Intelligent agent trajectory prediction method in unstructured environment in field of automatic driving

Abstract

The invention discloses an intelligent agent trajectory prediction method in an unstructured environment in the field of automatic driving, and the method comprises the steps: S1, carrying out the data preprocessing and coding of a historical trajectory, carrying out the multi-scale one-dimensional convolution operation of the historical trajectory data of an intelligent agent, and enabling a processed trajectory sequence to generate a hidden state; s2, interactive modeling is carried out, and interactive features of a target agent are generated through a distance-orientation perception filtering method; and S3, decoding the future trajectory, initializing the hidden state of a decoder, and predicting the future state of the target agent. According to the method, the interaction agent can be accurately identified, the directional interaction features are efficiently extracted, the model robustness is enhanced, and lightweight deployment is realized.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a method for predicting the trajectory of an intelligent agent in an unstructured environment in the field of autonomous driving. Background Technology

[0002] In the field of autonomous driving, determining which agents interact with the target agent is crucial for trajectory prediction. Existing methods are mainly divided into three categories: grid-based methods, graph-based methods, and gridless methods.

[0003] 1) Grid-based methods: These methods establish a spatial grid around the target agent, defined by a specific distance resolution, and map neighboring agents to corresponding grid cells based on their spatial relationships with the target agent. This representation restricts the area of ​​interest to a fixed and predefined range and limits the spatial resolution to the granularity of the grid cells. For example, Social LSTM constructs a grid for the target pedestrian, mapping neighboring pedestrians to corresponding grid cells based on their relative positions, while considering neighbors from all directions, resulting in high computational complexity. PV-CollisionGrid introduces a collision time metric in an unstructured environment primarily composed of pedestrians but also including a limited number of vehicles and cyclists. To address the diversity of agent types, PV-CollisionGrid constructs different grids for pedestrian-pedestrian interactions and vehicle-pedestrian interactions, but does not explicitly model cyclist-pedestrian interactions.

[0004] 2) Graph-based methods: These methods represent agents as nodes and their interactions as edges, allowing the application of Graph Neural Network (GNN) operations. While graph-based methods effectively preserve detailed information about interactions between agents, they can lead to high computational complexity when dealing with a large number of neighboring agents. Some studies address this challenge by employing rule-based strategies (such as distance thresholds or map prior knowledge) to prune edges.

[0005] 3) Meshless methods: Unlike mesh-based and graph-based methods, meshless methods do not rely on a fixed mesh structure but treat all agents as potential interactive entities. Social GAN ​​simulates social interaction behavior between pedestrians through a generator-discriminator adversarial mechanism, and can simultaneously predict the future trajectories of multiple pedestrians, explicitly modeling the mutual influence between them; however, it has high computational complexity, a large number of model parameters, and a strong dependence on high-quality data.

[0006] Therefore, how to establish a method for predicting the trajectory of intelligent agents in unstructured environments in the field of autonomous driving has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method for predicting the trajectory of intelligent agents in unstructured environments in the field of autonomous driving, which can accurately identify interactive intelligent agents, efficiently extract directional interaction features, enhance model robustness, and achieve lightweight deployment.

[0008] The present invention solves the technical problem by adopting the following technical solution:

[0009] A method for predicting agent trajectories in unstructured environments for autonomous driving includes the following steps:

[0010] Step S1, data preprocessing and encoding historical trajectories: first, the historical trajectory data of the agent is processed using multi-scale one-dimensional convolution operation, and then the processed trajectory sequence is used to generate hidden states.

[0011] Step S2, interaction modeling, generates the interaction features of the target agent through a distance-orientation awareness filtering method;

[0012] Step S3: Decode the future trajectory, initialize the hidden state of the decoder, and predict the future state of the target agent.

