Multi-mode dynamic cooperative unmanned ship cluster autonomous obstacle avoidance method

Abstract

The invention provides a multi-modal dynamic cooperative unmanned ship cluster autonomous obstacle avoidance method, which comprises the following steps: firstly, synchronizing LiDAR point cloud, camera images, radar data and AIS data through an IEEE 1588 protocol, and realizing cross-modal space alignment based on an external parameter matrix; secondly, a cross-modal attention mechanism and deformable convolution are combined to establish correlation between a channel and spatial dimensions, so that fused multi-modal features generate high-precision environmental semantic information; then, dynamically dividing unmanned ship task groups based on time-varying graph incremental modularity, and only updating a disturbed edge set to reduce calculation complexity; and finally, the unmanned ship leader optimizes a multi-target path by adopting an improved speed obstacle model (IVO-DWA), and the follower corrects the trajectory through a repulsive force potential field, so that real-time, safe and compliant cluster obstacle avoidance is realized.

Description

Technical Field

[0001] The present invention relates to the field of shipborne system information communication technology, and in particular to an autonomous obstacle avoidance method for a multi-modal dynamic collaborative unmanned boat cluster. Background Art

[0002] As a kind of intelligent water equipment capable of autonomous navigation, unmanned boat formations have attracted much attention because of their maneuverability, flexibility, economy and efficiency. They can enhance human initiative in facing unknown waters and replace humans in performing dangerous water operations. This has enabled unmanned boat formations to play a role in both public and civilian fields. In the civilian field, unmanned boat formations can be used for hydrological exploration, environmental maintenance, post-disaster search and rescue, etc., which can reduce labor expenses and improve operational efficiency; in the public field, unmanned boat formations can be used for anti-submarine and anti-missile, water blockade, regional patrol, etc., which can reduce casualties and improve cost-effectiveness.

[0003] For example, the patent document with Chinese patent application number 202410732541.7, classification number G05D1 / 43, and publication date 2024.07.09 discloses a method for tracking and avoiding obstacles of an unmanned boat formation, which belongs to the field of motion control technology of an unmanned boat formation. By measuring obstacles, unmanned boat formations, channel status information and desired target points sent by sensors and shore-based signals, a collision avoidance guidance law based on an improved artificial potential field method meets the collision avoidance requirements and complies with the International Maritime Collision Prevention Regulations to plan a feasible path, and plans the unmanned boat formation plan based on a formation transformation strategy of a hazard evaluation function. Finally, a control law is generated in real time based on an RBF neural network estimator and a backstepping controller and executed by an actuator to drive the unmanned boat 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 above-mentioned literature mainly adopts an improved artificial potential field method combined with a formation transformation strategy. By obtaining the motion state of the unmanned boat and the environmental obstacle information, a manipulation motion model is established, the attraction and repulsion functions are calculated, and a feasible path is planned. The RBF neural network and the backstepping method are used to generate control instructions to achieve adaptive tracking and obstacle avoidance of the formation. Since it directly processes data in different time and space such as cameras and radars, the heterogeneous sensors of different modes have large cross-modal feature alignment errors due to the difference in time and space references, which easily leads to low obstacle detection accuracy and cannot be applied to applications with high precision requirements. In addition, it is directly based on changing the triangular relationship between the leader and the follower when encountering an obstacle, but it cannot dynamically adjust the grouping of the community and then make the optimal path selection according to different grouping situations to achieve better path control. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-modal dynamic collaborative unmanned boat cluster autonomous obstacle avoidance method with high obstacle detection accuracy and improved unmanned boat cluster collaboration efficiency.

