A multi-agent intention recognition method and system under the condition of multi-source environmental information incompleteness

Abstract

The application discloses a multi-source environmental information incomplete condition multi-agent intention recognition method and system, relates to the multi-agent intention recognition technical field.The application is proposed in order to solve the problems of the adaptability of the existing method to the dynamic target in the complex scene or dynamic environment, the lack of dynamic weight adjustment mechanism affecting the situation interaction between the obstacles and the agent and the like.The method of the application encodes the interactive relationship of the agent and the influence of the environmental obstacles into spatial features by fusing the motion situation graph and the blocking situation graph, compared with the original trajectory input, more directly reflects the tactical game logic in the dynamic environment, so that the recognition accuracy is improved; the spatial structural characteristics of the situation graph avoid the interference of noise in the trajectory data, and the model robustness is enhanced.The application captures the local motion mutation caused by obstacle avoidance through an LSTM layer, and combines a Transformer encoder to model the correlation between the obstacle distribution and the global intention, so that the comprehensive performance is optimal in the complex scene.Experiments show that the accuracy of the method of the application in the obstacle-containing scene is higher than that of the prior art method model, and the technical advantages of the collaborative action of the situation graph and the hybrid model of the application are verified, so that a reliable solution is provided for the complex confrontation scene.

Description

Technical Field

[0001] This invention relates to the field of multi-agent intent recognition technology, and more specifically, to a method and system for multi-agent intent recognition under conditions of incomplete multi-source environmental information based on situational graph sequences and hybrid networks. Background Technology

[0002] With the widespread application of multi-agent systems, the scenarios of collaboration and confrontation between agents are becoming increasingly complex. In real-world scenarios such as drone swarm collaborative reconnaissance, autonomous vehicle fleet formation control, intelligent warehouse robot collaborative handling, and military tactical games, agents need to perceive, make decisions, and interact in real time within dynamic environments. Therefore, accurately identifying the intentions of agents has become crucial for achieving efficient decision-making and autonomous control. For example, in autonomous driving, predicting the lane-changing intentions of adjacent vehicles can help plan safe paths in advance; in military confrontation, identifying the encirclement strategy of enemy drones can trigger targeted countermeasures; and in smart factories, predicting the target workstation of handling robots can optimize overall scheduling efficiency.

[0003] Recently, research on multi-agent intent recognition has been increasing. For example, existing patent document CN119621916B provides a method and system for identifying implicit intent in open-ended questions based on multi-agent reasoning. Its technical solution involves constructing multilingual background knowledge using a multilingual translation agent and an information retrieval agent; performing topic analysis based on the constructed multilingual background knowledge to generate multi-domain expert agent information and dynamically constructing multi-domain expert agents; generating implicit intent and reasoning text content based on the guiding question posed by the moderator agent, with each multi-domain expert agent speaking according to the moderator agent's guiding question; and using the moderator agent to perform quality evaluation on the implicit intent and reasoning text content, treating the speech of each selected expert agent as a node to construct an implicit intent database. This effectively addresses the shortcomings in handling complex semantic understanding and more accurately and comprehensively identifies implicit intent in open-ended questions.

[0004] Existing patent document CN116757249A discloses a method for identifying the strategic intent of UAV swarms based on distributed reinforcement learning. This prior art aims to achieve efficient collaborative action among UAV swarms by focusing on competitive and cooperative relationships, human intent reasoning, and decision-making. Through training and optimization using deep neural networks and reinforcement learning algorithms, it can accurately identify the intent of individuals within the UAV swarm and improve decision-making and execution capabilities. It employs an asynchronous, multi-threaded training method to improve learning efficiency and performance. Flexible reward functions and state space definitions are used to adapt to different task requirements. Learning progress is monitored by evaluating indicators such as the UAV's intent learning rate, cumulative reward value, and the convergence speed of the learning algorithm. It is applicable to fields such as military operations and rescue operations, and has broad application value in UAV swarm communication, coordination, and resource sharing, providing important technical support for the intelligent development of UAV swarms.

[0005] However, existing technologies have significant limitations:

[0006] Insufficient adaptability to dynamic environments: Traditional intent recognition methods (based on rule reasoning or single sensor data) are difficult to effectively handle problems such as dynamic obstacles, sensor noise and missing real-time data. In complex scenarios, the irregular distribution of obstacles and sudden changes in the movement of the agent may cause gradient information loss or singular value problems in the situation map generated by the traditional artificial potential field method, thus affecting the accuracy of intent judgment.

[0007] Multi-source data fusion is inefficient: Existing methods typically overlay motion situation maps of multiple agents directly. However, when there is a significant difference in the number of agents on both sides, the situation information of a minority group is easily overwhelmed, resulting in the fused global situation map failing to reflect the behavioral characteristics of both sides in a balanced manner. Furthermore, the lack of a dynamic weight adjustment mechanism for the interaction between obstacles and agents further reduces the reliability of the fusion results.

[0008] Poor robustness in scenarios with incomplete information: When trajectories are partially missing due to occlusion or sensor failure, traditional models struggle to balance local temporal abrupt changes with long-distance dependencies. For example, while LSTM can capture short-term obstacle avoidance actions, it cannot effectively correlate global intent logic across discontinuous time steps; and while Transformer excels at modeling long-sequence dependencies, it lacks sensitivity to local dynamic changes, leading to a significant drop in recognition accuracy in scenarios with missing data.

[0009] Therefore, there is an urgent need for a new method that integrates dynamic environmental situation modeling, multi-source information compensation, and hybrid deep learning to overcome the limitations of existing technologies. Summary of the Invention

[0010] The technical problem to be solved by this invention:

[0011] To address the shortcomings of existing methods in adapting to dynamic targets in complex or dynamic environments, the lack of a dynamic weight adjustment mechanism affecting the situational interaction between obstacles and agents, the reduced reliability of fusion results when the number of agents on both sides differs significantly, and the lack of consideration for scenarios with incomplete information such as missing trajectories due to occlusion or sensor failure, this invention provides a method and system for multi-agent intent recognition under conditions of incomplete multi-source environmental information.

[0012] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0013] To address the aforementioned technical problems, this invention provides a method for multi-agent intent recognition under conditions of incomplete multi-source environmental information, comprising the following steps:

[0014] Step S1: Generation of single agent motion situation diagram: Based on the artificial potential field method, generate gravitational field and repulsive field for the target unit and the responding unit respectively;

[0015] The formula for calculating the unit repulsive field is as follows: Where η i Let q be the scale factor of the repulsive / gravitational field of the i-th agent. a Let ρ(q,q) be the position of the agent. a Let η be the distance between the point to be determined and the agent, μ be the scale parameter controlling the attenuation range of the potential field, and the magnitude of the repulsive field be limited to (0, η). i ];

[0016] The formula for calculating the target unit gravitational field is as follows:

[0017] The scale factor η i Through the mapping function α(P,S) i O) is dynamically determined, where P is the agent attribute, and S is the agent attribute. i O represents the state of the agent, and O represents obstacle information.

