within the perception range of the autonomous vehicle.

[0138] Optionally, a preset network is constructed based at least on the first encoder and the second encoder, including: constructing the preset network based on the first encoder, the second encoder and the target decoder, wherein the target decoder is used to predict the trajectory of the target agent based on the target interaction information output by the first encoder and the second encoder.

[0139] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs at least the following steps:

[0140] Step S201: Based on at least the first encoder and the second encoder, a preset network is constructed. The first encoder is used to perform cross-temporal and spatiotemporal one-way interaction modeling of the target intelligent agent in the first region; the second encoder is used to perform cross-temporal and spatiotemporal two-way interaction modeling of the target intelligent agent in the second region. The first region is a part of the second region. The preset network is used to predict the trajectory of the target intelligent agent.

[0141] Step S202: Using the target dataset, train the above-mentioned preset network to obtain the target network. The target dataset includes multiple sets of training data and the target results corresponding to each set of training data.

[0142] Step S203: At least the above-mentioned target network and real-time perception information are used to predict the trajectory of the real-time target intelligent agent. The above-mentioned real-time perception information is the environmental information and intelligent agent information perceived by the autonomous vehicle in the target area during the driving process.

[0143] Optionally, the process of the first encoder performing cross-spatial-space unidirectional interaction modeling of the target agents in the first region includes: an evaluation step, which uses a priority evaluator to evaluate the priorities of multiple target agents in the first region to obtain the priority order between any two target agents; a modeling step, which performs cross-spatial-space unidirectional interaction modeling of the corresponding two target agents based on the priority order between any two target agents; and a repetition step, which repeats the evaluation step and the modeling step at least once in sequence until cross-spatial-space unidirectional interaction modeling is performed for the target agents in the first region at all target times.

[0144] Optionally, after performing cross-spatial one-way interactive modeling of the target agents in the first region at all target times, the trajectory prediction method further includes: performing cross-spatial one-way interactive modeling of the target agents at multiple target times with the high-precision map respectively.

[0145] Optionally, the process of the second encoder performing cross-temporal bidirectional interactive modeling of the target intelligent agents in the second region includes: performing feature information modeling for each of the target intelligent agents in the second region, wherein the feature information includes at least velocity information, direction information and category information; and performing cross-temporal bidirectional interactive modeling for any two of the target intelligent agents in the second region at each target time.

[0146] Optionally, the process of determining the priority estimator is as follows: based on predetermined rules and high-precision maps, the priority order of the target agents in each predetermined area is determined to obtain a priority training set; based on the priority training set, a preset priority estimator is trained to obtain the priority estimator, which is constructed based on a multilayer perceptron.

[0147] Optionally, the process of determining the target intelligent agent in the second region includes: acquiring target perception information of the autonomous vehicle, wherein the target perception information is environmental information and intelligent agent information of the second region acquired by the target device, wherein the target device includes at least one of the following: image acquisition device, radar device, and the intelligent agent information is information of multiple intelligent agents in the second region; using a target intelligent agent selection algorithm and the target perception information to filter the intelligent agents in the second region to obtain the target intelligent agent, wherein the target intelligent agent selection algorithm is obtained based on neural network training, and the target intelligent agent is a person or vehicle that interacts with the autonomous vehicle within the perception range of the autonomous vehicle.

[0148] Optionally, a preset network is constructed based at least on the first encoder and the second encoder, including: constructing the preset network based on the first encoder, the second encoder and the target decoder, wherein the target decoder is used to predict the trajectory of the target agent based on the target interaction information output by the first encoder and the second encoder.

[0149] The devices mentioned in this article can be servers, PCs, tablets, mobile phones, etc.

[0150] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program having at least the following method steps:

[0151] Step S201: Based on at least the first encoder and the second encoder, a preset network is constructed. The first encoder is used to perform cross-temporal and spatiotemporal one-way interaction modeling of the target intelligent agent in the first region; the second encoder is used to perform cross-temporal and spatiotemporal two-way interaction modeling of the target intelligent agent in the second region. The first region is a part of the second region. The preset network is used to predict the trajectory of the target intelligent agent.

[0152] Step S202: Using the target dataset, train the above-mentioned preset network to obtain the target network. The target dataset includes multiple sets of training data and the target results corresponding to each set of training data.

[0153] Step S203: At least the above-mentioned target network and real-time perception information are used to predict the trajectory of the real-time target intelligent agent. The above-mentioned real-time perception information is the environmental information and intelligent agent information perceived by the autonomous vehicle in the target area during the driving process.

[0154] Optionally, the process of the first encoder performing cross-spatial-space unidirectional interaction modeling of the target agents in the first region includes: an evaluation step, which uses a priority evaluator to evaluate the priorities of multiple target agents in the first region to obtain the priority order between any two target agents; a modeling step, which performs cross-spatial-space unidirectional interaction modeling of the corresponding two target agents based on the priority order between any two target agents; and a repetition step, which repeats the evaluation step and the modeling step at least once in sequence until cross-spatial-space unidirectional interaction modeling is performed for the target agents in the first region at all target times.