[0006] To achieve the above objectives, the present invention provides a multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method, wherein the unmanned boat swarm includes a leader and a follower, comprising the following steps: S0 acquires multimodal sensor data and performs data preprocessing; In multimodal dynamic perception fusion, S1 fuses LiDAR bird's-eye view and camera perspective features through cross-modal attention mechanism and deformable convolution to generate a time-varying graph of the fused multimodal features; S2 updates the disturbed edge set of the dynamic community of the unmanned boat cluster through the incremental community discovery model on the time-varying graph of the fused multimodal features, and selects the leader based on the threat degree; The S3 leader plans multi-objective paths based on the improved speed obstacle model by defining the multi-objective functions of path length, smoothness and safety, selects the optimal path from the multi-objective paths based on the leader and follower paths, and avoids obstacles according to the optimal path.

[0007] The above method solves the spatiotemporal alignment problem of heterogeneous sensor data by fusing bird's-eye views, camera perspective images, radar, and other images in different time and space through the cross-modal attention mechanism and deformable convolution in multimodal dynamic perception fusion (CDPF) to improve obstacle detection accuracy. It also optimizes the changing communities and edge sets through incremental community discovery (DCDM). Then, the leader is re-determined and the follower paths are corrected based on the optimized communities. This allows the task groups to be re-determined according to the changes in the paths between multiple groups during operation, thereby improving the collaborative efficiency of the unmanned boat swarm. Finally, the improved velocity obstacle model (IVO-DWA) is used to define multiple target paths based on the relationship between path length, smoothness, and safety. Then, among the multiple target paths, the path with a relatively balanced relationship between path length, smoothness, and safety in the leader's community is selected as the optimal path, ensuring that the leader leads the followers to achieve real-time, safe, and compliant cluster obstacle avoidance according to the optimal path, resulting in better collaboration and higher obstacle avoidance efficiency.

[0008] Furthermore, step S0 includes: Multimodal sensor data is synchronized using the IEEE 1588 protocol. Under the condition of temporal alignment, the LiDAR-camera extrinsic parameter matrix is ​​calculated using the checkerboard calibration method for spatial alignment.

[0009] The above settings achieve spatiotemporal alignment between different modalities, thereby facilitating subsequent data fusion between different modalities.

[0010] Furthermore, in step S0, the multimodal sensor data includes LiDAR point cloud, camera image, radar data, and AIS data.

[0011] The above settings can form sensor data for data of multiple different dimensions.

[0012] Furthermore, step S1 includes: steps S1.1 to S1.2, S1.1 preset cross-modal attention weight matrix is ; S1.2 Through the cross-modal attention weight matrix , deformable convolution performs geometric feature fusion on the LiDAR bird's-eye view and the camera perspective image to form a fused multimodal feature .

[0013] The above settings are achieved through the cross-modal attention weight matrix And the deformable convolution can make the spatial dimension of the formed multimodal fusion features correlated with the channel, thereby improving the obstacle detection accuracy affected by the spatiotemporal reference differences of different sensor data.

[0014] Furthermore, in step S1.1, the cross-modal attention weight matrix Indicates modality Modal The attention weight of reflects the cross-modal feature correlation, as shown in formula (1): , in Indicates the The original features of the modes, Indicates the The original features of the modes, Represents a linear mapping function and generates queries and keys respectively, represents the feature dimension, Represents spatial dimensions; In step S1.2, the fused multimodal features , as shown in formula (2): (2), Indicates the number of feature sources involved in fusion, Represents a linear mapping function and generates a vector of values.

[0015] The above settings increase query and key weights through the attention mechanism, correlate multiple modalities, and jointly optimize the correlation between channel and spatial dimensions, thereby significantly improving obstacle detection accuracy.

[0016] Furthermore, the step S2 includes steps S2.1-S2.3, wherein step S2.1 dynamically updates the edge weight based on the incremental modularity of the time-varying graph, thereby updating the disturbed edge set of the dynamic community of the unmanned boat cluster, and the incremental modularity , as shown in formula (3): (3), , updated by the position of the unmanned boat and the similarity of the mission, Representation node The degree, Representation node The degree, represents the total number of edges in the timing graph, Represents the preset modularity resolution parameter, which is used to control the dynamic community granularity. Whether they belong to the same dynamic community, if they are the same dynamic community, then , S2.2 Based on threat level Selecting a leader in the UAV swarm; S2.3 Adjust the follower path in the repulsive potential field of the unmanned boat cluster through the potential field function.