[0018] Based on the repulsive and gravitational fields of a single responding unit and a target unit, given the situation map size [l] x ,l y and spatial sampling interval l s Sampling in space generates two-dimensional image data, which is the motion situation map of a single responding unit or target unit;

[0019] Step S2: The motion situation diagrams generated by both agents are normalized, and the attraction / repulsion factor is dynamically adjusted.

[0020] The normalization formula is:

[0021]

[0022] Where: Θ t For the final motion situation diagram, This is the motion situation diagram generated when the i-th agent acts alone. and These are superimposed motion situation diagrams of the two sides' intelligent agents, where M and N are the total number of the two sides' intelligent agents, respectively.

[0023] The formula for dynamically adjusting the attraction / repulsion factor is: Where d ij Let γ be the Euclidean distance between the i-th response unit and the j-th target unit, and let γ be the attenuation coefficient.

[0024] Step S3, Multi-source environmental situation map fusion: Construct an obstacle obstruction situation map and dynamically fuse it with the motion situation map;

[0025] The obstruction situation map calculates the range of influence of obstacles using a piecewise function:

[0026]

[0027] Where r(q,q) obs ) represents the Euclidean distance between the point to be determined and the boundary of the obstacle, l0 represents the minimum influence range of the obstacle's situation, ρ1 represents the maximum influence range of the obstacle's situation, and β represents the repulsive field influence factor corresponding to the obstacle.

[0028] Each obstacle in the obstacle situation map has a different impact on the agent's behavior. Therefore, its weight in the fusion process is related to the magnitude of its impact on the agent's behavior. The fusion weight factor is:

[0029]

[0030]

[0031] Where dist(P,S) i ) represents the distance between the agent and the center of the obstacle, size(S) i ) represents the size of the obstacle, and ε is a small constant to prevent division by zero, used to avoid instability in the calculation;

[0032] The dynamic fusion process is as follows: in The motion state at position (x, y) at time i after fusion. This represents the obstacle situation at position (x, y). To integrate the motion state at position (x, y) at time i before the previous time step;

[0033] Step S4: Intent recognition based on the LSTM (Long Short-Term Memory)-Transformer hybrid model:

[0034] The LSTM-Transformer hybrid model includes:

[0035] LSTM layer: The input dimension is (batch_size, seq_len, input_size), which extracts local temporal features and captures short-term motion changes caused by obstacle avoidance or occlusion. Here, batch_size is the dataset size, seq_len is the number of sample time steps, and input_size is the real-time coordinate unfolding scale of each agent (including the responding unit and the target unit).

[0036] Transformer encoder layer: Models long-distance dependencies of trajectories through self-attention mechanism, and associates the global behavioral logic of the starting point, ending point and discontinuous time steps;

[0037] Decoder layer: fuses the spatial-temporal features of multi-source situation maps and outputs the intent classification result;

[0038] The connection relationships and input-output relationships between the LSTM layer, Transformer encoder layer, and decoder layer are as follows:

[0039] The original trajectory features of the multi-agent system and the situation map obtained by steps S1-S3 are jointly input into the LSTM layer to extract local temporal dependencies and compress trajectory data.

[0040] The output of the LSTM layer serves as the input to the Transformer encoder layer, which models global behavioral logic and captures long-distance dependencies.

[0041] The output of the Transformer encoder layer is processed by the decoder layer to obtain the final intent recognition result;

[0042] Step S5: Real-time Intent Reasoning and Output:

[0043] The dynamically fused situation map sequence is updated online and input together with the original trajectory features into the LSTM-Transformer hybrid model, which outputs the intent classification results of the multi-agent in real time, supporting robust recognition in environments with occlusion, dynamic obstacles and sensor noise.

[0044] Furthermore, during the normalization process in step S2, except... In addition to the steps, it also includes the introduction of an S-shaped nonlinear function. Compressing the data reduces the influence of extreme values, resulting in the final motion situation diagram at time t.

[0045] Furthermore, step S3, multi-source environmental situation map fusion, also includes a trajectory prediction compensation mechanism:

[0046] When trajectory information is missing due to occlusion or sensor failure, a multi-probability predicted motion situation map is generated based on historical trajectory data: Where p i Let q be the probability of the i-th predicted position. ai The predicted agent position;

[0047] The predicted motion situation map is fused with the real-time observed motion situation map sequence to fill in the missing information.

[0048] Furthermore, the repulsion factor adjustment process in step S2 also includes a dynamic adjustment mechanism for the repulsion factor:

[0049] In the initial stage of occlusion, the repulsion factor is gradually reduced by using a fade-out function to reduce the influence of the obstacle's repulsion field. The repulsion factor is then smoothly adjusted using a time-varying function F(t,q) to avoid abrupt interference.

[0050]

[0051] Where F(t,q) determines the trend of the inflection point in the motion state diagram, and satisfies F(0,q)=0, F(t) max ,q)=η max , t max η is the time corresponding to the repulsive force factor reaching its maximum value. max η is the initial value of the repulsion factor. min This represents the minimum value of the repulsion factor.

[0052] If the target continues to disappear, the repulsion factor will first drop to as low as η. hold It is maintained for a period of time, after which the repulsion factor decays to zero; if the target is observed again, the repulsion factor is gradually restored; otherwise, the repulsion factor gradually decays to 0.

[0053]

[0054] Where t2 is the time threshold, η hold To maintain a small repulsive force factor during the residual motion state diagram process, the recovery function F3(t,q) and the decay function F4(t,q) determine the trend of the residual motion state diagram, which needs to satisfy the following constraints:

[0055] (1) F3(t,q) is monotonically increasing with respect to the time variable t, and F4(t,q) is monotonically decreasing with respect to the time variable t;

[0056] (2) At the time threshold t = t2, the function values ​​are smoothly connected to avoid abrupt changes in the repulsive factor, F3(t2,q) = F4(t2,q) = η hold ;

[0057] (3)

[0058] Furthermore, in step S1, the agent attribute P includes type identifier, velocity, and acceleration, and the state S... i This includes real-time location and heading angle; the obstacle information O includes location, shape, and size.