[0155] Optionally, after performing cross-spatial one-way interactive modeling of the target agents in the first region at all target times, the trajectory prediction method further includes: performing cross-spatial one-way interactive modeling of the target agents at multiple target times with the high-precision map respectively.

[0156] Optionally, the process of the second encoder performing cross-temporal bidirectional interactive modeling of the target intelligent agents in the second region includes: performing feature information modeling for each of the target intelligent agents in the second region, wherein the feature information includes at least velocity information, direction information and category information; and performing cross-temporal bidirectional interactive modeling for any two of the target intelligent agents in the second region at each target time.

[0157] Optionally, the process of determining the priority estimator is as follows: based on predetermined rules and high-precision maps, the priority order of the target agents in each predetermined area is determined to obtain a priority training set; based on the priority training set, a preset priority estimator is trained to obtain the priority estimator, which is constructed based on a multilayer perceptron.

[0158] Optionally, the process of determining the target intelligent agent in the second region includes: acquiring target perception information of the autonomous vehicle, wherein the target perception information is environmental information and intelligent agent information of the second region acquired by the target device, wherein the target device includes at least one of the following: image acquisition device, radar device, and the intelligent agent information is information of multiple intelligent agents in the second region; using a target intelligent agent selection algorithm and the target perception information to filter the intelligent agents in the second region to obtain the target intelligent agent, wherein the target intelligent agent selection algorithm is obtained based on neural network training, and the target intelligent agent is a person or vehicle that interacts with the autonomous vehicle within the perception range of the autonomous vehicle.

[0159] Optionally, a preset network is constructed based at least on the first encoder and the second encoder, including: constructing the preset network based on the first encoder, the second encoder and the target decoder, wherein the target decoder is used to predict the trajectory of the target agent based on the target interaction information output by the first encoder and the second encoder.

[0160] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0161] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0162] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0163] 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.

[0164] 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.

[0165] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0166] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0167] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0168] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0169] As can be seen from the above description, the embodiments of this application achieve the following technical effects:

[0170] 1) In the trajectory prediction method of this application, firstly, a preset network with at least a first encoder and a second encoder is constructed, wherein the first encoder is used to perform cross-temporal and spatiotemporal unidirectional interaction modeling of the target agent in the first region; the second encoder is used to perform cross-temporal and spatiotemporal bidirectional interaction modeling of the target agent in the second target region; then, the preset network is trained using the target dataset to obtain the target network; finally, the trajectory prediction of the corresponding real-time target agent is performed using at least the trained target network and real-time perception information. Compared with existing technologies that perform cross-spatial bidirectional interactive modeling of target agents in the first region, this application uses a first encoder to perform cross-spatial unidirectional interactive modeling of target agents in the first region (i.e., cross-spatial unidirectional interactive modeling of target agents in a local region) and a second encoder to perform cross-spatial bidirectional interactive modeling of target agents in the second region (i.e., cross-spatial bidirectional interactive modeling of target agents in the entire region), instead of performing interactive modeling of all agents in the first and second regions. This results in a smaller overall computational load and less information loss during the interactive modeling process, ensuring higher accuracy in trajectory prediction of target agents (and real-time target agents in practical applications). This solves the problem of large information loss and computational load in existing trajectory prediction technologies for interactive modeling between agents in traffic, thereby ensuring that the autonomous vehicle's subsequent trajectory planning based on the predicted real-time target agent trajectory is more reasonable and accurate, ensuring higher safety and intelligence of the autonomous vehicle.

[0171] 2) In the trajectory prediction device of this application, the construction unit is used to construct a preset network with at least a first encoder and a second encoder, wherein the first encoder is used to perform cross-temporal unidirectional interaction modeling of the target agent in the first region; the second encoder is used to perform cross-temporal bidirectional interaction modeling of the target agent in the second target region; the training unit is used to train the preset network using the target dataset to obtain the target network; the prediction unit uses at least the trained target network and real-time perception information to perform trajectory prediction of the corresponding real-time target agent. Compared with existing technologies that perform cross-spatial bidirectional interactive modeling of target agents in the first region, this application uses a first encoder to perform cross-spatial unidirectional interactive modeling of target agents in the first region (i.e., cross-spatial unidirectional interactive modeling of target agents in a local region) and a second encoder to perform cross-spatial bidirectional interactive modeling of target agents in the second region (i.e., cross-spatial bidirectional interactive modeling of target agents in the entire region), instead of performing interactive modeling of all agents in the first and second regions. This results in a smaller overall computational load and less information loss during the interactive modeling process, ensuring higher accuracy in trajectory prediction of target agents (and real-time target agents in practical applications). This solves the problem of large information loss and computational load in existing trajectory prediction technologies for interactive modeling between agents in traffic, thereby ensuring that the autonomous vehicle's subsequent trajectory planning based on the predicted real-time target agent trajectory is more reasonable and accurate, ensuring higher safety and intelligence of the autonomous vehicle.