[0017] The above settings determine the disturbed edge set based on the position and resolution between adjacent nodes and the variable weights of the unmanned boat position and task similarity updates, thereby realizing dynamic community discovery through local updates, only processing the changed edge set, and it is updated with the position of the unmanned boat and the task similarity. On the one hand, it can reduce the amount of calculation, and on the other hand, it can ensure the reliability of the edge set update.

[0018] Furthermore, in step S2.2, the threat level The calculation is shown in formula (4): (4), in is the detected obstacle speed change, is the module of the obstacle’s acceleration, reflecting the severity of the change in the obstacle’s motion state; Indicates the Current location of unmanned boats To all obstacles The minimum Euclidean distance; is a very small positive number, .

[0019] The above settings, through Current location of unmanned boats To all obstacles The minimum Euclidean distance is used to measure the proximity between the unmanned boat and the obstacle, which is obtained by comparing the acceleration module and the minimum Euclidean distance, so that the threat level is Identify the optimal leader in a swarm of unmanned watercraft to optimize the collaborative obstacle avoidance strategy.

[0020] Furthermore, in step S2.3, the potential field function is calculated as shown in formula (5): (5), in Represents the repulsive force vector on the follower unmanned boat, which is used to correct its path to avoid collision with other unmanned boats in the same group. is the preset repulsion coefficient; is the Gaussian decay function, Unmanned boat and Euclidean distance to the current location, is the preset attenuation coefficient; From an unmanned boat Pointing to unmanned boats The unit direction vector of .

[0021] The above settings reflect the task group through the repulsive vector The other unmanned boats in the vehicle have an obstacle avoidance correction effect on the current follower; 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, ensuring that the follower path correction improves reliability.

[0022] Furthermore, the step S3 further includes: steps S3.1-S3.2, S3.1 Define the multi-objective function of path length, smoothness, and safety in the improved speed barrier model (IVO-DWA); S3.2 determines the final heading angle of the follower in the unmanned boat cluster based on the current heading angle of the leader in the unmanned boat cluster, as shown in formula (6): (6), in Indicates the final heading angle of the follower unmanned vehicle, which is used to adjust its navigation direction to achieve collaborative obstacle avoidance; Indicates the current heading angle of the leader unmanned boat, which serves as the reference for path planning; represents the power function, Represents the time of the control cycle, which is used to convert the instantaneous effect of the repulsive force vector into an incremental adjustment of the heading angle.

[0023] The above settings, through the comprehensive distance attenuation effect and direction correction, dynamically adjust the follower trajectory to ensure the current task group The safety of all unmanned boats in the team's coordinated obstacle avoidance is ensured by combining the leader's heading reference and repulsion dynamic correction, so that the follower's heading can be updated in real time, ensuring that they follow the leader's path during the obstacle avoidance process while avoiding conflicts with the task group. The vehicle can avoid collisions with other unmanned boats in the vehicle, thus achieving real-time dynamic collaborative obstacle avoidance.

[0024] Furthermore, in step S3.1, the multi-objective function is calculated as shown in formula (7): (7), in They represent the heading angle and speed of the unmanned boat, which are the control variables that need to be optimized; Indicates the total number of steps in the planning time range, limiting the duration of the optimization process; in the summation term of the multi-objective function, it corresponds to different optimization goals: represents the path length weight, Indicates that the unmanned boat is at time The position vector of represents the target point position vector, In the form of vector norm, it represents the time of the unmanned boat Location To the destination The Euclidean distance in n-dimensional space; represents the path smoothness weight, Indicates the heading angle at time The value of Indicates the change amplitude of the heading angle in adjacent time steps, which is used to suppress sudden changes in heading; represents the security weight, Indicates the obstacle at time The position vector of Indicates the safety obstacle avoidance distance threshold. The obstacle position is lower than the safety threshold. When the penalty is imposed, the obstacle avoidance safety distance is forced to be met. Indicates the minimum distance between the unmanned boat and obstacles.