[0059] Furthermore, the normalization process in step S2 further includes:

[0060] Based on the difference in the number of agents on both sides, the components of the motion situation diagram for each side are normalized:

[0061]

[0062] When globally overlaying the normalized motion situation map, an adversarial weight coefficient λ is introduced. p and λ n :

[0063]

[0064] Where the adversarial weight coefficient λ p and λ n The system is dynamically set based on the ratio of the number of agents on both sides to ensure that the information of the side with fewer agents is not obliterated in the fusion situation map, and to evenly display the situation of the opposing agents.

[0065] Furthermore, the generation of the obstacle obstruction situation map in step S3 specifically includes:

[0066] Irregular static obstacles are simulated using randomly generated circular obstacles, with the obstacle radius R maximizing r. max and maximum value r min It follows a uniform distribution within the range;

[0067] The obstacle repulsive field influence factor β is set according to the obstacle size as β = k·size(S) i ), where k is an empirical coefficient;

[0068] In the piecewise function, the minimum influence range l0 is set to 1.2 times the obstacle radius R, and the maximum influence range ρ1 is set to 3 times the obstacle radius R.

[0069] Furthermore, in step S4, the hidden dimension of the LSTM layer is set to 64 or 128, the number of Transformer encoder layers is 2, the number of decoder layers is 6, and the activation function is the ReLU function.

[0070] Furthermore, in the multi-probability prediction motion situation map generated based on historical trajectory data, the probability p of the predicted position... i and the predicted agent position q ai It is obtained through the following steps:

[0071] Historical trajectory data input:

[0072] Extract the agent's trajectory data from T consecutive time steps prior to the missing time period, including coordinates, velocity, and heading angle, to form the input sequence X = {x}. t-T ,x t-T+1 ,L,x t-1};

[0073] Multiple candidate trajectory generation:

[0074] The input sequence is fed into a pre-trained LSTM-Transformer hybrid prediction module. The LSTM layer extracts local temporal features and generates initial predictions for K' candidate trajectories.

[0075] The Transformer encoder layer analyzes the global consistency between the K' candidate trajectories and historical trajectories through a self-attention mechanism, and selects the K most likely trajectory prediction results. K≤K', based on the agent's position and trajectory prediction results at the last moment before the missing time period of the agent's trajectory data. A predicted position q is calculated for each candidate trajectory. ai where i = 1, 2, ..., K;

[0076] Prediction probability calculation:

[0077] The confidence level is calculated based on the fitting error between the candidate trajectory and the historical trajectory, using the following formula:

[0078]

[0079] Where σ is the smoothing coefficient, used to adjust the degree of concentration of the probability distribution;

[0080] Normalize the confidence level to obtain the probability p of each predicted location. i ;

[0081] Generation of multi-probability motion situation diagrams:

[0082] K predicted positions q ai and its probability p i Substitute into the formula Generate a predicted motion pattern with superimposed probabilities to fill in missing information.

[0083] Furthermore, the present invention also provides a multi-agent intent recognition system under conditions of incomplete multi-source environmental information, characterized in that it includes:

[0084] a. Data acquisition module (91) is used to acquire the status information of multiple agents in real time, including position, velocity, heading angle and sensor data, and output the data to the situation map generation module;

[0085] b. A situation map generation module (92), connected to the data acquisition module, comprising:

[0086] Motion situation diagram generation unit (921): Based on the artificial potential field method, it generates an attractive field and a repulsive field for each agent, and the calculation formulas are as follows: and Where η i Dynamically adjust based on agent attributes;

[0087] Obstacle situation map generation unit (922): Generates the repulsive field of obstacles through a piecewise function U rep (q) Calculate the area of ​​influence of obstacles and output the results to the situation map fusion module;

[0088] c. A situation map fusion module (93), connected to the situation map generation module, comprising:

[0089] Normalization processing unit (931): Normalizes the motion state diagrams of both agent groups using the following formula:

[0090]

[0091] Dynamic fusion unit (932): via weighting factors and By fusing the motion situation map and the obstacle situation map, a global fused situation map sequence is generated;

[0092] d. An intent recognition model (94), connected to the situation map fusion module, is an LSTM-Transformer hybrid network architecture, comprising:

[0093] LSTM layer (941): The input dimension is batch_size×seq_len×input_dim, which extracts local temporal features;

[0094] Transformer encoder layer (942) and Transformer decoder layer (943): Model global dependencies through self-attention mechanism and fuse multi-source situation map information;

[0095] Output layer (944): Generates intent classification results and transmits them to the real-time processing module;

[0096] e. Real-time processing module (95), connected to the intent recognition model, is used to update the situation map sequence online and output the intent recognition result in real time in the dynamic environment;

[0097] f. The compensation and prediction module (96) interacts with the data acquisition module and the situation map fusion module respectively, including:

[0098] Trajectory prediction unit (961): When trajectory information is missing, a multi-probability predicted motion situation map is generated based on historical data;

[0099] Gravity / repulsion factor adjustment unit (962): The gravity / repulsion field of the obstacle is dynamically adjusted by time-varying functions F3(t,q) and F4(t,q) to avoid sudden changes in the situation map due to data loss;

[0100] g. Storage module (97), connected to all functional modules, used to cache historical situation map sequences, model parameters and real-time data.

[0101] Connection relationships between components:

[0102] The output of the data acquisition module (91) is connected to the input of the situation map generation module (92);

[0103] The output of the situation map generation module (92) is connected to the input of the situation map fusion module (93);

[0104] The output of the situation map fusion module (93) is connected to the input of the intent recognition model (94);

[0105] The output of the intent recognition model (94) is connected to the input of the real-time processing module (95);

[0106] The input end of the compensation and prediction module (96) receives the abnormal signal from the data acquisition module, and the output end is connected to the weight adjustment interface of the situation map fusion module (93).

[0107] The storage module (97) communicates bidirectionally with each module through the data bus, supporting data reading, writing and synchronous updates;

[0108] The real-time processing module (95) outputs the intent recognition result in real time.

[0109] Compared with the prior art, the beneficial effects of the present invention are:

[0110] The method of this invention integrates the motion situation map and the obstacle situation map (weighting factor). and This method encodes agent interactions and the influence of environmental obstacles into spatial features, which, compared to the original trajectory input, more intuitively reflects the tactical game logic in a dynamic environment, thereby improving recognition accuracy. The spatial structured characteristics of the situation map (such as the repulsive field amplitude being limited to (0, η)) further enhance its effectiveness. i Within a certain range, this method avoids noise interference in trajectory data, enhancing model robustness. This invention captures local motion abrupt changes caused by obstacle avoidance using an LSTM layer and combines it with a Transformer encoder to model the correlation between obstacle distribution and global intent, achieving optimal overall performance in complex scenarios. Experiments show that the accuracy of this invention's method in obstacle-containing scenarios is improved compared to existing methods, verifying the technical advantage of the synergistic effect of the situation map and hybrid model, and providing a reliable solution for complex adversarial scenarios.

