A Multimodal Dynamic Cooperative Unmanned Surface Vessel Swarm Autonomous Obstacle Avoidance Method

Abstract

The application provides a kind of multimodal dynamic cooperative unmanned ship cluster autonomous obstacle avoidance method, first, through IEEE 1588 protocol synchronization LiDAR point cloud, camera image, radar data, AIS data, and cross-modal space alignment is realized based on external parameter matrix;Second, the channel and spatial dimension correlation is established by combining cross-modal attention mechanism and deformable convolution, so that the fused multimodal feature generates high-precision environment semantic information;Then, based on the time-varying graph incremental module, the unmanned ship task group is dynamically divided, and only the disturbed boundary set is updated to reduce the computational complexity;Finally, the improved velocity obstacle model (IVO-DWA) is used for the leader of unmanned ship to optimize multi-objective path, and the follower corrects the trajectory through repulsive potential field, to realize real-time, safe and compliant cluster obstacle avoidance.

Description

Technical Field

[0001] This invention relates to the field of shipborne system information and communication technology, specifically to a multimodal dynamic cooperative unmanned surface vessel swarm autonomous obstacle avoidance method. Background Technology

[0002] Unmanned surface vessel (USV) swarms, as a type of intelligent waterborne equipment capable of autonomous navigation, have attracted much attention due to their maneuverability, cost-effectiveness, and ability to enhance human initiative in unknown waters and perform dangerous waterborne tasks in place of humans. This has led to their use in both public and civilian sectors. In the civilian sector, USV swarms can be used for hydrological exploration, environmental maintenance, and disaster search and rescue, reducing labor costs and improving operational efficiency. In the public sector, USV swarms can be used for anti-submarine and anti-missile operations, water blockade, and area patrols, reducing casualties and improving cost-effectiveness.

[0003] For example, Chinese patent application number 202410732541.7, classification number G05D1 / 43, and publication date July 9, 2024, discloses a method for unmanned surface vessel (USV) formation tracking and obstacle avoidance control, belonging to the field of USV formation motion control technology. This method uses sensor measurements and shore-based transmissions of obstacle, USV formation, and channel status information, along with the desired target point. Based on an improved artificial potential field method, a collision avoidance guidance law is used to plan feasible paths that meet collision avoidance requirements and comply with international maritime collision avoidance rules. Furthermore, a formation change strategy based on a hazard assessment function is used to plan the USV formation scheme. Finally, a control law is generated in real-time using an RBF neural network estimator and a backstepping controller, and executed by the actuator to drive the USV formation to track the desired target point. When there is a collision risk, formation collision avoidance is prioritized; when there is no collision risk, formation tracking and formation maintenance are achieved.

[0004] The aforementioned literature mainly employs an improved artificial potential field method combined with a formation transformation strategy. By acquiring information on the motion state and environmental obstacles of the unmanned surface vessel, a manipulation motion model is established, attractive and repulsive force functions are calculated, feasible paths are planned, and control commands are generated using an RBF neural network and backstepping method to achieve adaptive tracking and obstacle avoidance of the formation. However, because it directly processes data from cameras, radars, etc., under different spatiotemporal conditions, the cross-modal feature alignment error caused by the spatiotemporal reference differences of heterogeneous sensors of different modalities is large, which easily leads to low obstacle detection accuracy and is not suitable for applications requiring high precision. In addition, it directly changes the triangular relationship between the leader and followers when encountering obstacles, but it cannot dynamically adjust the grouping of the community and then select the optimal path based on different grouping conditions to achieve better path control. Summary of the Invention

[0005] The purpose of this invention is to provide a multimodal dynamic cooperative unmanned surface vessel (USV) swarm autonomous obstacle avoidance method, which has high obstacle detection accuracy and improves the cooperative efficiency of USV swarm.

[0006] To achieve the above objectives, this invention provides a multimodal dynamic cooperative unmanned surface vessel (USV) swarm autonomous obstacle avoidance method, wherein the USV swarm includes a leader and followers, and includes the following steps:

[0007] S0 acquires multimodal sensor data and performs data preprocessing;

[0008] S1 uses a cross-modal attention mechanism and deformable convolution to fuse LiDAR bird's-eye view and camera perspective features in multimodal dynamic perception fusion, generating a time-varying map of the fused multimodal features;

[0009] S2 updates the disturbed edge set of the dynamic community of the unmanned surface vessel swarm using an incremental community discovery model on the time-varying graph of the fused multimodal features, and selects the leader based on the threat level;

[0010] The S3 leader uses an improved speed obstacle model to define a multi-objective function that considers path length, smoothness, and safety to plan a multi-objective path. From the multi-objective paths, it selects the optimal path based on the leader's and followers' paths and performs obstacle avoidance based on the optimal path.

