Vehicle track prediction method and system based on human-like memory mechanism

Abstract

The invention belongs to the field of intelligent traffic and automatic driving, and relates to a vehicle track prediction method and system based on a human-like memory mechanism, and the method comprises the following steps: 1), carrying out the modeling of real-time interaction information in a current traffic scene in a time domain and a frequency domain through a working memory generator, so as to obtain real-time working memory knowledge; 2) retrieving information from the working memory through a long-term memory builder, and finally forming continuously accumulated driving experience through an iterative updating mechanism so as to obtain long-term memory knowledge; and 3) inputting the real-time working memory knowledge and the long-term memory knowledge into a physical driving trajectory predictor to generate a future motion trajectory. By introducing a human-like memory mechanism, the accuracy of trajectory prediction is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent transportation and automatic driving, and relates to a vehicle trajectory prediction method and system, in particular to a vehicle trajectory prediction method and system based on a human-like memory mechanism. BACKGROUND

[0002] To ensure the safe driving of autonomous vehicles in complex environments, it is crucial to accurately predict the future trajectories of surrounding traffic participants. Current popular trajectory prediction methods mainly rely on advanced deep learning techniques, which analyze the motion patterns of vehicles and environmental constraints to predict trajectories. These methods use advanced algorithms such as sequence networks, graph-based methods, and generative models to capture temporal dependencies and social interactions in dynamic traffic environments.

[0003] Although existing methods perform well in saliency modeling, they often perform poorly in scenarios that require high perceptual ability and a more comprehensive understanding of the environment. They mostly focus on modeling features in the time domain, while ignoring information that may exist in other representation spaces such as the frequency domain. Experienced human drivers can easily infer the trajectories of other traffic participants in highly dynamic, diverse, and random traffic environments, and this ability is partly due to their working memory (responsible for processing current environmental information) and long-term memory (accumulated driving experience) mechanisms. However, most current vehicle trajectory prediction methods fail to effectively simulate these two memory mechanisms, resulting in a gap in prediction performance compared to human drivers.

[0004] Therefore, in view of the defects in the prior art described above, it is necessary to develop a new vehicle trajectory prediction method and system. SUMMARY

[0005] To overcome the defects of the prior art, the present application proposes a vehicle trajectory prediction method and system based on a human-like memory mechanism, which significantly improves the accuracy of trajectory prediction by introducing a human-like memory mechanism.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A vehicle trajectory prediction method based on a human-like memory mechanism, characterized by the following steps:

[0008] 1) Model real-time interaction information in the current traffic scene in the time domain and frequency domain through a working memory generator to obtain real-time working memory knowledge;

[0009] 2) Retrieve information from the working memory through a long-term memory builder and update it iteratively to form continuously accumulated driving experience, thereby obtaining long-term memory knowledge;

[0010] 3) Real-time working memory knowledge and long-term memory knowledge are input into the physically driven trajectory predictor to generate future motion trajectories.

[0011] Preferably, the step 1) specifically comprises:

[0012] 11) Vehicle trajectory vector V i , lane vector M j and raw multi-view images acquired by vehicle-mounted cameras to obtain embedding features H i of vehicle trajectories, embedding features P j of lanes, and embedding features U t of vehicle multi-view bird's-eye views, and to obtain homologous interaction features H' i , P' j and U' t based on H i , P j and U t ;

[0013] 12) to obtain heterologous interaction features based on the homologous interaction features H' i and P' j and U' t ; and

[0014] 13) to splice the heterologous interaction features and to obtain aggregated features Z;

[0015] 14) to perform time-to-frequency conversion based on the aggregated features Z to obtain real-time working memory knowledge.

[0016] Preferably, the step 2) specifically comprises:

[0017] 21) to randomly initialize long-term memory knowledge S in the form of feature maps;

[0018] 22) to continuously optimize and update the long-term memory knowledge S in each forward propagation through continuous learning, so that the long-term memory knowledge gradually evolves into a dynamic memory bank with the long-term memory knowledge at each forward propagation and the feedback working memory features at each forward propagation in the memory bank;

[0019] 23) to convert the long-term memory knowledge S a-1 at the a-1th forward propagation and the feedback working memory features B a-1 at the a-1th forward propagation to the frequency domain space to obtain and

[0020] 24) long-term memory knowledge converted to frequency domain space and feedback working memory features modulated to obtain more robust cues and

[0021] 25) by concatenation and and applying an MLP to generate long-term memory knowledge of the a-th iteration.