[0111] This invention adjusts the weight of the obstacle repulsion field in real time through a dynamic repulsion factor adjustment mechanism, avoiding sudden changes in the situation map. It solves the problem of gradient information loss or singular values ​​caused by fixed weights in the traditional artificial potential field method in dynamic environments, and improves the adaptability to dynamic targets in complex or dynamic scenarios.

[0112] Using normalization techniques, the difference in the number of agents on both sides of the adversarial conflict is addressed through the formula: Normalization is performed to ensure that the situational information of minority groups is not overwhelmed, thereby improving the balance and reliability of the fusion results.

[0113] By combining the LSTM-Transformer hybrid model with a trajectory prediction compensation mechanism, the LSTM layer captures the local temporal features of short-term obstacle avoidance actions, solving the problem of missing trajectory segments caused by sensor noise or occlusion. The Transformer encoder models long-distance dependencies through a self-attention mechanism, associating information from discontinuous time steps. The prediction compensation unit generates a multi-probability situation map, and by fusing historical and predicted data with confidence weights, it completes the missing trajectory, enhancing the robustness of multi-agent intent recognition in scenarios with incomplete information.

[0114] This invention is applicable to multi-agent systems, including drones, unmanned vehicles, robots, etc. Attached Figure Description

[0115] Figure 1 This is a flowchart of the multi-agent intent recognition method under incomplete multi-source environmental information in Embodiment 1 of the present invention;

[0116] Figure 2 This is a schematic diagram of the gravitational and repulsive fields of the motion state of a single intelligent agent based on the artificial potential field method in Embodiment 1 of the present invention;

[0117] Figure 3 (a) is a motion situation diagram of multi-agent fusion in Embodiment 1 of the present invention;

[0118] Figure 3 (b) is a diagram of the obstruction situation in Embodiment 1 of the present invention;

[0119] Figure 4 This is a flowchart illustrating the dynamic fusion of obstacle and motion states in Embodiment 1 of the present invention.

[0120] Figure 5 This is a multi-source environmental fusion motion situation diagram in Embodiment 1 of the present invention;

[0121] Figure 6 This is a schematic diagram of the multi-agent intent recognition system in Embodiment 2 of the present invention;

[0122] Figure 7 This is a schematic diagram of four typical intentions of multi-agents (encircle, intercept, track, and irrelevant) in Embodiment 3 of the present invention;

[0123] Figure 8 (a) is a comparison chart of training set accuracy in the scenario with incomplete information (obstacles and trajectory occlusion) in Embodiment 3 of the present invention;

[0124] Figure 8 (b) is a comparison chart of test set accuracy in the scenario of incomplete information (obstacles and trajectory occlusion) in Embodiment 3 of the present invention;

[0125] Figure 9 (a) is a comparison chart of training loss in the scenario of incomplete information (obstacles and trajectory occlusion) in Embodiment 3 of the present invention;

[0126] Figure 9 (b) is a comparison chart of test losses in the scenario of incomplete information (obstacles and trajectory occlusion) in Embodiment 3 of the present invention. Detailed Implementation

[0127] Example 1

[0128] like Figure 1 As shown, this invention provides a method for multi-agent intent recognition under conditions of incomplete multi-source environmental information, characterized by comprising the following steps:

[0129] Step S1: Generation of Single-Agent Motion Situation Diagram: Based on the artificial potential field method, gravitational and repulsive fields are generated for the target unit and the responding unit respectively, as shown in the diagram. Figure 2 As shown;

[0130] The formula for calculating the unit repulsive field is as follows: Where η i Let q be the scale factor of the repulsive / gravitational field of the i-th agent.a Let ρ(q,q) be the position of the agent. a Let η be the distance between the point to be determined and the agent, μ be the scale parameter controlling the attenuation range of the potential field, and the magnitude of the repulsive field be limited to (0, η). i ];

[0131] The formula for calculating the target unit gravitational field is as follows:

[0132] The scaling factor η i Through the mapping function α(P,S) i O) is dynamically determined, where P is the agent attribute, and S is the agent attribute. i O represents the state of the agent, and O represents obstacle information.

[0133] Based on the repulsive and gravitational fields of a single responding unit and a target unit, given the situation map size [l] x ,l y and spatial sampling interval l s Sampling in space generates two-dimensional image data, which is the motion situation map of a single responding unit or target unit;

[0134] Step S2: The motion situation diagrams generated by both agents are normalized, and the attraction / repulsion factor is dynamically adjusted.

[0135] The normalization formula is:

[0136]

[0137] Where: Θ t For the final motion situation diagram, This is the motion situation diagram generated when the i-th agent acts alone. and These are superimposed motion situation diagrams of the two sides' intelligent agents, where M and N are the total number of the two sides' intelligent agents, respectively.

[0138] The formula for dynamically adjusting the attraction / repulsion factor is: Where d ij Let γ be the Euclidean distance between the i-th response unit and the j-th target unit, and let γ be the attenuation coefficient.

[0139] The resulting multi-agent fusion motion situation diagram is shown in the figure below. Figure 3 As shown in (a);

[0140] Step S3, Multi-source environmental situation map fusion: Construct an obstacle obstruction situation map and dynamically fuse it with the motion situation map;

[0141] The obstruction situation map calculates the range of influence of obstacles using a piecewise function:

[0142]

[0143] Where r(q,q) obs Let ρ be the Euclidean distance between the point to be determined and the obstacle boundary, l0 be the minimum influence range of the obstacle situation, ρ1 be the maximum influence range of the obstacle situation, and β be the repulsive force field influence factor corresponding to the obstacle. The resulting obstacle situation diagram is shown in the figure below. Figure 3 As shown in (b);

[0144] Each obstacle in the obstacle situation map has a different impact on the agent's behavior. Therefore, its weight in the fusion process is related to the magnitude of its impact on the agent's behavior. The fusion weight factor is:

[0145]

[0146] Where dist(P,S) i ) represents the distance between the agent and the center of the obstacle, size(S) i ) represents the size of the obstacle, and ε is a small constant to prevent division by zero, used to avoid instability in the calculation.