[0011] The above methods address the spatiotemporal alignment problem of heterogeneous sensor data by fusing cross-modal attention mechanisms and deformable convolutions with images from different time and space, such as bird's-eye views, camera perspectives, and radar, in Multimodal Dynamic Perception Fusion (CDPF) to improve obstacle detection accuracy. Incremental Community Discovery (DCDM) optimizes changing communities and edge sets, then redetermines the leader and corrects the paths of followers based on the optimized communities. This allows for the re-determination of task groups based on changes in paths between multiple groups during operation, improving the collaborative efficiency of the unmanned surface vessel (USV) swarm. Finally, an improved velocity obstacle model (IVO-DWA) defines multiple target paths that balance path length, smoothness, and safety. Among these multiple target paths, the optimal path is selected based on the balance of these three factors within the leader's community. This ensures that the leader leads followers to achieve real-time, safe, and compliant swarm obstacle avoidance, resulting in better collaboration and higher obstacle avoidance efficiency.

[0012] Furthermore, step S0 includes:

[0013] Multimodal sensor data is synchronized via the IEEE 1588 protocol; spatial alignment is achieved by calculating the LiDAR-camera extrinsic parameter matrix using a checkerboard calibration method under time alignment conditions.

[0014] The above settings achieve spatiotemporal alignment between different modalities, thereby facilitating subsequent data fusion between different modalities.

[0015] Furthermore, in step S0, the multimodal sensor data includes LiDAR point clouds, camera images, radar data, and AIS data.

[0016] The above settings can generate sensor data from multiple different dimensions of data.

[0017] Furthermore, step S1 includes: steps S1.1 to S1.2,

[0018] S1.1 The preset cross-modal attention weight matrix is ​​as follows: ;

[0019] S1.2 uses a cross-modal attention weight matrix Deformable convolution fuses geometric features from LiDAR bird's-eye view and camera perspective view to form fused multimodal features. .

[0020] The above settings are achieved through a cross-modal attention weight matrix. Furthermore, deformable convolution enables the spatial dimension of the formed multimodal fusion features to be correlated with the channels, thereby improving the obstacle detection accuracy affected by the spatiotemporal reference differences of different sensor data.

[0021] Furthermore, in step S1.1, the cross-modal attention weight matrix Representing modes For modes The attention weights reflect the cross-modal feature correlation, as shown in formula (1):

[0022] ,

[0023] in Indicates the first The original features of each modality Indicates the first The original features of each modality Represent a linear mapping function and generate queries and keys respectively. Representing feature dimension, Indicates spatial dimension;

[0024] In step S1.2, the fused multimodal features As shown in formula (2):

[0025] (2),

[0026] This indicates the number of feature sources participating in the fusion. It represents a linear mapping function and generates a value vector.

[0027] The above settings, through the attention mechanism, increase the weight of queries and keys, and associate multiple modalities with each other, jointly optimizing the correlation between channels and spatial dimensions, thereby significantly improving obstacle detection accuracy.

[0028] Furthermore, step S2 includes steps S2.1-S2.3. Step S2.1 dynamically updates the edge weights based on the incremental modularity of the time-varying graph, thereby updating the disturbed edge set of the dynamic community of the unmanned surface vessel swarm. As shown in formula (3):

[0029] (3),

[0030] Updated based on the location of the unmanned surface vessel and mission similarity. Represents a node The degree, Represents a node The degree, This represents the total number of edges in the sequence diagram. This represents the preset modularity resolution parameter, used to control the dynamic community granularity. Whether they belong to the same dynamic community. If they belong to the same dynamic community, then... ,

[0031] S2.2 Based on Threat Level Select the leader in the unmanned surface vessel swarm;

[0032] S2.3 adjusts the follower path in the repulsive potential field of the unmanned surface vessel swarm using a potential field function.

[0033] The above settings determine the disturbed edge set by updating the variable weights based on the position and resolution of adjacent nodes, the position of the unmanned vessel, and the similarity of the task. This enables dynamic community detection by processing only the changed edge set through local updates, and these updates are based on the position and task similarity of the unmanned vessel. This reduces the amount of computation and ensures the reliability of the edge set updates.

[0034] Furthermore, in step S2.2, the threat level... The calculation is shown in formula (4):

[0035] (4),

[0036] in For the detected changes in the speed of the obstacle, The magnitude of the obstacle's acceleration reflects the degree of drastic change in the obstacle's motion state; Indicates the first Current location of the unmanned surface vessel To all obstacle locations The minimum Euclidean distance; It is a very small positive number. .

[0037] The above settings are configured through the first step. Current location of the unmanned surface vessel To all obstacle locations The minimum Euclidean distance is used to measure the proximity of the unmanned surface vessel (USV) to an obstacle. It is obtained by comparing the acceleration modulus with the minimum Euclidean distance, thereby determining the threat level. Identify the optimal leader in the unmanned surface vessel swarm to optimize the collaborative obstacle avoidance strategy.

[0038] Furthermore, in step S2.3, the potential field function is calculated as shown in formula (5):

[0039] (5),

[0040] in This represents the repulsive force vector experienced by the follower UAV, used to correct its path to avoid collisions with other UAVs in the same group. The preset repulsion coefficient; It is a Gaussian decay function. unmanned surface vessel and The Euclidean distance to the current location. The preset attenuation coefficient; From unmanned surface vessel Pointing to unmanned boat The unit direction vector.

[0041] The above settings reflect the task group through the repulsion vector. The system can correct obstacles for other unmanned surface vessels (USVs) that are currently following the target; and the repulsion vector is determined by the Euclidean distance, the repulsion coefficient, and the preset attenuation coefficient. The force function is related to the distance, which ensures that the path correction of the target follower is more reliable.