[0022] Preferably, the step 3) specifically comprises:

[0023] 31) generating anchor multi-modal trajectories with the real-time working memory knowledge;

[0024] 32) optimizing the anchor multi-modal trajectories with the long-term memory knowledge to correct the anchor multi-modal trajectories to obtain future motion trajectories.

[0025] Furthermore, the present application also provides a vehicle trajectory prediction system based on a human-like memory mechanism, characterized in that it comprises:

[0026] a working memory generator for modeling real-time interaction information in a current traffic scene in time domain and frequency domain to obtain real-time working memory knowledge;

[0027] a long-term memory constructor for retrieving information from working memory and finally forming continuously accumulated driving experience through an iterative updating mechanism to obtain long-term memory knowledge;

[0028] a physically driven trajectory predictor for generating future motion trajectories based on the real-time working memory knowledge and long-term memory knowledge.

[0029] Moreover, the present application also provides a vehicle trajectory prediction device based on a human-like memory mechanism, characterized in that it comprises:

[0030] one or more processors;

[0031] a memory for storing one or more programs;

[0032] when the one or more programs are executed by the one or more processors, the one or more processors implement the vehicle trajectory prediction method based on a human-like memory mechanism as described above.

[0033] Finally, the present application also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the steps of the vehicle trajectory prediction method based on a human-like memory mechanism as described above.

[0034] Compared with the prior art, the vehicle trajectory prediction method and system based on the human-like memory mechanism have one or more of the following beneficial technical effects:

[0035] 1、The application significantly improves the accuracy of trajectory prediction by introducing a human-like memory mechanism. Traditional methods mainly focus on time domain feature modeling, while the method proposed in the application innovatively introduces the concepts of working memory and long-term memory, and through discrete Fourier transform and inverse discrete Fourier transform, important features are modeled in time and frequency domains. This multi-domain feature modeling method enhances the understanding of complex traffic scenarios, making trajectory prediction more accurate, especially in scenarios that require high perception and extensive context understanding.

[0036] 2、The application effectively enhances robustness through a synergistic modulation strategy. During the construction of long-term memory, an iterative update mechanism and a synergistic modulation strategy are used, which enables adaptive retrieval of working memory information and gradual accumulation of driving experience. More importantly, the synergistic modulation strategy helps to better filter out irrelevant noise when processing information, retaining meaningful representations, thereby significantly improving stability and robustness in complex and variable driving environments.

[0037] 3、The application designs an efficient working memory generator and a long-term memory constructor, and applies a two-stage training strategy, which can better meet the real-time requirements of autonomous vehicles, providing strong support for achieving safer and smarter autonomous driving. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a flowchart of the vehicle trajectory prediction method based on the human-like memory mechanism of the application;

[0039] Figure 2 is a constituent schematic diagram of the vehicle trajectory prediction system based on the human-like memory mechanism of the application. DETAILED DESCRIPTION

[0040] Before any embodiments of the application are explained in detail, it is to be understood that the application is not limited in its application to the details of construction and the arrangement of components set forth in the following description or illustrated in the following drawings. The application is capable of other embodiments and of being practiced or being carried out in various ways. Also, it is to be understood that the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of "including," "comprising," or "having" and variations thereof herein is meant to encompass the items listed thereafter and equivalents thereof as well as additional items. Unless specified or limited otherwise, the terms "mounted," "connected," "supported," and "coupled" and variations thereof are used broadly and encompass both direct and indirect mountings, connections, supports, and couplings. Further, "connected" and "coupled" are not restricted to physical or mechanical connections or couplings.

[0041] Also, in the disclosure of the present application, the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the above terms cannot be understood as a limitation of the present application; secondly, the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of one element can be one, and in another embodiment, the number of the element can be multiple, the term "one" cannot be understood as a limitation of the number.