[0147] The dynamic fusion process is as follows: in The motion state at position (x, y) at time i after fusion. This represents the obstacle situation at position (x, y). To fuse the motion state at position (x, y) at time i, the flowchart for dynamically fusing the hindering state and the motion state is as follows: Figure 4 As shown in the diagram, the resulting multi-source environmental fusion motion situation map is as follows: Figure 5 As shown;

[0148] Step S4: Intent recognition based on the LSTM (Long Short-Term Memory)-Transformer hybrid model:

[0149] The LSTM-Transformer hybrid model includes:

[0150] LSTM layer: The input dimension is (batch_size, seq_len, input_size), which extracts local temporal features and captures short-term motion changes caused by obstacle avoidance or occlusion. Here, batch_size is the dataset size, seq_len is the number of sample time steps, and input_size is the real-time coordinate unfolding scale of each agent (including the responding unit and the target unit).

[0151] Transformer encoder layer: Models long-distance dependencies of trajectories through self-attention mechanism, and associates the global behavioral logic of the starting point, ending point and discontinuous time steps;

[0152] Decoder layer: fuses the spatial-temporal features of multi-source situation maps and outputs the intent classification result;

[0153] The connection relationships and input-output relationships between the LSTM layer, Transformer encoder layer, and decoder layer are as follows:

[0154] The original trajectory features of the multi-agent system and the situation map obtained by steps S1-S3 are jointly input into the LSTM layer to extract local temporal dependencies and compress trajectory data.

[0155] The output of the LSTM layer serves as the input to the Transformer encoder layer, which models global behavioral logic and captures long-distance dependencies.

[0156] The output of the Transformer encoder layer is processed by the decoder layer to obtain the final intent recognition result;

[0157] Step S5: Real-time Intent Reasoning and Output:

[0158] The dynamically fused situation map sequence is updated online and input together with the original trajectory features into the LSTM-Transformer hybrid model, which outputs the intent classification results of the multi-agent in real time, supporting robust recognition in environments with occlusion, dynamic obstacles and sensor noise.

[0159] Furthermore, during the normalization process in step S2, except... In addition to the steps, it also includes the introduction of an S-shaped nonlinear function. Compressing the data reduces the influence of extreme values, resulting in the final motion situation diagram at time t.

[0160] Furthermore, step S3, multi-source environmental situation map fusion, also includes a trajectory prediction compensation mechanism:

[0161] When trajectory information is missing due to occlusion or sensor failure, a multi-probability predicted motion situation map is generated based on historical trajectory data: Where p i Let q be the probability of the i-th predicted position. ai The predicted agent position;

[0162] The predicted motion situation map is fused with the real-time observed motion situation map sequence to fill in the missing information.

[0163] Furthermore, the repulsion factor adjustment process in step S2 also includes a dynamic adjustment mechanism for the repulsion factor:

[0164] In the initial stage of occlusion, the repulsion factor is gradually reduced by using a fade-out function to reduce the influence of the obstacle's repulsion field. The repulsion factor is then smoothly adjusted using a time-varying function F(t,q) to avoid abrupt interference.

[0165]

[0166] Where F(t,q) determines the trend of the inflection point in the motion state diagram, and satisfies F(0,q)=0, F(t) max ,q)=η max , t max η is the time corresponding to the repulsive force factor reaching its maximum value. max η is the initial value of the repulsion factor. min It is the minimum value of the repulsion factor;

[0167] If the target continues to disappear, the repulsion factor will first drop to as low as η. hold It is maintained for a period of time, after which the repulsion factor decays to zero; if the target is observed again, the repulsion factor is gradually restored; otherwise, the repulsion factor gradually decays to 0.

[0168]

[0169] Where t2 is the time threshold, η hold To maintain a small repulsive force factor during the residual motion state diagram process, the recovery function F3(t,q) and the decay function F4(t,q) determine the trend of the residual motion state diagram, which needs to satisfy the following constraints:

[0170] (1) F3(t,q) is monotonically increasing with respect to the time variable t, and F4(t,q) is monotonically decreasing with respect to the time variable t;

[0171] (2) At the time threshold t = t2, the function values ​​are smoothly connected to avoid abrupt changes in the repulsive factor, F3(t2,q) = F4(t2,q) = η hold ;

[0172] (3)

[0173] Furthermore, in step S1, the agent attribute P includes type identifier, velocity, and acceleration, and the state S... i This includes real-time location and heading angle; the obstacle information O includes location, shape, and size.

[0174] Furthermore, the normalization process in step S2 further includes:

[0175] Based on the difference in the number of agents on both sides, the components of the motion situation diagram for each side are normalized:

[0176]

[0177] When globally overlaying the normalized motion situation map, an adversarial weight coefficient λ is introduced. p and λ n :

[0178]

[0179] Where the adversarial weight coefficient λ p and λ n The system is dynamically set based on the ratio of the number of agents on both sides to ensure that the information of the side with fewer agents is not obliterated in the fusion situation map, and to evenly display the situation of the opposing agents.

[0180] Furthermore, the generation of the obstacle obstruction situation map in step S3 specifically includes:

[0181] Irregular static obstacles are simulated using randomly generated circular obstacles, with the obstacle radius R maximizing r. max and maximum value r min It follows a uniform distribution within the range;

[0182] The obstacle repulsive field influence factor β is set according to the obstacle size as β = k·size(S) i ), where k is an empirical coefficient;

[0183] In the piecewise function, the minimum influence range l0 is set to 1.2 times the obstacle radius R, and the maximum influence range ρ1 is set to 3 times the obstacle radius R.

[0184] Furthermore, in step S4, the hidden dimension of the LSTM layer is set to 64 or 128, the number of Transformer encoder layers is 2, the number of decoder layers is 6, and the activation function is the ReLU function.

[0185] Furthermore, in the multi-probability prediction motion situation map generated based on historical trajectory data, the probability p of the predicted position... i and the predicted agent position q ai It is obtained through the following steps:

[0186] Historical trajectory data input:

[0187] Extract the agent's trajectory data from T consecutive time steps prior to the missing time period, including coordinates, velocity, and heading angle, to form the input sequence X = {x}. t-T ,x t-T+1 ,L,x t-1};

[0188] Multiple candidate trajectory generation:

[0189] The input sequence is fed into a pre-trained LSTM-Transformer hybrid prediction module. The LSTM layer extracts local temporal features and generates initial predictions for K' candidate trajectories.

[0190] The Transformer encoder layer analyzes the global consistency between the K' candidate trajectories and historical trajectories through a self-attention mechanism, and selects the K most likely trajectory prediction results. K≤K', based on the agent's position and trajectory prediction results at the last moment before the missing time period of the agent's trajectory data. A predicted position q is calculated for each candidate trajectory. ai where i = 1, 2, ..., K;

[0191] Prediction probability calculation:

[0192] The confidence level is calculated based on the fitting error between the candidate trajectory and the historical trajectory, using the following formula:

[0193]

[0194] Where σ is the smoothing coefficient, used to adjust the degree of concentration of the probability distribution;

[0195] Normalize the confidence level to obtain the probability p of each predicted location. i ;

[0196] Generation of multi-probability motion situation diagrams:

[0197] K predicted positions q ai and its probability p i Substitute into the formula Generate a predicted motion pattern with superimposed probabilities to fill in missing information.