[0042] Furthermore, step S3 also includes: steps S3.1-S3.2,

[0043] S3.1 Defines a multi-objective function for path length, smoothness, and safety in the Improved Velocity-Boundary Model (IVO-DWA);

[0044] S3.2 The final heading angle of the followers in the unmanned surface vessel (USV) swarm is determined based on the current heading angle of the leader in the USV swarm, as shown in formula (6):

[0045] (6),

[0046] in This indicates the final heading angle of the follower unmanned surface vessel, used to adjust its course to achieve cooperative obstacle avoidance; The current heading angle of the leader unmanned surface vessel is used as a reference for path planning; Represents the power function, It represents the control cycle time, used to convert the instantaneous effect of the repulsive force vector into incremental adjustments to the heading angle.

[0047] The above settings, by combining distance decay and direction correction, dynamically adjust the follower's trajectory to ensure the current task group's safety. The system ensures safe obstacle avoidance for all unmanned surface vessels (USVs) by combining the leader's heading reference with dynamic repulsion corrections. This allows for real-time updates to the followers' headings, ensuring they follow the leader's path while avoiding collisions with the task force during obstacle avoidance. It can avoid collisions with other unmanned surface vessels within the vessel, thus achieving real-time dynamic collaborative obstacle avoidance.

[0048] Furthermore, in step S3.1, the multi-objective function is calculated as shown in formula (7):

[0049] (7),

[0050] in Let represent the heading angle and velocity of the unmanned surface vessel, respectively, which are the control variables that need to be optimized; This represents the total number of steps within the planning timeframe, limiting the duration of the optimization process; different optimization objectives are represented within the summation term of the multi-objective function: Indicates path length weight. Indicates the time of the unmanned surface vessel The position vector, Represents the target point position vector. In the form of vector norm, it represents the unmanned surface vessel in time. Location To the target point Euclidean distance in n-dimensional space; Indicates the path smoothness weight. Indicates the heading angle in time The value, This indicates the magnitude of the change in heading angle between adjacent time steps, used to suppress abrupt changes in heading. Represents security weights. Indicates the time of the obstacle The position vector, This indicates the safe obstacle avoidance distance threshold; the obstacle's position is below the safe threshold. Imposing penalties at the time to force compliance with obstacle avoidance safety distances. This indicates the minimum distance between the unmanned surface vessel and an obstacle.

[0051] The above settings can determine the leader's navigation angle and speed by using a weighted summation of multiple objective functions, based on the minimum Euclidean distance and travel distance corresponding to different objective points. This can comprehensively consider path length, balance, and safety to find the optimal path, improve the efficiency of unmanned surface vessel (USV) path optimization in path planning, enhance stability through smooth turning of the USV, and ensure safety through compliant dynamic real-time obstacle avoidance. Thus, it achieves dynamic balance optimization of path and real-time obstacle avoidance among multiple objectives. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of a scenario for collaborative obstacle avoidance by a swarm of unmanned surface vessels according to the present invention.

[0053] Figure 2 This is a diagram of the CoDAC algorithm framework of the present invention.

[0054] Figure 3 This is a flowchart of the CoDAC algorithm of the present invention.

[0055] Figure 4 This is a schematic diagram illustrating the decomposition and verification results of the core modules in this invention.

[0056] Figure 5 This is a schematic diagram illustrating the evolution of population stability under dynamic obstacles in this invention.

[0057] Figure 6 This is a schematic diagram of the sensor noise and large-scale cluster robustness verification results in this invention.

[0058] Figure 7 This is a schematic diagram showing the performance comparison results of path planning in this invention.

[0059] Figure 8 This is a schematic diagram illustrating the comparative analysis results of obstacle avoidance trajectories in a multi-ship intersection scenario in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1-3 As shown, a multimodal dynamic cooperative unmanned surface vessel (USV) swarm autonomous obstacle avoidance method is proposed. The USV swarm includes a leader and followers, and includes the following specific steps:

[0062] S0 acquires multimodal sensor data and performs data preprocessing. The multimodal sensor data includes LiDAR point clouds, camera images, and radar data. For data from different sensors in the multimodal dataset, under the condition of time alignment, the LiDAR-camera extrinsic parameter matrix is ​​calculated using the checkerboard calibration method for spatial alignment. Spatiotemporal alignment is achieved between different modalities, which facilitates subsequent data fusion between different modalities. The checkerboard calibration method is an existing technology and will not be described in detail here.

[0063] In multimodal dynamic perception fusion (CDPF), step S1 synchronizes multimodal sensor data via the IEEE 1588 protocol. In this embodiment, time synchronization via the IEEE 1588 protocol is existing technology and will not be elaborated further. A time-series map of the fused multimodal features is generated by fusing LiDAR bird's-eye view (BEV) and camera perspective (RV) features through a cross-modal attention mechanism and deformable convolution. Steps S1.1 to S1.2 are also included.