[0172] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A trajectory prediction method for an intelligent agent, characterized in that, include: A preset network is constructed based at least on a first encoder and a second encoder. The first encoder is used to perform cross-spatial unidirectional interaction modeling of the target agent in a first region, and the second encoder is used to perform cross-spatial bidirectional interaction modeling of the target agent in a second region. The first region is a part of the second region, and the preset network is used to predict the trajectory of the target agent. The target network is obtained by training the preset network using the target dataset, wherein the target dataset includes multiple sets of training data and the target results corresponding to each set of training data; At least the target network and real-time perception information are used to predict the trajectory of the real-time target intelligent agent. The real-time perception information is the environmental information and intelligent agent information in the target area perceived by the autonomous vehicle during the driving process. The process of the first encoder performing cross-spatial-unidirectional interaction modeling of the target agents in the first region includes: an evaluation step, which uses a priority evaluator to evaluate the priorities of multiple target agents in the first region to obtain the priority order between any two target agents; a modeling step, which performs cross-spatial-unidirectional interaction modeling of the corresponding two target agents based on the priority order between any two target agents, wherein the cross-spatial-unidirectional interaction modeling is performed between the two target agents according to the influence relationship between the target agent with higher priority and the target agent with lower priority, while the cross-spatial-unidirectional interaction modeling is not performed between two target agents with the same priority; and a repetition step, which repeats the evaluation step and the modeling step at least once in sequence until cross-spatial-unidirectional interaction modeling is performed for the target agents in the first region at all target times.

2. The trajectory prediction method according to claim 1, characterized in that, After performing cross-spatial-temporal unidirectional interaction modeling on the target agent in the first region at all target times, the trajectory prediction method further includes: Multiple target intelligent agents at multiple target times are respectively modeled to perform cross-temporal and spatiotemporal one-way interactive interactions with high-precision maps.

3. The trajectory prediction method according to claim 1, characterized in that, The process by which the second encoder performs cross-spatial bidirectional interaction modeling of the target agent in the second region includes: Feature information modeling is performed on each target intelligent agent in the second region, and the feature information includes at least velocity information, orientation information, position information, and category information; Perform cross-temporal bidirectional interaction modeling for any two target agents in the second region at each target time.

4. The trajectory prediction method according to claim 1, characterized in that, The process of determining the priority evaluator: Based on predetermined rules and high-precision maps, the priority order of the target intelligent agents in each predetermined area is determined to obtain a priority training set; Based on the priority training set, a preset priority estimator is trained to obtain the priority estimator, which is constructed based on a multilayer perceptron.

5. The trajectory prediction method according to any one of claims 1 to 4, characterized in that, The process of identifying the target agent in the second region includes: Acquire target perception information of an autonomous vehicle, wherein the target perception information is environmental information and agent information of the second region acquired by a target device, wherein the target device includes at least one of the following: an image acquisition device or a radar device, and the agent information is information of multiple agents in the second region; The target intelligent agent is obtained by using a target intelligent agent selection algorithm and the target perception information to filter the intelligent agents in the second region. The target intelligent agent selection algorithm is based on neural network training. The target intelligent agent is a person or vehicle that interacts with the autonomous vehicle within the perception range of the autonomous vehicle.

6. The trajectory prediction method according to any one of claims 1 to 4, characterized in that, Based at least on the first encoder and the second encoder, a preset network is constructed, including: Based on the first encoder, the second encoder, and the target decoder, the preset network is constructed. The target decoder is used to predict the trajectory of the target agent based on the target interaction information output by the first encoder and the second encoder.

7. A trajectory prediction device for an intelligent agent, characterized in that, include: The construction unit is used to construct a preset network based at least on a first encoder and a second encoder. The first encoder is used to perform cross-spatial unidirectional interaction modeling of a target agent in a first region, and the second encoder is used to perform cross-spatial bidirectional interaction modeling of the target agent in a second region. The first region is a part of the second region, and the preset network is used to predict the trajectory of the target agent. The training unit is used to train the preset network using the target dataset to obtain the target network. The target dataset includes multiple sets of training data and the target results corresponding to each set of training data. The prediction unit is used to predict the trajectory of a real-time target agent by at least the target network and real-time perception information, wherein the real-time perception information is environmental information and agent information in the target area perceived by the autonomous vehicle during driving. The construction unit includes an evaluation module, a first modeling module, and a repetition module. The evaluation module performs an evaluation step, using a priority evaluator to evaluate the priorities of multiple target agents in the first region, obtaining a priority order between any two target agents. The first modeling module performs a modeling step, based on the priority order between any two target agents, performing cross-spatial-space unidirectional interaction modeling on corresponding two target agents. Specifically, the cross-spatial-space unidirectional interaction modeling is performed between two target agents based on the influence relationship between the higher-priority target agent and the lower-priority agent, while cross-spatial-space unidirectional interaction modeling is not performed between two target agents of the same priority. The repetition module performs a repetition step, repeating the evaluation step and the modeling step at least once, until cross-spatial-space unidirectional interaction modeling is performed on the target agents in the first region at all target times.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the trajectory prediction method of the intelligent agent according to any one of claims 1 to 6.

9. A processor, characterized in that, The processor is used to run a program, wherein the program executes the trajectory prediction method of the intelligent agent according to any one of claims 1 to 6.

Images