[0198] Example 2

[0199] This invention also provides a multi-agent intent recognition system under conditions of incomplete multi-source environmental information, such as... Figure 6 As shown, the feature is that it includes:

[0200] a. Data acquisition module (91) is used to acquire the status information of multiple agents in real time, including position, velocity, heading angle and sensor data, and output the data to the situation map generation module;

[0201] b. A situation map generation module (92), connected to the data acquisition module, comprising:

[0202] Motion situation diagram generation unit (921): Based on the artificial potential field method, it generates an attractive field and a repulsive field for each agent, and the calculation formulas are as follows: and Where ηi Dynamically adjust based on agent attributes;

[0203] Obstacle situation map generation unit (922): Generates the repulsive field of obstacles through a piecewise function U rep (q) Calculate the area of ​​influence of obstacles and output the results to the situation map fusion module;

[0204] c. A situation map fusion module (93), connected to the situation map generation module, comprising:

[0205] Normalization processing unit (931): Normalizes the motion state diagrams of both agent groups using the following formula:

[0206]

[0207] Dynamic fusion unit (932): via weighting factors and By fusing the motion situation map and the obstacle situation map, a global fused situation map sequence is generated;

[0208] d. An intent recognition model (94), connected to the situation map fusion module, is an LSTM-Transformer hybrid network architecture, comprising:

[0209] LSTM layer (941): The input dimension is batch_size×seq_len×input_dim, which extracts local temporal features;

[0210] Transformer encoder layer (942) and Transformer decoder layer (943): Model global dependencies through self-attention mechanism and fuse multi-source situation map information;

[0211] Output layer (944): Generates intent classification results and transmits them to the real-time processing module;

[0212] e. Real-time processing module (95), connected to the intent recognition model, is used to update the situation map sequence online and output the intent recognition result in real time in the dynamic environment;

[0213] f. The compensation and prediction module (96) interacts with the data acquisition module and the situation map fusion module respectively, including:

[0214] Trajectory prediction unit (961): When trajectory information is missing, a multi-probability predicted motion situation map is generated based on historical data;

[0215] Gravity / repulsion factor adjustment unit (962): The gravity / repulsion field of the obstacle is dynamically adjusted by time-varying functions F3(t,q) and F4(t,q) to avoid sudden changes in the situation map due to data loss;

[0216] g. Storage module (97), connected to all functional modules, used to cache historical situation map sequences, model parameters and real-time data.

[0217] Connection relationships between components:

[0218] The output of the data acquisition module (91) is connected to the input of the situation map generation module (92);

[0219] The output of the situation map generation module (92) is connected to the input of the situation map fusion module (93);

[0220] The output of the situation map fusion module (93) is connected to the input of the intent recognition model (94);

[0221] The output of the intent recognition model (94) is connected to the input of the real-time processing module (95);

[0222] The input end of the compensation and prediction module (96) receives the abnormal signal from the data acquisition module, and the output end is connected to the weight adjustment interface of the situation map fusion module (93).

[0223] The storage module (97) communicates bidirectionally with each module through the data bus, supporting data reading, writing and synchronous updates; the output of the real-time processing module (95) outputs the intent recognition result in real time.

[0224] Example 3

[0225] This embodiment verifies the effectiveness of the method of the present invention in information-incomplete environments with obstacles and trajectory occlusion compared to existing technologies such as Transformer models, CNNs (Convolutional Neural Networks), MLPs (Multilayer Perceptrons), and LSTMs (Long Short-Term Memory). This embodiment generates a multi-agent trajectory dataset with 1000 samples, featuring obstacles and trajectory occlusion, and includes four typical intentions (encircle, intercept, track, and irrelevant). The definitions of the four intentions are as follows:

[0226] (1) Encirclement: The responding unit forms a blockade around the target unit, ensuring that there is a responding unit in any direction, thereby restricting the target unit's mobility and making it difficult for it to break through easily;

[0227] (2) Interception: In order to prevent the target unit from breaking through the blockade line, the responding units should form a semi-circular interception formation in the direction of advance to maximize the blocking of the target unit's escape route;

[0228] (3) Tracking: The response units are positioned behind the target unit, forming a semi-circular tracking formation to ensure that as many response units as possible maintain effective tracking when the target suddenly accelerates or changes direction.

[0229] (4) Irrelevant: Each unit moves independently based on environmental characteristics and does not make any adaptive adjustments for a specific target.

[0230] Figure 7 Schematic diagrams of four typical intentions of multi-agent interactions (encircle, intercept, track, and irrelevant) are presented. Red represents responding units, blue represents target units, green circles represent obstacles hindering movement, and yellow circles represent obstacles blocking detection. Figure 7 (a) is a schematic diagram of the multi-agent encirclement intent. Figure 7 (b) is a schematic diagram of a multi-agent interception intent. Figure 7 (c) is a schematic diagram of multi-agent tracking intent. Figure 7 (d) is a schematic diagram of multi-agent-independent intent.

[0231] Set the hidden dimension of LSTM and Transformer to 128, and compare the models (Transformer, CNN, MLP and LSTM) with the parameters set according to the standard configuration.

[0232] The original trajectory data in scenarios with incomplete information (obstacles and trajectory occlusion) is processed according to the method of this invention to obtain a situation map after multi-source environmental situation fusion. This map is used as the input to the LSTM-Transformer hybrid model of this invention and the comparison model. The comparison chart of the training set accuracy of the method of this invention and the existing technology method in scenarios with incomplete information (obstacles and trajectory occlusion) is shown below. Figure 8 As shown in (a), the test set accuracy comparison chart is as follows: Figure 8 As shown in (b), the training loss comparison diagram is as follows: Figure 9 As shown in (a), the test loss comparison chart is as follows: Figure 9 As shown in (b), the final accuracy of the method of the present invention on the test set is 93.39%, which is higher than that of the Transformer model (92.76%), CNN (91.50%), MLP (91.57%) and LSTM (91.13%).

[0233] From such Figure 8 (a) Figure 8 (b) Figure 9 (a) Figure 9(b) shows that the final recognition accuracy of the method of the present invention on the training set and the test set is significantly higher than that of the existing technology method. The training loss and test loss of the method of the present invention are significantly lower than those of the existing technology method. Moreover, compared with the existing technology method, the method of the present invention has a faster convergence speed, which verifies the superiority of the method of the present invention in performing intent recognition under incomplete information conditions and proves that the method of the present invention has a strong adaptability to dynamic and complex scenarios.