[0064] S1.1 The preset cross-modal attention weight matrix is ​​as follows: Cross-modal attention weight matrix Representing modes For modes The attention weights reflect the cross-modal feature correlation, as shown in formula (1):

[0065] ,

[0066] in Indicates the first The original features of each modality Indicates the first The original features of each modality This represents a linear mapping function and generates queries and keys respectively. Representing feature dimension, In this embodiment, spatial dimension is represented. ;

[0067] S1.2 uses the cross-modal attention weight matrix as follows: Deformable convolutions fuse geometric features from LiDAR bird's-eye view (BEV) and camera perspective view (RV) to form fused multimodal features. As shown in formula (2):

[0068] (2),

[0069] This indicates the number of feature sources participating in the fusion. It represents a linear mapping function and generates a value vector.

[0070] By fusing cross-modal weights with deformable convolutional geometric features, the joint optimization channel after multimodal fusion is correlated with the spatial dimension, thereby improving the obstacle detection accuracy affected by the spatiotemporal benchmark differences of different sensor data. At the same time, the fused multimodal features provide high-precision environmental semantic information for subsequent path planning of unmanned surface vessels.

[0071] S2 updates the perturbed edge set of the dynamic community of the unmanned surface vessel (USV) swarm using incremental community discovery (DCDM) on the time-varying graph of the fused multimodal features. In this embodiment of incremental community discovery (DCDM), the edge weights are dynamically updated based on the incremental modularity of the time-varying graph, thereby updating the perturbed edge set of the dynamic community of the USV swarm. The incremental modularity is set to... As shown in formula (3):

[0072] (3),

[0073] The update is based on the unmanned surface vessel's (USV) location and mission similarity, specifically determined by a pre-defined USV location-mission similarity table. Represents a node The degree, Represents a node The degree, This represents the total number of edges in the sequence diagram. This represents the modularity resolution parameter, with a value of 1.2, used to control the dynamic community granularity. Represents a node and Whether they belong to the same dynamic community. If they belong to the same dynamic community, then... ;otherwise ,

[0074] S2.2 Based on Threat Level Select the leader in the unmanned surface vessel swarm.

[0075] After establishing cooperation rules within the task group, each dynamic community's unmanned surface vessel swarm is evaluated based on threat level. Select the leader in the unmanned surface vessel swarm, threat level The calculation is shown in formula (4):

[0076] (4),

[0077] in The magnitude of the obstacle's acceleration is the speed of the obstacle, which can be detected by a speed sensor, reflecting the degree of change in the obstacle's motion state. Indicates the first Current location of the unmanned surface vessel To all obstacle locations The minimum Euclidean distance is used to measure the proximity of an unmanned surface vessel to an obstacle. It is a very small positive number. To avoid the denominator being zero in equation (4).

[0078] S2.3 adjusts the follower path in the repulsive potential field of the unmanned surface vessel swarm using a potential field function.

[0079] In this embodiment, the follower path in the repulsive potential field of the unmanned surface vessel cluster is adjusted by the potential field function to avoid collision between followers within the group. This is also achieved by formula (5), except that both are followers.

[0080] The S3 leader plans the path based on the Improved Velocity-DWA model, and the followers correct their trajectories through a repulsive potential field, collaboratively completing multi-target real-time obstacle avoidance. This also includes steps S3.1-S3.3.

[0081] S3.1 defines a multi-objective function for path length, smoothness, and safety in the Improved Velocity-DWA model.

[0082] In the Improved Velocity-DWA model, a multi-objective function is defined to determine path length, smoothness, and safety. The calculation of the multi-objective function is shown in Equation (7):

[0083] (7),

[0084] in Let represent the heading angle and velocity of the unmanned surface vessel, respectively, which are the control variables that need to be optimized; This represents the total number of steps within the planning timeframe, limiting the duration of the optimization process; different optimization objectives are represented within the summation term of the multi-objective function:

[0085] This indicates the preset path length weight. Indicates the time of the unmanned surface vessel The position vector, Represents the target point position vector. In the form of vector norm, it represents the unmanned surface vessel in time. Location To the target point The Euclidean distance in n-dimensional space is used to minimize the navigation distance; the distance between the current position and the target point can be obtained through a distance detection sensor.

[0086] This represents the preset path smoothness weight. Indicates the heading angle in time The value, This indicates the magnitude of the change in heading angle between adjacent time steps, used to suppress abrupt changes in heading.

[0087] This indicates the preset security weight. Indicates the time of the obstacle The position vector, This indicates the safe obstacle avoidance distance threshold; the obstacle's position is below the safe threshold. Imposing penalties at the time to force compliance with obstacle avoidance safety distances. In this embodiment, the minimum distance between the unmanned surface vessel and the obstacle is indicated. The position vector of the obstacle is also obtained through the obstacle detection sensor.

[0088] This approach, through the weighted summation of multiple objective functions, can improve the efficiency of unmanned surface vessel (USV) path optimization in path planning, enhance stability through smooth turning of USVs, and ensure safety through compliant dynamic real-time obstacle avoidance. This achieves dynamic balance among multiple objectives, optimizing paths and real-time obstacle avoidance, enabling leaders to optimize multi-objective paths through multi-objective functions and leading followers to achieve real-time, safe, and compliant cluster obstacle avoidance.