[0013] Furthermore, in step S1, the specific methods for data preprocessing and encoding historical trajectories are as follows:

[0014] The historical trajectory data of agent A is embedded using multi-scale temporal convolution along the time dimension to achieve smooth trajectory and reduce high-frequency noise, outputting a smooth feature e:

[0015] e = MTCE(x) + CE(c) (1)

[0016] Where MTCE(x) is a multi-scale temporal convolutional embedding function module, and CE(c) represents the category embedding function for agent category c;

[0017] The processed trajectory sequence is then input into a gated recurrent unit to generate a hidden state h, which is used to capture the temporal dynamic features of trajectory prediction.

[0018] Furthermore, in step S2, the specific method for interaction modeling is as follows:

[0019] Based on the relative raw features of all agents A relative to the target agent TA in the scene snapshot of the last observation time step, all agents that have interacted are selected through a distance-orientation-aware filtering method. An interaction feature map of the target agent TA is then constructed, and its interaction features are generated through distance-orientation convolutional pooling. Finally, the interaction features are added to the hidden state h encoded in step S1, resulting in:

[0020] Furthermore, the distance-orientation sensing filtering method is as follows:

[0021] In unstructured environments, trajectory prediction of traffic participants requires accurate identification of interactive agents (IAs) around the target agent (TA) that may affect its motion. A dynamic filtering mechanism based on distance and azimuth angle is proposed, which retains only interactive agents (IAs) within a specific distance and angle range in front of the target agent (TA) and ignores irrelevant agents. A fan-shaped region is constructed based on the position and heading angle of the target agent (TA).

[0022] Filtering criteria: Maximum interaction threshold range D, azimuth range β, formula:

[0023]

[0024] Where IA represents an agent that has a certain interactive influence on the target agent TA; a represents a specific agent index; OA represents all other agents in the scene besides the target agent TA; p a and p TA θ represents the position coordinates of agents a and TA, respectively; ||·|| is the Euclidean distance; ∠(·) is used to calculate the angle between the two vectors; g(θ) TA )=(cosθ TA sinθ TA ) is the heading angle θ of TA. TA Convert to a unit vector; D is the maximum interaction threshold range; β is the azimuth range.

[0025] Furthermore, distance-azimuth convolutional pooling includes:

[0026] Distance axis convolution (result of layer l): Slide a one-dimensional convolution kernel along the distance dimension to capture the continuity of close-range interactions;

[0027]

[0028] In this step, The above l-1 layer convolution result S (l-1) The distance axis convolution result calculated for the parameters, S (l-1) This is the result of a (l-1) layer convolution; the size of the convolution kernel k is (k d ,1), indicating that the convolution operation is performed only on the distance axis; the stride p is (2,1), indicating that the movement is 2 units on the distance axis and 1 unit on the azimuth axis; the padding mode s is (1,0), which means that 1 pixel is padded on the top side of the azimuth axis.

[0029] Azimuth axis convolution (result of layer l): Slide a one-dimensional convolution kernel along the azimuth dimension to model directional dependence;

[0030]

[0031] In this step, S (l) The result of convolution with the l-th distance axis The convolution result of this layer is calculated using the parameters; the size of the convolution kernel k is... This indicates that the convolution operation is performed only on the azimuth axis; the stride p is (1,2), which means that the movement is 1 unit on the distance axis and 2 units on the azimuth axis; the padding s is (0,1), which means that 1 pixel is padded on the top of the azimuth axis.

[0032] Feature aggregation: Compact interactive feature vectors are generated by compressing the spatial dimension through global average pooling.

[0033] Furthermore, multi-scale temporal convolutional embeddings include:

[0034]

[0035] Input consists of historical trajectory sequences Includes T h Two-dimensional coordinates with a time step; then perform dual-path convolution: short-time convolution kernel k=3, used to capture local motion details, stride p=1 to maintain sequence length; long-time convolution kernel k=5, used to extract global trends, stride p=2 sequence; then concatenate the two features to form a multi-scale temporal representation and output it;

[0036] Simultaneously, multi-scale temporal convolutional embeddings are combined with class embeddings to generate the final smooth features:

[0037] e = MTCE(x) + CE(c) (9).