[0234] The method of this invention integrates the motion situation map and the obstacle situation map (weighting factor). and This method encodes agent interactions and the influence of environmental obstacles into spatial features, which, compared to the original trajectory input, more intuitively reflects the tactical game logic in a dynamic environment, thereby improving recognition accuracy. The spatial structured characteristics of the situation map (such as the repulsive field amplitude being limited to (0, η)) further enhance its effectiveness. i Within a certain range, this method avoids noise interference in trajectory data, enhancing model robustness. This invention captures local motion abrupt changes caused by obstacle avoidance using an LSTM layer and combines it with a Transformer encoder to model the correlation between obstacle distribution and global intent, achieving optimal overall performance in complex scenarios. Experiments show that the accuracy of this invention's method in obstacle-containing scenarios is improved compared to existing methods, verifying the technical advantage of the synergistic effect of the situation map and hybrid model, and providing a reliable solution for complex adversarial scenarios.

Claims

1. A method for multi-agent intent recognition under conditions of incomplete multi-source environmental information, characterized in that, Includes the following steps: Step S1: Generation of single agent motion situation diagram: Based on the artificial potential field method, generate gravitational field and repulsive field for the target unit and the responding unit respectively; The formula for calculating the unit repulsive field is as follows: ,in Let be the scale factor of the repulsive / gravitational field of the i-th agent. For the location of the agent, Let be the distance between the point to be found and the agent. To control the scale parameter of the potential field decay range, the repulsive field amplitude is limited to ; The formula for calculating the target unit gravitational field is: ; The repulsive / gravitational field scale factor Through mapping function Dynamically determined, where P represents agent attributes, S... i O represents the state of the agent, and O represents obstacle information. Based on the repulsive and gravitational fields of a single responding unit and a target unit, given the situation map size... and spatial sampling interval Sampling in space generates two-dimensional image data, which is the motion situation map of a single responding unit or target unit; Step S2: The motion situation diagrams generated by both agents are normalized, and the scale factor of the repulsive field / gravitational field is dynamically adjusted. The normalization formula is: ， ， ； in: For the final motion situation diagram, This is the motion situation diagram generated when the i-th agent acts alone. and These are superimposed motion situation diagrams of the two sides' intelligent agents, where M and N are the total number of the two sides' intelligent agents, respectively. The formula for dynamically adjusting the scale factor of the repulsive / gravitational field is as follows: , where d ij Let γ be the Euclidean distance between the i-th response unit and the j-th target unit, and let γ be the attenuation coefficient. Step S3, Multi-source environmental situation map fusion: Construct an obstacle obstruction situation map and dynamically fuse it with the motion situation map; The obstruction situation map calculates the range of influence of obstacles using a piecewise function: ， in Let be the Euclidean distance between the point to be determined and the boundary of the obstacle. To minimize the impact of the obstacle's situation, The maximum impact range of the obstacle situation. The influencing factor of the repulsive force field of the obstacle; Each obstacle in the obstacle situation map has a different impact on the agent's behavior. Therefore, its weight in the fusion process is related to the magnitude of its impact on the agent's behavior. The fusion weight factor is: in The distance between the agent and the center of the obstacle. The size of the obstacle, A small constant used to prevent division by zero, thus avoiding instability in calculations; The dynamic fusion process is as follows: ,in For the fusion The time position is The movement status at the location, For position is The obstructive situation at the scene, Before integration The time position is The movement status at the location; Step S4: Intent recognition based on the LSTM-Transformer hybrid model: The LSTM-Transformer hybrid model includes: LSTM layer: The input dimension is batch_size×seq_len×input_size. It extracts local temporal features and captures short-term motion changes caused by obstacle avoidance or occlusion. Here, batch_size is the dataset size, seq_len is the number of sample time steps, and input_size is the scale of real-time coordinate expansion for each agent. Transformer encoder layer: Models long-distance dependencies of trajectories through self-attention mechanism, and associates the global behavioral logic of the starting point, ending point and discontinuous time steps; Decoder layer: fuses the spatial-temporal features of multi-source situation maps and outputs the intent classification result; The connection relationships and input-output relationships between the LSTM layer, Transformer encoder layer, and decoder layer are as follows: The original trajectory features of the multi-agent system and the fused situation map obtained by steps S1-S3 are jointly input into the LSTM layer to extract local temporal dependencies and compress trajectory data. The output of the LSTM layer serves as the input to the Transformer encoder layer, which models global behavioral logic and captures long-distance dependencies. The output of the Transformer encoder layer is processed by the decoder layer to obtain the final intent recognition result; Step S5: Real-time Intent Reasoning and Output: The dynamically fused situation map sequence is updated online and input together with the original trajectory features into the LSTM-Transformer hybrid model, which outputs the intent classification results of the multi-agent in real time, supporting robust recognition in environments with occlusion, dynamic obstacles and sensor noise.

2. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 1, characterized in that, During the normalization process in step S2, except... , , In addition to the steps, it also includes the introduction of an S-shaped nonlinear function. Compressing the data reduces the influence of extreme values, resulting in the final motion situation diagram at time t. .

3. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 2, characterized in that, The multi-source environmental situation map fusion step S3 also includes a trajectory prediction and compensation mechanism: When trajectory information is missing due to occlusion or sensor failure, a multi-probability predicted motion situation map is generated based on historical trajectory data: , where p i Let q be the probability of the i-th predicted position. ai The predicted location of the agent; The predicted motion situation map is fused with the real-time observed motion situation map sequence to fill in the missing information.

4. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 3, characterized in that, The repulsive / gravitational field scale factor adjustment process in step S2 also includes a dynamic adjustment mechanism for the repulsive / gravitational field scale factor: In the initial stage of occlusion, the scale factor of the repulsive / gravitational field is gradually reduced by using a fade-out function, thereby reducing the influence of the obstacle's repulsive field. This is achieved using a time-varying function. Smoothly adjust the scale factor of the repulsive / gravitational field to avoid abrupt changes. ， in The trend of the gradual ingress of the motion pattern is determined, and the following conditions are met: , , This represents the time corresponding to the repulsive / gravitational field scale factor reaching its maximum value. This represents the maximum value of the scale factor for the repulsive / gravitational field. This represents the minimum scale factor of the repulsive / gravitational field. If the target continues to disappear, the repulsive / gravitational field scale factor first decreases to as low as... It is maintained for a period of time. After time t2, the repulsive / gravitational field scale factor decays to zero. If the target is observed again, the repulsive / gravitational field scale factor is gradually restored; otherwise, the repulsive / gravitational field scale factor gradually decays to 0. Where t2 is the time threshold. The recovery function is the small repulsive / gravitational field scale factor maintained when the motion state diagram remains. With decay function The following constraints must be met to determine the residual trend of the motion pattern: (1) The time variable t is monotonically increasing. The time variable t is monotonically decreasing; (2) At the time threshold At this point, the function values ​​are smoothly connected to avoid abrupt changes in the scale factor of the repulsive / gravitational field. ; （3） ， 。 5. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 4, characterized in that, In step S1, the agent attribute P includes type identifier, velocity, acceleration, and state S. i This includes real-time location and heading angle, and obstacle information O including location, shape, and size.

6. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 5, characterized in that, The normalization process in step S2 further includes: Based on the difference in the number of agents on both sides, the components of the motion situation diagram for each side are normalized: ， ； When globally overlaying the normalized motion situation map, an adversarial weight coefficient is introduced. and : Among them, the adversarial weight coefficient and The system is dynamically set based on the ratio of the number of agents on both sides to ensure that the information of the side with fewer agents is not obliterated in the fusion situation map, and to evenly display the situation of the opposing agents.

7. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 6, characterized in that, The generation of the obstacle obstruction situation map in step S3 specifically includes: Irregular static obstacles are simulated using randomly generated circular obstacles, with the obstacle radius R maximizing r. max and minimum value r min It follows a uniform distribution within the range; The obstacle repulsive field influence factor β is set according to the obstacle size. , where k is an empirical coefficient; In the piecewise function, the minimum influence range Set to 1.2 times the obstacle radius R, maximum influence range Set to 3 times the obstacle radius R.

8. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 7, characterized in that, In step S4, the hidden dimension of the LSTM layer is set to 64 or 128, the number of Transformer encoder layers is 2, the number of Transformer decoder layers is 6, and the Transformer activation function is the ReLU function.

9. The method for multi-agent intent recognition under conditions of incomplete multi-source environmental information as described in claim 8, characterized in that, In the multi-probability predictive motion map generated based on historical trajectory data, the probability p of the predicted position... i and the predicted agent position q ai It is obtained through the following steps: Historical trajectory data input: Extract the agent's trajectory data from T consecutive time steps prior to the missing time period, including coordinates, velocity, and heading angle, to form the input sequence. ; Multiple candidate trajectory generation: The input sequence is fed into a pre-trained LSTM-Transformer hybrid prediction module. The LSTM layer extracts local temporal features and generates initial predictions for K' candidate trajectories. The Transformer encoder layer analyzes the global consistency between the K' candidate trajectories and historical trajectories through a self-attention mechanism, and selects the K most likely trajectory prediction results. If K≤K', the prediction result is based on the agent's position and trajectory at the last moment before the missing time period of the agent's trajectory data. A predicted position q is calculated for each candidate trajectory. ai where i = 1, 2, ..., K; Prediction probability calculation: The confidence level is calculated based on the fitting error between the candidate trajectory and the historical trajectory, using the following formula: Where σ is the smoothing coefficient, used to adjust the concentration of the probability distribution; the confidence level is normalized to obtain the probability p of each predicted location. i ; Generation of multi-probability motion situation diagrams: K predicted positions q ai and its probability p i Substitute into the formula It generates a predicted motion pattern with superimposed probabilities to fill in missing information.

10. A multi-agent intent recognition system under conditions of incomplete multi-source environmental information, characterized in that, include: a. Data acquisition module (91) is used to acquire the status information of multiple agents in real time, including position, velocity, heading angle and sensor data, and output the data to the situation map generation module; b. A situation map generation module (92), connected to the data acquisition module, including: Motion situation diagram generation unit (921): Based on the artificial potential field method, it generates gravitational and repulsive fields for each agent, and the calculation formulas are as follows: and ,in Dynamically adjust based on agent attributes; in Let be the scale factor of the repulsive / gravitational field of the i-th agent. For the location of the agent, Let be the distance between the point to be found and the agent. To control the scale parameter of the potential field decay range, the repulsive field amplitude is limited to ; Obstacle situation map generation unit (922): Generates the repulsive field of obstacles through a piecewise function U rep (q) Calculate the area of ​​influence of obstacles and output the results to the situation map fusion module; c. A situation map fusion module (93), connected to the situation map generation module, including: Normalization processing unit (931): Normalizes the motion situation diagrams of both agent groups respectively, using the formula: , , ; in: For the final motion situation diagram, This is the motion situation diagram generated when the i-th agent acts alone. and These are superimposed motion situation diagrams of the two sides' intelligent agents, where M and N represent the total number of intelligent agents on both sides, respectively. and To counteract the weighting coefficient; Dynamic fusion unit (932): through weighting factors and By fusing the motion situation map and the obstacle situation map, a global fused situation map sequence is generated; d. An intent recognition model (94), connected to the situation map fusion module, is an LSTM-Transformer hybrid network architecture, comprising: LSTM layer (941): The input dimension is batch_size×seq_len×input_size, which extracts local temporal features; Where batch_size is the dataset size, seq_len is the number of time steps for the sample, and input_size is the scale of the real-time coordinate unfolding for each agent; Transformer encoder layer (942) and Transformer decoder layer (943): Model global dependencies through self-attention mechanism and fuse multi-source situational information; Output layer (944): Generates intent classification results and transmits them to the real-time processing module; e. Real-time processing module (95), connected to the intent recognition model, is used to update the situation map sequence online and output the intent recognition result in real time in the dynamic environment; f. The compensation and prediction module (96) interacts with the data acquisition module and the situation map fusion module, respectively, including: Trajectory prediction unit (961): When trajectory information is missing, it generates a multi-probability predicted motion situation map based on historical data; Repulsive / gravitational field scale factor adjustment unit (962): via recovery function With decay function Dynamically adjust the gravitational / repulsive fields of obstacles to avoid sudden changes in the situation map due to data loss; g. Storage module (97), connected to all functional modules, used to cache historical situation map sequences, model parameters and real-time data; Connection relationships between components: The output of the data acquisition module (91) is connected to the input of the situation map generation module (92); The output of the situation map generation module (92) is connected to the input of the situation map fusion module (93); The output of the situation map fusion module (93) is connected to the input of the intent recognition model (94); The output of the intent recognition model (94) is connected to the input of the real-time processing module (95); The input end of the compensation and prediction module (96) receives the abnormal signal from the data acquisition module, and the output end is connected to the weight adjustment interface of the situation map fusion module (93). The storage module (97) communicates bidirectionally with each module through the data bus, supporting data reading, writing and synchronous updates; The real-time processing module (95) outputs the intent recognition result in real time.

Images