[0089] S3.2 The final heading angle of the followers in the unmanned surface vessel (USV) swarm is determined based on the current heading angle of the leader in the USV swarm, as shown in formula (6):

[0090] (6),

[0091] in This indicates the final heading angle of the follower unmanned surface vessel, used to adjust its course to achieve cooperative obstacle avoidance; The current heading angle of the leader unmanned surface vessel is used as a reference for path planning; Let represent the force function, i.e., the repulsive force vector experienced by the follower unmanned surface vessel; the current heading angle is obtained through equation (7). It represents the control cycle time, used to convert the instantaneous effect of the repulsive force vector into incremental adjustments to the heading angle.

[0092] In this embodiment, The follower path in the repulsive potential field of the unmanned surface vessel swarm is adjusted by the potential field function, which is calculated as shown in formula (5):

[0093] (5),

[0094] in This represents the repulsive force vector experienced by the follower UV, used to correct its path to avoid collisions with other UVs in the same group, reflecting the mission group's... The obstacle avoidance correction effect of other unmanned surface vessels on the current follower; This is the repulsion coefficient, used to control the intensity of the repulsive force. It is a Gaussian decay function. unmanned surface vessel and The Euclidean distance to the current location. The attenuation coefficient represents the rate at which the repulsive force weakens with increasing distance. From unmanned surface vessel Pointing to unmanned boat The unit direction vector.

[0095] The key technical indicators of this invention are compared with existing technologies (VO / MPC / DRL) as shown in Table 1 below.

[0096] Table 1 Comparison of Technical Indicators

[0097]

[0098] It can be concluded that the present invention significantly improves the perception accuracy detection, obstacle avoidance distance detection, and COLREGs compliance of unmanned surface vessels.

[0099] In this embodiment, the core module contribution decomposition verification results are as follows: Figure 4 As shown, by systematically disabling the core module, the key role of each module in obstacle avoidance performance was verified.

[0100] Multimodal dynamic perception fusion (CDPF) was disabled in experimental group G1, such as Figure 4 As shown in (a), the minimum obstacle avoidance distance (MAD) of the unmanned surface vessel is reduced by 22.9% compared to the present invention (CoDAC) (32.1m vs 53.6m, p=0.003), as Figure 4 As shown in Figure (d), the cross-modal feature alignment error (FE) increases significantly by 171% (2.5px → 6.7px) compared to the present invention (CoDAC). The I-shaped symbol at the top of each bar in the figure represents the fluctuation range of the corresponding value. Figure 4 In (d), the maximum value of the present invention (CoDAC) is 2.5px, and the maximum value of experimental group G1 is 6.7px, indicating that multimodal dynamic perception fusion (CDPF) can effectively improve the feature fusion accuracy of heterogeneous sensors through the dual attention mechanism, providing a reliable perception basis for obstacle localization.

[0101] In experimental group G2, the Dynamic Community Discovery Mechanism (DCDM) was disabled, such as Figure 4As shown in (c), the Task Group Partition Consistency (TGCI) of the unmanned surface vessel decreased by 20.9% (0.91→0.72), which confirms that the Dynamic Community Discovery Mechanism (DCDM) updates the edge weights by incremental modularity and then updates the disturbed edge set of the dynamic community, thereby optimizing the dynamic task group partitioning and reducing communication complexity.

[0102] In experimental group G3, the traditional speed obstacle course method was used, such as... Figure 4 In the middle (e), the Rule Compliance Score (RCS) of unmanned surface vessels plummeted by 30.6% (100%→69.4%), and the emergency obstacle avoidance success rate dropped from 96.7% to 72.3%, highlighting the decisive role of the Improved Speed ​​Obstacle Model (IVO-DWA) and COLREGs coding in maritime rule compliance.

[0103] In experimental group G4, when both multimodal dynamic perception fusion (CDPF) and dynamic community detection (DCDM) are disabled, the performance degradation of the unmanned surface vessel exhibits a non-linear, cumulative effect. Figure 4 In (b), the collision rate (CR) of experimental group G4 increased by 12 times (0.3%→3.6%) relative to the present invention, indicating that the multi-module collaboration forms a closed-loop optimization mechanism of perception-decision-coordination.

[0104] The above significance analysis (ANOVA, p<0.01) further verified the overall contribution of the core modules: Dynamic Community Detection Mechanism (DCDM) (TGCI improved by 26.4%), Improved Velocity Barrier Model (IVO-DWA) (RCS improved by 30.6%), and Multimodal Dynamic Perception Fusion (CDPF) (MAD improved by 22.9%), which provide multi-dimensional technical support for unmanned surface vessel swarm collaboration in dynamic marine environments.

[0105] like Figure 5 As shown, the stability evolution of an unmanned surface vessel (USV) swarm under dynamic obstacles is illustrated in this embodiment, where N = 100 USVs, demonstrating the evolution of community stability of a 100 USV swarm in a dynamic obstacle environment.

[0106] This invention is based on a dynamic community discovery mechanism (DCDM) with incremental modularity optimization (e.g., Figure 5 (as shown by the solid line) and the experimental group G2 fixed grouping strategy (as shown by the solid line) Figure 5 As shown by the dashed line, the Community Stability Index (CSI) is quantified using the Jaccard similarity metric to reflect the degree of overlap among community members at adjacent times. In the no-mutation phase (t=0-19), the mean CSI of the Dynamic Community Discovery Mechanism (DCDM) is equivalent to 0.92±0.03, which is 0.85±0.05 in the experimental group G2. This indicates that the dynamic mechanism can maintain higher consistency under steady state, and the dynamically adjusted community structure significantly improves the robustness of the system.