[0038] Furthermore, in step S3, the specific method for decoding the future trajectory is as follows:

[0039] First, initialize the hidden state of the decoder as follows: Predict the future state of the target agent TA within a time step from 0 to t:

[0040]

[0041] Where, φ s It is a multilayer perceptron module. It is the predicted displacement, s t Indicates spatial embedding, Let t be the state of the decoder. Let W be the state of the decoder at time t-1. d1 and W d2 Represents the learnable weights;

[0042] The output distribution follows a double Gaussian variable distribution, with the mean parameter being... The variance and covariance parameters are These parameters are obtained through a multilayer perceptron module that operates on each hidden state:

[0043]

[0044] Where, φ o It is a multilayer perceptron module, W o These are learnable weights.

[0045] This invention discloses a method for predicting the trajectory of an intelligent agent in an unstructured environment in the field of autonomous driving, which has the following beneficial effects:

[0046] This invention addresses the following four main problems in the intelligent agent trajectory prediction framework.

[0047] 1. Accurately identify interactive intelligent agents (IA): Combine distance and azimuth to dynamically filter relevant interactive objects and simulate human decision-making logic.

[0048] 2. Efficient extraction of directional interaction features: Asymmetric convolution separates the distance / orientation dimension, reducing the number of parameters and improving feature representation ability.

[0049] 3. Enhance model robustness: Design a multi-scale temporal convolution module (MTCE) to mitigate the impact of sensor failure or data loss.

[0050] 4. Achieve lightweight deployment: Modular design is compatible with low computing power platforms, and the number of parameters is controlled at 329.9K. Attached Figure Description

[0051] Figure 1 This is a modular encoder-interaction-decoder framework diagram.

[0052] Figure 2 This is a schematic diagram of interactive intelligent agent filtering.

[0053] Figure 3 This is a schematic diagram of the proposed DA-ConvPool module.

[0054] Figure 4 These are trajectory prediction performance curves under different levels of data missing. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] This invention provides a method for predicting the trajectory of an intelligent agent in an unstructured environment in the field of autonomous driving, comprising the following steps:

[0058] Step S1, data preprocessing and encoding historical trajectories: first, the historical trajectory data of the agent is processed using multi-scale one-dimensional convolution operation, and then the processed trajectory sequence is used to generate hidden states.

[0059] The specific methods for data preprocessing and encoding historical trajectories are as follows:

[0060] The historical trajectory data of agent A is processed using multi-scale temporal convolutional embedding (MTCE) along the time dimension to smooth the trajectory and reduce high-frequency noise, outputting a smooth feature e:

[0061] e = MTCE(x) + CE(c) (1)

[0062] Where MTCE(x) is the Multiscale Temporal Convolutional Embedding module, and CE(c) represents the category embedding function for agent class c;

[0063] The processed trajectory sequence is then input into a gated recurrent unit (GRU) to generate a hidden state h, which is used to capture the temporal dynamic features of trajectory prediction.

[0064] Step S2, interaction modeling, generates the interaction features of the target agent through a distance-orientation awareness filtering method;

[0065] The specific methods for interaction modeling are as follows:

[0066] Based on the relative raw features (e.g., position and heading angle) of all agents A relative to the target agent TA in the scene snapshot of the last observation time step, all agents (IAs) that have interacted are selected using the distance-aware-fiber (IAF) filtering method. Then, an interaction feature map of the target agent TA is constructed, and the interaction features of the target agent TA are generated through distance-aware-fiber convolutional pooling (DA-ConvPool). Finally, the interaction features are added to the hidden state h encoded in step S1, resulting in...

[0067] This invention proposes a dynamic filtering mechanism based on distance and azimuth angle, retaining only agents within a specific distance and angle range in front of the target agent (TA), while ignoring irrelevant agents. By using distance-azimuth constraints, redundant computations are avoided (e.g., filtering out agents without interaction behind or to the side), computational efficiency is improved, and irrelevant interference is reduced, thus enabling priority processing of nearby agents. The distance-azimuth perception filtering method is as follows:

[0068] In unstructured environments, trajectory prediction of traffic participants (such as pedestrians, vehicles, and bicycles) requires accurate identification of interactive agents (IAs) around the target agent (TA) that may influence its motion. Traditional methods typically rely on fixed-distance or omnidirectional filtering, which is inefficient and may introduce irrelevant interference. This proposal suggests a dynamic filtering mechanism based on distance and azimuth angle, retaining only interactive agents (IAs) within a specific distance and angle range in front of the target agent (TA) while ignoring irrelevant agents. Figure 2 As shown, a fan-shaped region is constructed based on the target agent's TA position and heading angle;

[0069] Filtering criteria: Maximum interaction threshold range D (default 10m), azimuth range β (default 90°), formula:

[0070]

[0071] Where IA (Interacting Agents) refers to agents that have a certain interactive influence on the target agent TA; a is the specific agent index; OA (Other Agents) refers to all other agents in the scene besides the target agent TA; p a and p TA θ represents the position coordinates of agents a and TA, respectively; ||·|| is the Euclidean distance; ∠(·) is used to calculate the angle between the two vectors; g(θ) TA )=(cosθ TA sinθ TA ) is the heading angle θ of TA. TA Convert to a unit vector; D is the maximum interaction threshold range; β is the azimuth range (e.g., 90°).

[0072] Advantages: Eliminating lateral / rear interference, as shown in Table 1, our method demonstrates superior performance across all evaluation metrics. Specifically, among the pedestrian categories that constitute the majority of the dataset, our method achieves state-of-the-art results in both Average Displacement Error (ADE) and Final Displacement Error (FDE), outperforming the suboptimal baseline method PV-CollisionGrid by 21.0% and 12.7%, respectively. Even without distinguishing agent categories, our method achieves state-of-the-art overall performance, exceeding the suboptimal baseline SocialVAE by 1.0% in ADE and by 1.2% in FDE. These results demonstrate the significant effectiveness and robustness of our method in modeling interactive behavior and predicting trajectories in unstructured environments.

[0073] Table 1 Performance comparison of the HBS dataset under deterministic strategies.

[0074]

[0075] Quantitative evaluation was conducted using Table 1; the evaluation index was ADE / FDE (the lower the better).

[0076] DA-ConvPool, for example Figure 3 As shown, this is an interaction feature extraction module for multi-agent trajectory prediction. It extracts directional interaction features along the distance and azimuth axes through asymmetric convolution operations, and finally generates the interaction features of the target agent (TA) through global average pooling (GAP). This method decouples distance and azimuth features, improving the ability to model spatial heterogeneous interactions. It reduces the number of convolution parameters to 33.4% of the Sym-Conv method, while reducing ADE by 1.7%, as shown in Table 2, which compares experimental results with the number of parameters.

[0077] Table 2 Comparison of performance and parameter count for different pooling operations (HBS dataset)

[0078]

[0079] Distance-orientation convolutional pooling includes:

[0080] Distance-axis convolution (result of layer l): Slide a one-dimensional convolution kernel (size k) along the distance dimension (vertical direction). d *1) Capture the continuity of close-range interactions (such as the risk of sudden braking) to obtain the convolution result.

[0081]

[0082] In this step, The result S of the (l-1) convolution layer above. (l-1)The distance axis convolution result calculated for the parameters, S (l-1) This is the result of a (l-1) layer convolution; the size of the convolution kernel k is (k d ,1), indicating that the convolution operation is performed only on the distance axis; the stride p is (2,1), indicating that the movement is 2 units on the distance axis and 1 unit on the azimuth axis; the padding mode s is (1,0), which means that 1 pixel is padded on the top side of the azimuth axis.

[0083] Azimuth axis convolution (result of layer l): Slide a one-dimensional convolution kernel (size) along the azimuth dimension (horizontal direction). Modeling directional dependencies (such as the intention to overtake laterally);

[0084]

[0085] In this step, S (l) The result of convolution with the l-th distance axis The parameter is used to calculate the convolution result of this layer; the size of the convolution kernel k is... This indicates that the convolution operation is performed only on the azimuth axis; the stride p is (1,2), which means that the movement is 1 unit on the distance axis and 2 units on the azimuth axis; the padding s is (0,1), which means that 1 pixel is padded on the top of the azimuth axis.