[0025] The above settings can determine the leader's navigation angle and speed through the weighted summation of multi-objective functions, the minimum Euclidean distance corresponding to different target points, the driving distance, etc., so that the path length, balance and safety can be comprehensively considered to find the optimal path. In path planning, the efficiency of unmanned boat path optimization can be improved, and the stability can be improved through the smooth steering of the unmanned boat. At the same time, it can comply with dynamic real-time obstacle avoidance to ensure safety, thereby achieving dynamic balance between multiple objectives and optimizing the path and avoiding obstacles in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1This is a schematic diagram of the cooperative obstacle avoidance scenario of the unmanned boat cluster of the present invention.

[0027] Figure 2 This is the framework diagram of the CoDAC algorithm of the present invention.

[0028] Figure 3 This is a flow chart of the CoDAC algorithm of the present invention.

[0029] Figure 4 This is a schematic diagram of the decomposition verification results of the core module contribution in the present invention.

[0030] Figure 5 Schematic diagram of the evolution of group stability under dynamic obstacles in the present invention.

[0031] Figure 6 Schematic diagram of the verification results of sensor noise and large-scale cluster robustness in the present invention.

[0032] Figure 7 Schematic diagram of the path planning performance comparison results in the present invention.

[0033] Figure 8 This is a schematic diagram of the comparative analysis results of obstacle avoidance trajectories in the multi-ship intersection scenario in the present invention. DETAILED DESCRIPTION

[0034] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] like Figure 1-3 As shown in FIG, a multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method, wherein the unmanned boat swarm includes a leader and a follower, includes the following specific steps: S0 obtains multimodal sensor data and performs data preprocessing. Multimodal sensor data includes LiDAR point cloud, camera image, and radar data. For the data of different sensors in the multimodality, under the condition of time alignment, the checkerboard calibration method is used to calculate the LiDAR-camera extrinsic parameter matrix for spatial alignment. By achieving spatiotemporal alignment between different modalities, it facilitates subsequent data fusion between different modalities. The checkerboard calibration method is an existing technology and will not be described in detail here.

[0036] In multimodal dynamic perception fusion (CDPF), S1 synchronizes multimodal sensor data through the IEEE 1588 protocol. In this embodiment, time synchronization through the IEEE 1588 protocol is a prior art and will not be repeated here. The LiDAR bird's-eye view (BEV) and camera perspective view (RV) features are fused through a cross-modal attention mechanism and deformable convolution to generate a time sequence diagram of the fused multimodal features. It also includes steps S1.1 to S1.2. S1.1 preset cross-modal attention weight matrix is , cross-modal attention weight matrix Indicates modality Modal The attention weight of reflects the cross-modal feature correlation, as shown in formula (1): , in Indicates the The original features of the modes, Indicates the The original features of the modes, Represents a linear mapping function and generates queries and keys respectively. represents the feature dimension, represents the spatial dimension. In this embodiment, ; S1.2 uses the cross-modal attention weight matrix as , deformable convolution performs geometric feature fusion on LiDAR bird's eye view (BEV) and camera perspective view (RV) to form fused multimodal features , as shown in formula (2): (2), Indicates the number of feature sources involved in fusion, Represents a linear mapping function and generates a value vector.

[0037] 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 temporal and spatial reference differences of different sensor data. At the same time, the fused multimodal features provide high-precision environmental semantic information for subsequent unmanned boat path planning.

[0038] S2 updates the disturbed edge set of the dynamic community of the unmanned boat cluster through incremental community discovery (DCDM) on the time-varying graph of the fused multimodal features. In the incremental community discovery (DCDM) of this embodiment, the edge weights are dynamically updated by the incremental modularity based on the time-varying graph, and then the disturbed edge set of the dynamic community of the unmanned boat cluster is updated. The incremental modularity is set to , as shown in formula (3): (3), , is updated by the unmanned boat position and task similarity, which is specifically determined by the preset unmanned boat position and task similarity relationship table, Representation node The degree, Representation node The degree, represents the total number of edges in the timing graph, Indicates the modularity resolution parameter, with a value of 1.2, which is used to control the granularity of dynamic communities. Representation node and Whether they belong to the same dynamic community, if they are the same dynamic community, then ;otherwise , S2.2 Based on threat level Select the leader of the UAV swarm.