[0107] When encountering a sudden change in obstacle velocity (Δv=5m / s², t=20 / 35 steps), the CSI of the dynamic community detection mechanism (CoDAC) only dropped to 0.87±0.04, while it plummeted to 0.58±0.07 in experimental group G2, confirming that dynamic reorganization strategies (such as local modularity gain calculation) can mitigate the impact of sudden disturbances.

[0108] The core advantages of the Dynamic Community Detection (DCDM) mechanism are reflected in two aspects: First, the incremental update efficiency, which only processes the affected disturbed edge set (ΔE), reduces the number of recombinations by 58.3% compared to the experimental group G2 (5 vs 12 times), and reduces the computational complexity from O(N²) to O(|ΔE|logN), meeting the real-time requirements; Second, the structural adaptability, which balances the community size and communication load by using the resolution parameter γ=1.2, and quickly restores stability after mutation (CSI rises to 0.88±0.03 after t=35-49 steps), avoiding the oscillation decay of fixed groups.

[0109] Error band analysis further revealed the robustness differences. The standard deviation of the dynamic community discovery mechanism (CoDAC) (0.03-0.04) was significantly lower than that of the experimental group G2 (0.05-0.07), verifying the tolerance of the dynamic mechanism to noise. This result is consistent with the "diversity-stability" theory in ecology: the modular network architecture maintains global stability by enhancing intraspecific interactions (such as potential field correction), thus providing a theoretical basis for the engineering deployment of large-scale clusters. It has been applied to port patrol scenarios (reducing communication load by 68%).

[0110] like Figure 6 As shown, sensor noise and robustness to large-scale clusters were verified. A hybrid noise model (LiDAR ranging noise σ=0.3m, camera randomly occludes 40%-60% of pixels) was constructed using measured parameters from Velodyne VLP-16 and compared with DRL-CA and DynaMOCO methods.

[0111] like Figure 6 As shown in (a), the experimental results indicate that:

[0112] 1) In terms of cross-modal complementarity, multimodal dynamic perception fusion (CDPF) can effectively suppress single-modal noise interference through cross-modal attention mechanism. The cross-modal feature alignment error (FE) under LiDAR noise is significantly better than the error increase of DRL-CA.

[0113] 2) In terms of dynamic feature repair capability, the spatial-channel dual attention (SCA) of the present invention (CoDAC(CR)) still maintains a collision rate CR<1.5% under 50% occlusion rate, which is 57% lower than DRL-CA (3.6%→1.5%, p=0.007).

[0114] 3) Noise sensitivity analysis shows that the minimum obstacle avoidance distance (MAD) in mixed noise environment is strongly linearly correlated with noise intensity (R²=0.93). However, the slope of the present invention (CoDAC (MAD) in the figure) is only 48% of that of DRL-CA (0.25m / σ), which proves that the dynamic community detection mechanism (DCDM) has stronger noise robustness and verifies the adaptability advantage of the cross-modal dynamic sensing architecture in complex environments.

[0115] like Figure 6 As shown in (b), this is a large-scale cluster test.

[0116] In large-scale cluster testing, this embodiment constructs a simulation scenario involving 10-100 heterogeneous USVs (unmanned surface vessels) (thruster power differences ±30%), introducing 30% communication packet loss and ±20Mbps bandwidth fluctuation interference conditions. Comparative analysis is performed with centralized MPC (model predictive control) and distributed PSO (particle swarm optimization) methods. Test results show that:

[0117] 1) In terms of distributed scalability, the Dynamic Community Discovery Mechanism (DCDM) achieves sublinear growth in planning delay through incremental modularity optimization. When the number of nodes increases from 50 to 100, the delay slope k=1.8ms / ship (R²=0.98), which is 65% lower than the linear growth of MPC (k=5.2ms / ship).

[0118] 2) In terms of anti-interference capability, under the scenario of 30% packet loss, the leader election mechanism (Equation 9) maintains the consistency of the task group TGCI>0.75, and the communication load GCL only increases by 12% (25%→28%), which is significantly better than the 45% increase of PSO;

[0119] 3) In terms of resource efficiency, the memory usage was optimized to 48MB for a scale of 100 vessels, which is 9.2% of the MPC solution (520MB→48MB). The maximum vertical column height of MPC is 520MB, which meets the deployment requirements of edge computing devices (such as Jetson Xavier) and verifies the engineering practicality of the hierarchical collaborative architecture in complex sea conditions.

[0120] like Figure 7 As shown, a path planning performance comparison was conducted (multi-ship intersection scenario, N=20 unmanned surface vessels), demonstrating the path planning performance comparison of different obstacle avoidance algorithms in the multi-ship intersection scenario.