[0042] Figure 1 A flow chart of the vehicle trajectory prediction method based on the human-like memory mechanism of the present application is shown. As Figure 1 shown, the vehicle trajectory prediction method based on the human-like memory mechanism of the present application includes the following steps:

[0043] 1. Real-time memory knowledge acquisition.

[0044] The real-time interaction information in the current traffic scene is modeled in the time domain and the frequency domain by the working memory generator to obtain real-time working memory knowledge.

[0045] 2. Long-term memory knowledge acquisition.

[0046] Information is retrieved from the working memory by the long-term memory constructor, and through an iterative updating mechanism, the driving experience is finally accumulated to obtain long-term memory knowledge.

[0047] 3. Motion trajectory generation.

[0048] The real-time working memory knowledge and the long-term memory knowledge are input into the physical driving trajectory predictor to generate the future motion trajectory.

[0049] Figure 2 A schematic diagram of the vehicle trajectory prediction system based on the human-like memory mechanism of the present application is shown. As Figure 2As shown, the vehicle trajectory prediction system based on human-like memory mechanism of the present application is composed of three core components: working memory generator, long-term memory constructor and physically driven trajectory predictor. Among them, the working memory generator is responsible for modeling various interactions between scene elements in time and frequency domains. Further, the long-term memory constructor retrieves information from the working memory and finally forms continuously accumulated driving experience through an iterative updating mechanism. Finally, after fusion, the two types of memories are input into the physically driven trajectory predictor to generate future motion trajectories.

[0050] 1. Working memory generator.

[0051] The working memory generator aims to understand different types of interactions by encoding motion information and scene information of traffic participants. The working memory generator is composed of multiple neural networks responsible for processing input vectorized samples and outputting high-level features. The working memory generator first models the working memory elements in the time domain. Assuming that the embedding features H i of the agent i (the i-th traffic participant, i.e., the vehicle), the embedding features P j of the lane, and the embedding features U t of the vehicle multi-view overhead view are all from the historical state:

[0052]

[0053] where V i represents the trajectory vector of the vehicle, M j represents the lane vector, represents the original multi-view image obtained by the vehicle-mounted camera, m is the camera number, t is the historical observation timestamp, know represents the LSTM layer, and represents the MLP network, represents the MobileNet layer capable of generating a series of image features, represents the ViewTransformer layer capable of converting the generated original image features into overhead view features.

[0054] The embedding features H i of the vehicle, the embedding features P j of the lane, and the embedding features U t of the vehicle multi-view overhead view are input into the multi-head attention network respectively to extract homologous interaction features P' j , H' i and U' t :

[0055] P' j= MultiAtt(P j ), H′ i = MultiAtt(H i ), U′ t = MultiAtt(U t ).

[0056] To further model the spatial heterogeneous interaction relationship between traffic participants and lanes, the working memory generator also employs a cross-attention network with residual connection. Then, the extracted features are normalized and input into the MLP layer to obtain the heterogeneous interaction features and The specific process is as follows:

[0057] P′ j = LayerNorm(CrossAtt(P′ j , H′ i ) + P′ j ),

[0058]

[0059] U′ t = LayerNorm(CrossAtt(U′ t , H′ i ) + U t ),

[0060]

[0061]

[0062] wherein LayerNorm(·) represents a layer normalization operation, and CrossAtt(·) represents a cross-attention mechanism.

[0063] The heterogeneous interaction features and are concatenated to obtain the aggregated features Z = {z1, …, z l}, l e {i, j).

[0064] The working memory formed by hippocampus in human driving process is not only limited to time domain dimension, but also can be extended to other dimensions such as frequency domain to support deeper reasoning process. Based on this, the Fourier GNN is introduced as a key module to enhance the environmental understanding ability of the autonomous vehicle. It should be emphasized that the technical concept of extending the modeling dimension of working memory proposed in the application has universality, and the modeling range can break through the frequency domain limit and extend to other representation spaces. In the application, the modeling method of working memory elements in the frequency domain space is mainly studied. By converting the time domain features to the frequency domain space, the working memory generator can obtain a global scene perspective, thereby realizing the simultaneous attention to all traffic participants and scene elements. In addition, the frequency domain representation method is beneficial to learning more discriminative motion patterns, and significantly improves the reasoning ability of complex dependence relationships in the driving scene.