[0039] After forming cooperation rules in the task group, the unmanned boat cluster of each dynamic community is divided into two groups based on the threat level. Select the leader in the unmanned boat swarm, threat level The calculation is shown in formula (4): (4), in The speed of the obstacle can be detected by the speed sensor and reflects the severity of the change in the obstacle's motion state. Indicates the Current location of unmanned boats To all obstacles The minimum Euclidean distance is used to measure the proximity between the unmanned boat and the obstacle; is a very small positive number, , to avoid the denominator in formula (4) being zero.

[0040] S2.3 Adjust the follower path in the repulsive potential field of the unmanned boat cluster through the potential field function.

[0041] In this embodiment, the follower path in the repulsive potential field of the unmanned boat cluster is adjusted by the potential field function to avoid collisions between followers in the group. This is also achieved by formula (5), except that both are followers.

[0042] The S3 leader plans the path based on the improved velocity obstacle model (IVO-DWA), and the follower corrects the trajectory through the repulsive potential field, collaboratively completing multi-target real-time obstacle avoidance, which also includes: steps S3.1-S3.3, S3.1 Define the multi-objective functions of path length, smoothness, and safety in the improved speed barrier model (IVO-DWA).

[0043] In the improved speed barrier model (IVO-DWA), a multi-objective function of path length, smoothness and safety is defined. The calculation of the multi-objective function is shown in formula (7): (7), in They represent the heading angle and speed of the unmanned boat, which are the control variables that need to be optimized; Indicates the total number of steps in the planning time range, limiting the duration of the optimization process; in the summation term of the multi-objective function, it corresponds to different optimization goals: represents the preset path length weight, Indicates that the unmanned boat is at time The position vector of represents the target point position vector, In the form of vector norm, it represents the time of the unmanned boat Location To the destination The Euclidean distance in the n-dimensional space is used to minimize the navigation distance; the distance between the current position and the target point position can be obtained by the distance detection sensor.

[0044] Represents the preset path smoothness weight, Indicates the heading angle at time The value of Indicates the change amplitude of the heading angle in adjacent time steps, which is used to suppress sudden changes in heading; Indicates the preset security weight, Indicates the obstacle at time The position vector of Indicates the safety obstacle avoidance distance threshold. The obstacle position is lower than the safety threshold. When the penalty is imposed, the obstacle avoidance safety distance is forced to be met. Indicates the minimum distance between the unmanned boat and the obstacle. In this embodiment, ,The position vector of the obstacle is also obtained through the obstacle detection sensor.

[0045] In this way, the efficiency of unmanned boat path optimization can be improved in path planning through the weighted summation of multi-objective functions, and the stability can be improved through the smooth steering of the unmanned boat. At the same time, it can comply with dynamic real-time obstacle avoidance to ensure safety, thereby achieving dynamic balance between multiple objectives and optimizing paths for real-time obstacle avoidance. The leader optimizes the multi-objective path through the multi-objective function, driving the followers to achieve real-time, safe and compliant cluster obstacle avoidance.

[0046] S3.2 determines the final heading angle of the follower in the unmanned boat cluster based on the current heading angle of the leader in the unmanned boat cluster, as shown in formula (6): (6), in Indicates the final heading angle of the follower unmanned vehicle, which is used to adjust its navigation direction to achieve collaborative obstacle avoidance; Indicates the current heading angle of the leader unmanned boat, which serves as the reference for path planning; represents the force function, that is, the repulsive force vector of the follower unmanned boat; the current heading angle is obtained by formula (7), Represents the time of the control cycle, which is used to convert the instantaneous effect of the repulsive force vector into an incremental adjustment of the heading angle.