[0086] Feature aggregation: Compact interactive feature vectors are generated by compressing the spatial dimension through global average pooling.

[0087] Advantages: As shown in Table 2, the proposed DA-Conv achieves the best performance compared to the No-Conv configuration, improving ADE and FDE by 1.7% and 1.9%, respectively. Furthermore, it reduces the number of parameters by 66.6% compared to the Sym-Conv configuration. These findings highlight the effectiveness of the proposed asymmetric convolution in capturing subtle spatial dependencies, thereby improving trajectory prediction accuracy. In addition, the parameter efficiency of DA-Conv makes it particularly suitable for small datasets, achieving a balance between performance and computational cost.

[0088] Multi-scale temporal convolutional embeddings include:

[0089] MTCE smooths and extracts features from historical trajectory data through multi-scale convolutional operations. Specifically, it mainly uses parallel 1D convolutional layers of different scales to capture local short-term dynamics (such as acceleration changes) and long-term trends (such as motion direction stability) in the trajectory data, thereby enhancing the robustness of the trajectory prediction model. Its core design includes:

[0090]

[0091] Input consists of historical trajectory sequences Includes T h Two-dimensional coordinates with a time step; then perform dual-path convolution: short-time convolution kernel k=3, used to capture local motion details (such as instantaneous velocity changes), stride p=1 to maintain sequence length; long-time convolution kernel k=5, used to extract global trends (such as turning intentions), stride p=2 sequence; then concatenate the two features (using the Concat function) to form a multi-scale temporal representation and output it;

[0092] Simultaneously, a multi-scale temporal convolutional embedding encoder is combined with class embedding (CE) to generate the final smooth features:

[0093] e = MTCE(x) + CE(c) (9).

[0094] Advantages: It performs one-dimensional convolution only on two paths, featuring a lightweight design, high computational efficiency, and support for deployment. When agent category is not distinguished, the introduction of MTCE improves the average displacement error (ADE) and final displacement error (FDE) by 1.7% and 1.3%, respectively, as shown in Table 3, with particularly significant performance improvements for pedestrians and vehicles. However, the prediction performance for cyclists is slightly lower, possibly due to more pronounced data loss in cyclist trajectories compared to other categories.

[0095] Table 3 Quantitative analysis of the impact of MTCE on incomplete trajectories

[0096]

[0097] The evaluation metric in Table 3 is ADE / FDE, and the lower the value, the better.

[0098] MTCE smooths and extracts features from historical trajectory data through multi-scale convolutional operations. Specifically, it mainly uses parallel 1D convolutional layers of different scales to capture local short-term dynamics (such as acceleration changes) and long-term trends (such as motion direction stability) in the trajectory data, enhancing robustness to sensor failures or data loss. When the masking ratio (data loss rate) increases to 40%, ADE (average displacement error) and FDE (final displacement error) only show gradual degradation; after exceeding this threshold, performance drops sharply—at a 50% masking ratio, ADE increases by 33.6% and FDE increases by 25.9%, as shown below. Figure 4 As shown, Figure 4 The horizontal axis represents the percentage of missing data, and the vertical axis displays the ADE and FDE indicators; Figure 4 (a) indicates the trend of ADE. Figure 4 (b) indicates the trend of FDE change.

[0099] Step S3: Decode the future trajectory, initialize the hidden state of the decoder, and predict the future state of the target agent.

[0100] In step S3, the specific method for decoding the future trajectory is as follows:

[0101] First, initialize the hidden state of the decoder as follows: Predict the future state of the target agent TA within a time step from 0 to t:

[0102]

[0103] Where, φ s It is a multilayer perceptron (MLP) module. It is the predicted displacement, s t Indicates spatial embedding, Let t be the state of the decoder. Let W be the state of the decoder at time t-1. d1 and W d2 Represents the learnable weights;

[0104] Finally, we assume that the output distribution follows a bigaussian variable distribution with mean parameters as follows: The variance and covariance parameters are These parameters are obtained through a multilayer perceptron module that operates on each hidden state:

[0105]

[0106] Where, φ o It is a multilayer perceptron (MLP) module, W o These are learnable weights.