[0121] Experimental data show that the CoDAC method of this invention is significantly superior to traditional methods in terms of safety, efficiency, and compliance, specifically:

[0122] like Figure 7 As shown in Figure (a), the minimum obstacle avoidance distance (MAD) of this invention (CoDAC) reaches 53.6 meters, which is 63.9% higher than that of the traditional speed obstacle avoidance (VO) method. Figure 7 (a) By using an improved Triangular Obstacle Zone (TOZ) model, the position of dynamic obstacle vertices can be predicted, thus expanding the safe avoidance boundary.

[0123] like Figure 7 As shown in Figure (b), the collision rate (CR) of this invention (CoDAC) is only 0.3%, which is 66.7% lower than the suboptimal method DynaMOCO. Figure 7 (b) is attributed to the task group division of the Dynamic Community Discovery Mechanism (DCDM), which reduces the probability of conflict within the task group through modularity optimization.

[0124] like Figure 7 As shown in (c), the path length ratio (PLR) of this invention (CoDAC) is 1.12 ( Figure 7 The results (c) show that multimodal perception fusion improves obstacle localization accuracy (AP@0.5IoU=89.7%) and reduces path redundancy, outperforming all the comparison methods.

[0125] like Figure 7 As shown in (d), the real-time performance of this invention (CoDAC) is reflected in the computational latency metric. Figure 7 In the middle (d), the computation latency of CoDAC is only 95ms, which is 70.3% lower than that of centralized MPC. Because its distributed architecture only needs to process the disturbed edge set (ΔE), the complexity is reduced from O(N²) to O(|ΔE|logN).

[0126] like Figure 7 As shown in (e), COLREGs compliance ( Figure 7 In the IVO-DWA model, CoDAC achieves zero violations, while traditional VO suffers from a violation density of 4.7 violations per m² due to the lack of rule encoding. This is thanks to the rule priority matrix embedded in the IVO-DWA model and the NSGA-II multi-objective optimization, which simultaneously balances path smoothness and rule constraints under velocity window constraints.

[0127] In summary, CoDAC provides a highly secure, efficient, and strictly compliant obstacle avoidance solution for large-scale USV clusters through a closed-loop optimization mechanism of perception-planning-coordination.

[0128] like Figure 8As shown, a comparative analysis of obstacle avoidance trajectories in a multi-ship intersection scenario is presented, demonstrating the comparative analysis of obstacle avoidance trajectories of three different ship collision avoidance technologies (CoDAC, VO, and DRL-CA) in a multi-ship intersection scenario. Figure 8 The invention (CoDAC), VO, and DRL-CA methods were evaluated from different dimensions:

[0129] from Figure 8 As can be seen in Figure a, the trajectory of the blue solid line 1 of the present invention (CoDAC) is relatively smooth, and it follows the rule well when approaching “COLREGsRule 15 Boundary” (located at a longitudinal position of 150m); the red dashed line 2 of the VO method fluctuates greatly, and the lateral position changes drastically in the longitudinal position range of 100-200m, deviating from the expected rule; the green dotted line 3 of the DRL-CA method has obvious lateral position fluctuations near the longitudinal position of 150m, and the trajectory stability is not as good as CoDAC.

[0130] from Figure 8 As can be seen from Figure b, the blue solid line 1 of the present invention (CoDAC) consistently maintains a minimum distance of approximately 53.6m, staying away from the conflict zone below 50m; the red dashed line 2 of the VO method fluctuates between 20 and 35m in minimum distance within the time range of 0-60s, repeatedly entering the conflict zone; while the green dotted line 3 of the DRL-CA method stably maintains a minimum distance of 100m, demonstrating high safety.

[0131] from Figure 8 As can be seen from the data, the dynamic safety margin (light blue filled area) around the trajectory of the blue solid line 1 of the present invention (CoDAC) is reasonable, with the margin fluctuating between -10m and 10m within the longitudinal position of 0-300m; the safety margin of the red dashed line 2 of the VO method is narrower, and the trajectory fluctuates greatly; the safety margin (light green filled area) of the green dotted line 3 of the DRL-CA method is within the range of -15m to 15m within the longitudinal position of 0-300m, which can better ensure safety.

[0132] The working principle of this invention is as follows: First, LiDAR point clouds, camera images, radar data, and AIS data are synchronized via the IEEE 1588 protocol, and cross-modal spatial alignment is achieved based on the extrinsic parameter matrix. Second, a cross-modal attention mechanism and deformable convolution are combined to establish the correlation between channels and spatial dimensions, enabling the fused multimodal features to generate high-precision environmental semantic information. Next, the unmanned surface vessel (USV) task groups are dynamically divided based on the incremental modularity of the time-varying graph, and only the disturbed edge set is updated to reduce computational complexity. Finally, the USV leader uses an improved velocity obstacle avoidance model (IVO-DWA) to optimize the multi-target path, and the followers correct their trajectories through a repulsive potential field, achieving real-time, safe, and compliant cluster obstacle avoidance.