[0047] In this embodiment, The follower path in the repulsive potential field of the unmanned boat cluster is adjusted by the potential field function. The potential field function is calculated as shown in formula (5): (5), in Represents the repulsive force vector of the follower unmanned boat, which is used to correct its path to avoid collision with other unmanned boats in the same group, reflecting the task group The obstacle avoidance correction effect of other unmanned boats on the current follower; is the repulsion coefficient, which is used to control the intensity of the repulsive force; is the Gaussian decay function, Unmanned boat and Euclidean distance to the current location, is the attenuation coefficient, which characterizes the rate at which the repulsive force weakens with increasing distance; From an unmanned boat Pointing to unmanned boats The unit direction vector of .

[0048] The present invention is compared with the existing technology (VO / MPC / DRL), and the key technical indicators are shown in Table 1 below. Table 1 Comparison of technical indicators

[0049] It can be concluded that the present invention has significantly improved the perception accuracy detection, obstacle avoidance distance detection and COLREGs compliance of the unmanned boat.

[0050] In this embodiment, the core module contribution decomposition verification result is as follows: Figure 4 As shown in Figure 3, by systematically disabling core modules, the key role of each module in obstacle avoidance performance is verified.

[0051] In experimental group G1, multimodal dynamic perception fusion (CDPF) is disabled, as shown in Figure 4 As shown in (a), the minimum obstacle avoidance distance (MAD) of the unmanned boat is reduced by 22.9% compared with the present invention (CoDAC) (32.1m vs 53.6m, p=0.003). Figure 4 As can be seen in (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 of the bar, as shown in Figure 2. Figure 4 The maximum value of the present invention (CoDAC) in (d) is 2.5px, and the maximum value of the 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 positioning.

[0052] Disable the dynamic community discovery mechanism (DCDM) in the experimental group G2, such as Figure 4 As shown in (c), the task group division consistency (TGCI) of the unmanned boat decreased by 20.9% (0.91→0.72), confirming that the dynamic community discovery mechanism (DCDM) updates the edge weights through incremental modularity and then updates the disturbed edge set of the dynamic community, thereby optimizing the dynamic task group division and reducing communication complexity.

[0053] In experimental group G3, the traditional speed barrier method was used, e.g. Figure 4 In (e), the UAV’s rule compliance score (RCS) dropped 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 compliance with maritime regulations.

[0054] In experimental group G4, both the multimodal dynamic perception fusion (CDPF) and the dynamic community discovery mechanism (DCDM) were disabled, and the performance degradation of the unmanned boat was nonlinearly superimposed, e.g. Figure 4 In (b), the collision rate (CR) of the experimental group G4 increased by 12 times compared with the present invention (0.3%→3.6%), indicating that multiple modules collaborate to form a closed-loop optimization mechanism of perception-decision-collaboration.

[0055] The above-mentioned significance analysis (ANOVA, p<0.01) further verified the overall contribution of the core modules: dynamic community discovery mechanism (DCDM) (TGCI increased by 26.4%), improved speed obstacle model (IVO-DWA) (RCS increased by 30.6%), and multimodal dynamic perception fusion (CDPF) (MAD increased by 22.9%), providing multi-dimensional technical support for the collaboration of unmanned boat clusters in dynamic ocean environments.

[0056] like Figure 5As shown, the stability evolution of the unmanned vehicle cluster under dynamic obstacles. In this example, N = 100 unmanned vehicles, showing the evolution process of the community stability of a 100-unit unmanned vehicle (USV) cluster in a dynamic obstacle environment.