[0107] The core architecture of this invention is a modular encoder-interaction-decoder framework.

[0108] like Figure 1 As shown, the proposed framework employs a concise and efficient encoder-interaction-decoder framework, enhanced by a modular global interaction modeling layer. This design offers flexible deployment capabilities. The interaction modeling layer, as a plug-and-play component, can be easily replaced or updated according to the offline or online requirements of different scenarios.

[0109] The present invention can achieve the following effects:

[0110] 1. Improved accuracy:

[0111] HBS dataset: As shown in Table 1, the average pedestrian displacement error is 0.233 (a 21% decrease compared to the PV-CollisionGrid method), and the final displacement error is 0.566 (a 12.7% decrease compared to the PV-CollisionGrid method).

[0112] SDD dataset: As shown in Table 5, two sampling strategies were employed in the tests: sample size = 1 and sample size = 20. As illustrated, although our method did not achieve optimal performance on this dataset, it was competitive with the listed baselines. This highlights the robustness and adaptability of our method across different datasets.

[0113] Table 5 Performance Comparison of SDD Dataset (Quantitative Evaluation under Two Sampling Strategies)

[0114]

[0115] 2. Robustness:

[0116] Table 4 summarizes the experimental results. The model exhibits strong robustness: when the masking ratio (input data loss rate) increases to 40%, ADE (mean displacement error) and FDE (final displacement error) only show gradual degradation; performance drops sharply after exceeding this threshold—ADE increases by 33.6% and FDE increases by 25.9% at a masking ratio of 50%. Figure 3 The visualization presents this trend, revealing the critical threshold at which prediction accuracy is significantly compromised.

[0117] Table 4. Trajectory prediction performance (ADE / FDE) under different data loss rates

[0118]

[0119] 3. Computational efficiency:

[0120] By filtering interactive agents using distance-orientation awareness (IAF), redundant computations are reduced, and the number of parameters is reduced by 66.6% compared to the Sym-Conv method, as shown in Table 2.

[0121] 4. Low computing power requirement:

[0122] The modular design is adapted to low-computing-power automotive platforms (with measured inference latency of <10ms on the Nvidia 4070Ti platform).

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for trajectory prediction of an agent in an unstructured environment in the field of autonomous driving, the method comprising: The method comprises the following steps: Step S1, data preprocessing and encoding historical trajectory, first using multi-scale one-dimensional convolution operation on the historical trajectory data of the agent, and then generating a hidden state from the processed trajectory sequence; Step S2, interaction modeling, generating the interaction features of the target agent by a distance-azimuth perception filtering method; Step S3, decoding future trajectory, initializing the hidden state of the decoder, and predicting the future state of the target agent. 2.The trajectory prediction method for an agent in an unstructured environment in an autonomous driving field according to claim 1, wherein, In step S1, the specific method of data preprocessing and encoding historical trajectory is: The historical trajectory data of the agent A is embedded by using multi-scale time convolution along the time dimension to smooth the trajectory and reduce high-frequency noise, and the smoothed feature e is output: e =MTCE(x)+CE(c) (1) Wherein, MTCE(x) is a multi-scale time convolution embedding function module, and CE(c) represents a category embedding function for the agent category c; The processed trajectory sequence is then input into a gated recurrent unit to generate a hidden state h, which is used to capture the time dynamic characteristics of trajectory prediction.