[0065] Specifically, the application first constructs a full connection graph G, which is defined as:

[0066]

[0067] Wherein, is a node set, and represents the complex representation of the i th node;E={e1,e2,…,e L} is an edge set, and satisfies the condition E={1} L×L , wherein L is the number of agents and lanes.

[0068] In order to obtain The application projects and transforms the feature z by a full connection layer parameterized by a learnable parameter matrix W z The mathematical expression is defined as:

[0069]

[0070] Wherein, δ(·) is a linear transformation layer.

[0071] Further, the processed Is converted into frequency domain features by using discrete Fourier transform. The features of traffic participants and lanes in the frequency domain can be obtained in the following way:

[0072]

[0073] Wherein, F(·) represents discrete Fourier transform, represents the spectral component at frequency 2πq / L, and u is the imaginary unit. Through the above transformation, the complex feature tensor Can be obtained, wherein D is the dimension of the hidden layer.

[0074] ​Based on the full connection graph G, the application is in the Fourier space of a total of K layers of structure to the feature tensor With the Fourier graph operator A 0：k Recursive multiplication operation is performed, and each layer operation is identified by index k:

[0075]

[0076] In the formula, A=F(κ)∈C L×D×D , σ(·) is an activation function, b k is a learnable bias parameter, κ is a Green kernel function, satisfying κ: [L]×[L]→R D×D . Specifically, κ is derived from the weight matrix W∈R D×D , and satisfies the constraint condition And κ[i, j]: = κ[i-j].

[0077] In order to improve the stability of training, residual connection is used in each layer operation, which can effectively alleviate the gradient vanishing problem and speed up the iterative convergence process. In addition, in order to avoid the phenomenon of feature over-smoothing, the SoftShrink activation function is applied in the eigenvalue spectrum.

[0078] Finally, the application converts the frequency domain representation Q G back to the time domain space by inverse discrete Fourier transform:

[0079]

[0080] Where F -1 (·) represents the inverse discrete Fourier transform operation.

[0081] 2. Long-term memory constructor.

[0082] Another key innovation of the application is its driving experience learning ability, which is inspired by the observation that experienced human drivers can handle diverse scenarios with ease through accumulated skills. To achieve this goal, the application introduces a long-term memory constructor containing an iterative update mechanism and a collaborative regulation strategy. The long-term memory constructor models the long-term memory through an iterative process of extracting optimized representations from working memory, which is inspired by the feedback theory of Wiener control theory, which proposes that system stability can be achieved through feedback loops and has been successfully applied in the field of driver attention modeling. Following this paradigm, the application conceptualizes human driving behavior as a feedback regulation process, in which new observation information and prior experience are dynamically integrated to guide the decision-making process. Based on the above viewpoint, the application adopts an iterative optimization scheme driven by the cyclic feedback of working memory and the gradual update of long-term memory. This design supports continuous knowledge accumulation process, thereby significantly improving the performance of complex trajectory prediction tasks.

[0083] In the iterative updating mechanism, the application first randomly initializes the long-term memory knowledge S in the form of feature map. Through continuous learning, the long-term memory knowledge S is continuously optimized and updated in each forward propagation, and gradually evolves into a dynamic memory bank. The process can be mathematically expressed as: B a ←{S a-1 , B a-1}, wherein S a-1 represents the long-term memory knowledge at the (a-1)th iteration, B a-1 represents the feedback working memory feature at the (a-1)th iteration. It is worth noting that the iterative process of the application is carried out in the frequency domain space, because the frequency domain representation can more efficiently extract effective components and filter out redundant information from a global perspective.