Claims

1. A multi-modal dynamic cooperative unmanned surface vehicle (USV) swarm autonomous obstacle avoidance method, the USV swarm comprising a leader and followers, the method comprising the following steps: S0: obtaining multi-modal sensor data and performing data preprocessing; S1: generating a time-varying map of fused multi-modal features by fusing LiDAR bird's-eye view and camera perspective view features through a cross-modal attention mechanism and deformable convolution in multi-modal dynamic perception fusion; S2: updating a disturbed edge set of a dynamic community of the USV swarm through incremental community discovery based on the time-varying map of fused multi-modal features, optimizing division of task groups, and selecting a leader based on threat degree; S3: defining a multi-objective function of path length, smoothness, and safety based on an improved velocity obstacle model to plan a multi-objective path, selecting an optimal path from the multi-objective path according to the leader and the followers, and performing obstacle avoidance according to the optimal path.

2. The method of claim 1, wherein: In step S0, the method comprises: synchronizing multi-modal sensor data through IEEE 1588 protocol; and performing spatial alignment by calculating a LiDAR-camera extrinsic parameter matrix using a checkerboard calibration method under the condition of time alignment.

3. The method of claim 1, wherein: In step S0, the multi-modal sensor data comprises LiDAR point cloud, camera image, and radar data.

4. The method of claim 1, wherein: Step S1 comprises steps S1.1-S1.2, S1.1 preset cross-modal attention weight matrix is ; S1.2 is formed by a cross-modal attention weight matrix for , deformable convolution fuses geometric features of LiDAR bird's eye view and camera perspective view to form a fused multi-modal feature .

5. The method of claim 4, wherein: In step S1.1, a cross-modal attention weight matrix represents the attention weight of the modality to the modality , reflecting the cross-modal feature correlation, as shown in equation (1): ， wherein represents the original features of the th modality, represents the original features of the th modality, represents a linear mapping function and generates queries, keys, respectively, represents the feature dimension, represents the spatial dimension; In step S1.2, the fused multi-modal feature As shown in equation (2): （2）， denotes the number of feature sources participating in the fusion, denotes a linear mapping function and generates a value vector.

6. The method of claim 1, wherein: The step S2 includes steps S2.1-S2.3, the step S2.1 updates the edge weight based on the incremental modularity of the time-varying graph, and further updates the disturbed edge set of the unmanned ship cluster dynamic community As shown in formula (3): （3）， wherein representing a node with at time is updated by the position of the unmanned ship and the task similarity, representing a node degree, representing a node degree, representing the total number of edges of the time sequence diagram, representing a preset module degree resolution parameter, used for controlling the dynamic community granularity, representing a node with whether they belong to the same dynamic community, if the same dynamic community, then , S2.2 based on threat level Select a leader in the swarm of unmanned surface vehicles; S2.3: adjusting a follower path in a repulsive potential field of the USV swarm through a potential field function.

7. The method of claim 6, wherein: In step S2.2, the threat degree is calculated as shown in equation (4): （4）， wherein is the detected obstacle velocity change, is the modulus of the obstacle acceleration, reflecting the degree of change in the obstacle motion state; represents the first current position of the unmanned ship to the minimum Euclidean distance from all obstacle positions for measuring the proximity of the unmanned ship to the obstacle; is a very small positive number, .

8. The method of claim 1, wherein: The step S3 further comprises steps S3.1-S3.2, S3.1: defining a multi-objective function of path length, smoothness, and safety in an improved velocity obstacle model (IVO-DWA); S3.2: determining a final heading angle of a follower in the USV swarm according to a current heading angle of a leader in the USV swarm, specifically as shown in formula (6): （6）， wherein represents the final heading angle of the follower USV for adjusting its sailing direction to achieve the cooperative obstacle avoidance; represents the current heading angle of the leader USV as the path planning reference; represents the potential field function, represents the time of the control period for converting the instantaneous action of the repulsive force vector into the incremental adjustment of the heading angle.

9. The method of claim 6, wherein: In step S2.3, the potential field function is calculated as shown in formula (5): （5）， wherein represents the repulsion vector that the follower UAV receives to correct its path to avoid collision with other UAVs in the same group, reflecting the task group other UAVs' obstacle avoidance correction effect on the current follower; is the repulsion coefficient, used to control the strength of the repulsion effect; is a Gaussian decay function, represents the Euclidean distance between the current position of the UAV and the position of the other UAVs in the same group; is the decay coefficient, representing the weakening rate of the repulsion with the increase of the distance; is the unit direction vector from the UAV to the other UAVs in the same group; .

10. The method of claim 8, wherein: In step S3.1, the multi-objective function is calculated as shown in formula (7): （7）， wherein respectively represent the heading angle and velocity of the USV, which are control variables to be optimized; N represents the total number of steps in the planning time range, which limits the duration of the optimization process; Different optimization objectives correspond to the summation term of the multi-objective function: Indicates path length weight. Indicates the time of the unmanned surface vessel The position vector, Represents the target point position vector. Indicates the time of the unmanned surface vessel Location To the target point The Euclidean distance is used to minimize the sailing distance; Indicates the path smoothness weight. Indicates the heading angle in time The value, This indicates the magnitude of the change in heading angle between adjacent time steps, used to suppress abrupt changes in heading. Represents security weights. Indicates the time of the obstacle The position vector, This indicates the safe obstacle avoidance distance threshold; the obstacle's position is below the safe threshold. Imposing penalties at the time to force compliance with obstacle avoidance safety distances. This indicates the minimum distance between the unmanned surface vessel and an obstacle.

Images