3. The trajectory prediction method for an agent in an unstructured environment in the field of autonomous driving according to claim 2, characterized in that, In step S2, the specific method of interaction modeling is: According to the relative original features of all agents A in the scene snapshot at the last observation time step relative to the target agent TA, all agents that interact are filtered out from the relative original features by a distance-orientation perception filtering method, and an interaction feature map of the target agent TA is constructed and generated, and an interaction feature of the target agent TA is generated by distance-orientation convolution pooling. Then, the interaction feature is added to the hidden state h after encoding in step S1, and the result is 4. The trajectory prediction method for an agent in an unstructured environment in the field of autonomous driving according to claim 3, characterized in that, The distance-azimuth perception filtering method is: In an unstructured environment, the trajectory prediction of a traffic participant needs to accurately identify the interactive agents IAs around the target agent TA that may affect its movement; a dynamic screening mechanism based on distance and azimuth angle is proposed, which only retains the interactive agents IAs within a certain distance and angle range in front of the TA, and ignores irrelevant agents; A fan-shaped region is constructed based on the position and heading angle of the target agent TA; The screening conditions are: maximum interaction threshold range D, azimuth angle range β, formula: Wherein, IA is the intelligent agent with certain interaction influence on the target intelligent agent TA; a is the specific intelligent agent index; OA is all other intelligent agents in the scene except the target intelligent agent TA; p a and p TA are the position coordinates of the intelligent agent a and the intelligent agent TA respectively; ||·|| is the Euclidean distance; ∠(·) is the calculation of the angle between two vectors; g(θ TA )=(cosθ TA , sinθ TA ) is the conversion of the heading angle θ TA of TA into a unit vector; D is the maximum interaction threshold range; β is the azimuth angle range.

5. The trajectory prediction method for an agent in an unstructured environment in the field of autonomous driving according to claim 4, characterized in that, The distance-azimuth convolution pooling includes: Distance axis convolution (the result of the lth layer): sliding a one-dimensional convolution kernel along the distance dimension to capture the continuity of close-range interaction; In this step, S is the (l-1) layer convolution result; the size of the convolution kernel k is (k (l-1) S is the distance axis convolution result calculated by the parameters (l-1) S is the (l-1) layer convolution result; the size of the convolution kernel k is (k d ,1), which means that the convolution operation is only performed on the distance axis; the stride p is (2,1), which means that every time the distance axis is moved by 2 units and the azimuth axis is moved by 1 unit; the padding s is (1,0), that is, 1 pixel is filled on the upper side of the azimuth axis; Azimuth axis convolution (the result of the lth layer): sliding a one-dimensional convolution kernel along the azimuth dimension to model the directional dependence; In this step, S (l) is the convolution result of the lth layer with the distance axis is the convolution result of the lth layer calculated by parameters; the size of the convolution kernel k is (1, ), which means that the convolution operation is only performed on the azimuth axis; the step p is (1, 2), which means that 1 unit is moved on the distance axis and 2 units are moved on the azimuth axis each time; the padding s is (0, 1), that is, 1 pixel is filled on the upper side of the azimuth axis; Feature aggregation: compress the spatial dimension by global average pooling to generate a compact interaction feature vector.

6. The trajectory prediction method for an agent in an unstructured environment in the field of autonomous driving according to claim 5, characterized in that, The multi-scale time convolution embedding includes: Input is a sequence of historical trajectories Contains T h two-dimensional coordinates at each time step; then a two-path convolution is performed: a short-time convolution kernel k=3 is used to capture local motion details, and a stride p=1 is used to maintain the sequence length; a long-time convolution kernel k=5 is used to extract global trends, and a stride p=2 is used to sequence; finally, the two-path features are concatenated to form a multi-scale time series representation and output. At the same time, the multi-scale time convolution embedding encoder is combined with the category embedding to generate the final smoothed feature: e=MTCE(x)+CE(c) (9).

7. The trajectory prediction method for an agent in an unstructured environment in the field of autonomous driving according to claim 6, characterized in that, In step S3, the specific method of decoding future trajectory is: The hidden state of the decoder is first initialized as The future state of the target agent TA is predicted over time steps 0 to t: where φ s is a multi-layer perceptron module, is a predicted displacement amount, s t denotes a spatial embedding, is a state of the decoder at time t, is a state of the decoder at time t-1, W d1 and W d2 denote learnable weights; The output distribution follows a bivariate Gaussian distribution with mean parameters The variance and covariance parameters are These parameters are obtained through a multi-layer perceptron module acting on each step hidden state: where φ o is a multi-layer perceptron module, W o are learnable weights.

Images