[0084] The application first converts S a-1 and B a-1 to the frequency domain space through discrete Fourier transform:

[0085]

[0086] It should be particularly pointed out that not all working memory elements contribute to the formation of long-term memory. If all working memory features are directly included, it may introduce redundant components unrelated to driving experience, thereby weakening the effectiveness of the long-term memory model. Therefore, the application introduces a synergistic regulation strategy for purifying feedback features from working memory. Specifically, by introducing two learnable parameter matrices Ω s and Ω B , the original features are modulated to obtain more robust cues and Further, the modulated features are subjected to average pooling operation for dimension reduction, and then the synergistic modulation signal is generated through linear transformation. The above overall process can be formulated as:

[0087]

[0088] wherein η S represents the previous memory synergistic modulation signal, η B represents the synergistic modulation signal of the feedback feature, f linear (·) and f′ linear (·) represent linear layers, ⊙ represents element-wise multiplication, and AvgPool(·) represents average pooling operation.

[0089] Further, the application extracts the frequency spectrum feature through the common selection mechanism of the synergistic modulation signal:

[0090]

[0091] After feature selection, the present application uses inverse discrete Fourier transform to map back to time domain space: and

[0092]

[0093] Finally, the long-term memory representation of the a-th iteration is generated by concatenating and and applying MLP. The above iterative update mechanism can be formally expressed as:

[0094]

[0095] wherein, denotes the feature concatenation operation, f MLP (·) represents the MLP layer. In particular, a Dropout mechanism is introduced during the iteration process to alleviate the risk of overfitting that may be caused by the accumulation of long-term memory.

[0096] Through continuous iterative update, long-term memory gradually accumulates driving experience knowledge, and finally evolves into an incremental driving experience library. It is worth noting that this memory update process is only performed in the training stage, and S will remain frozen after the model is deployed, just like a human driver who uses existing experience to guide driving behavior after completing the learning stage. In the test stage, long-term memory is like a theoretical knowledge system of a skilled driver, providing experience support for decision-making in complex scenarios.

[0097] 3. Physical driving trajectory predictor.

[0098] After learning the human-like memory representation, the present application applies it to the physical driving trajectory predictor. In order to fully exploit the potential of working memory knowledge and long-term memory knowledge, the present application adopts a two-stage optimization strategy: first, use working memory knowledge to generate anchor trajectories, and then correct prediction bias through long-term memory knowledge. It is worth noting that in the anchor generation stage, long-term memory knowledge is not directly involved in prediction, but it is still continuously constructed and updated.

[0099] The first stage aims to generate anchor multi-modal trajectories. For each prediction instance, the physical driving trajectory predictor receives the features z generated by the working memory generator and as input, and uses the MLP layer to output kinematic control variables:

[0100]

[0101] wherein, v i(j+h) denotes the acceleration of agent i, γ i(j+h) denotes the steering angle, and t = j + h denotes a certain time in the future.

[0102] ​From which the x and y coordinate values of the agent i at the next time t = j + h + 1 can be derived:

[0103]

[0104] The state derivative in the formula can be derived from the following formula:

[0105]

[0106] In the formula, β represents the included angle between the speed v and the heading and l fi and l ri respectively represent the front and rear wheel distance of the vehicle, and the two values can be derived from the historical state.

[0107] The second stage aims to optimize the initial predicted trajectory using long-term memory knowledge, simulating the process of novice drivers continuously correcting prediction errors under the guidance of experience. The optimization predictor is composed of two MLP layers, the input includes long-term memory knowledge, working memory knowledge and complete trajectory features encoded by GRU and MLP, and the output is the prediction offset:

[0108] Δy0 = y0 - y'0,

[0109] In the formula, y'0 represents the real trajectory of the target agent, and y0 represents the initial predicted trajectory of the target agent.

[0110] The present application proposes an innovative vehicle trajectory prediction system, which integrates human-like working memory and long-term memory into the prediction process, enabling autonomous vehicles to deeply understand complex driving environments and continuously accumulate driving experience. Among them, the working memory is responsible for processing real-time interaction information in the current traffic scene, while the long-term memory simulates the decision-making process of human drivers by continuously accumulating and optimizing driving experience. At the same time, the present application models memory elements in the time domain and frequency domain, and captures the dynamic characteristics of the traffic environment more comprehensively, thereby enhancing the representation learning ability of the proposed framework. In order to obtain accumulated long-term memory, the present application proposes an efficient iterative updating mechanism, and combines a cooperative modulation strategy to ensure that the stored driving experience can be continuously improved as the driving scene changes. Moreover, the physical driving trajectory predictor proposed by the present application embeds physical knowledge into the trajectory predictor to generate more accurate and reliable future motion trajectories.

[0111] In addition, the present application also provides a vehicle trajectory prediction device based on human-like memory, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the human-like memory-based vehicle trajectory prediction method as described above.

[0112] Finally, the application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the steps of the vehicle trajectory prediction method based on human-like memory.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the scope of protection of the present application. Based on the idea of the present application, those skilled in the art can modify or equivalently replace the technical solutions of the present application without departing from the essence and scope of the present application.

Claims

1. A vehicle trajectory prediction method based on human-like memory mechanism, characterized in that, Includes the following steps: 1) Model the real-time interactive information in the current traffic scenario in the time and frequency domains using a working memory generator to obtain real-time working memory knowledge; 2) Information is retrieved from working memory through a long-term memory builder, and through an iterative update mechanism, driving experience is continuously accumulated to acquire long-term memory knowledge; 3) Real-time working memory knowledge and long-term memory knowledge are input into the physical-driven trajectory predictor to generate future motion trajectories.

2. The vehicle trajectory prediction method based on human-like memory mechanism according to claim 1, characterized in that, Step 1) specifically includes: 11) Based on the vehicle's trajectory vector V i Lane vector M j Raw multi-view images acquired by vehicle-mounted cameras Obtain the embedding features H of the vehicle trajectory i Lane embedding features P j And the embedded features U of the multi-view bird's-eye view of the vehicle t And based on H i P j and U t Obtain the homology interaction feature H′ i 、P′ j and U′ t ; 12) Based on the aforementioned homologous interaction feature H′ i and P′ j and U′ t Obtain heterogeneous interaction features and 13) The heterogeneous interaction features and The aggregated feature Z is obtained by concatenating the features. 14) Perform time-domain frequency conversion based on the aggregated feature Z to obtain real-time working memory knowledge.

3. The vehicle trajectory prediction method based on human-like memory mechanism according to claim 1, characterized in that, Step 2) specifically includes: 21) Randomly initialize long-term memory knowledge S in the form of feature maps; 22) Through continuous learning, the long-term memory knowledge S is continuously optimized and updated in each forward propagation, gradually evolving into a dynamic memory bank, which contains the long-term memory knowledge at each forward propagation and the feedback working memory features at each forward propagation. 23) Transfer the long-term memory knowledge S from the (a-1)th forward propagation. a-1 and the feedback working memory feature B during the (a-1)th forward propagation a-1 Transform to the frequency domain to obtain and 24) Long-term memory knowledge converted to the frequency domain and feedback working memory characteristics Modulation is performed to obtain more robust cues. and 25) By splicing and And apply MLP to generate long-term memory knowledge for the a-th iteration.

4. The vehicle trajectory prediction method based on human-like memory mechanism according to claim 1, characterized in that, Step 3) specifically includes: 31) Generate anchor point multimodal trajectories using the aforementioned real-time working memory knowledge; 32) Optimize the anchor point multimodal trajectory using the long-term memory knowledge to correct the anchor point multimodal trajectory, thereby obtaining the future motion trajectory.

5. A vehicle trajectory prediction system based on a human-like memory mechanism, characterized in that, include: A working memory generator is used to model real-time interactive information in the current traffic scenario in the time and frequency domains to obtain real-time working memory knowledge. Long-term memory builder, which retrieves information from working memory and, through an iterative update mechanism, ultimately forms continuously accumulating driving experience to acquire long-term memory knowledge; A physics-driven trajectory predictor is used to generate future motion trajectories based on the real-time working memory knowledge and long-term memory knowledge.

6. A vehicle trajectory prediction device based on a human-like memory mechanism, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the vehicle trajectory prediction method based on human-like memory mechanism as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the vehicle trajectory prediction method based on human-like memory mechanism as described in any one of claims 1-4.

Images