[0057] The present invention is based on the dynamic community discovery mechanism (DCDM) of incremental modularity optimization (such as Figure 5 The solid line in the middle) and the fixed grouping strategy of the experimental group G2 (as shown in Figure 5 The community stability index (CSI) is quantified by the Jaccard similarity, reflecting the degree of overlap of community members at adjacent times. In the no-mutation stage (t=0-19), the average CSI of the dynamic community discovery mechanism (DCDM) is equivalent to 0.85±0.05 in the experimental group G2, reaching 0.92±0.03, indicating that the dynamic mechanism can maintain higher consistency in the steady state, and the dynamically adjusted community structure significantly improves the robustness of the system.

[0058] However, when encountering a sudden change in speed (Δv=5m / s², t=20 / 35 steps), the CSI of the dynamic community discovery mechanism (CoDAC) only dropped to 0.87±0.04, while it dropped sharply to 0.58±0.07 in the experimental group G2, confirming that dynamic reorganization strategies (such as local modularity gain calculation) can alleviate the impact of sudden disturbances.

[0059] The core advantages of the Dynamic Community Discovery Mechanism (DCDM) are reflected in two aspects: first, incremental update efficiency. By processing only the affected perturbed edge set (ΔE), the number of reorganizations is reduced by 58.3% compared to the experimental group G2 (5 vs. 12 times), and the computational complexity is reduced from O(N²) to O(|ΔE|logN), meeting real-time requirements. Second, structural adaptability. By balancing community size and communication load through a resolution parameter γ = 1.2, stability is quickly restored after mutations (the CSI returns to 0.88±0.03 between steps t = 35 and 49), avoiding the oscillation and attenuation of fixed groups.

[0060] Error band analysis further revealed differences in robustness. 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 dynamic mechanism's tolerance to noise. This result is consistent with the "diversity-stability" theory in ecology: the modular network architecture maintains global stability by enhancing negative intraspecific interactions (such as potential field correction), thereby providing a theoretical basis for the engineering deployment of large-scale clusters. It has been applied in port patrol scenarios (communication load reduced by 68%).

[0061] like Figure 6As shown in the figure, the robustness of sensor noise and large-scale clustering is verified. The measured parameters of Velodyne VLP-16 are used to build a hybrid noise model (LiDAR ranging noise σ = 0.3m, camera randomly occludes 40%-60% of pixels), and compared with DRL-CA and DynaMOCO methods for verification.

[0062] like Figure 6 As shown in (a), the experimental results show that: 1) In terms of cross-modal complementarity, multimodal dynamic perception fusion (CDPF) can effectively suppress unimodal noise interference through a cross-modal attention mechanism. The cross-modal feature alignment error (FE) under LiDAR noise is significantly better than the error increase of DRL-CA. 2) In terms of dynamic feature restoration capabilities, the spatial-channel dual attention (SCA) of the present invention (CoDAC (CR)) maintains a collision rate CR < 1.5% even at a 50% occlusion rate, a 57% reduction compared to DRL-CA (3.6% → 1.5%, p = 0.007); 3) Noise sensitivity analysis shows that the minimum obstacle avoidance distance (MAD) in a mixed noise environment is strongly linearly correlated with the noise intensity (R²=0.93). However, the decline slope k=0.12m / σ of the proposed method (such as CoDAC (MAD) in the figure) is only 48% of that of DRL-CA (0.25m / σ), demonstrating that the dynamic community discovery mechanism (DCDM) has stronger noise robustness and verifies the adaptability of the cross-modal dynamic perception architecture in complex environments.

[0063] like Figure 6 As shown in (b), large-scale cluster test.

[0064] In large-scale cluster testing, this embodiment constructed a simulation scenario containing 10-100 heterogeneous USVs (unmanned submersible vehicle clusters) (with propeller power differences of ±30%), introduced interference conditions of 30% communication packet loss and ±20Mbps bandwidth fluctuation, and compared and analyzed the centralized MPC (based on model predictive control) and distributed PSO (based on particle swarm optimization). The test results showed that: 1) In terms of distributed scalability, the Dynamic Community Discovery Mechanism (DCDM) achieves sublinear growth in planning latency through incremental modularity optimization. When the number of nodes increases from 50 to 100 ships, the latency slope k = 1.8ms / ship (R² = 0.98), a 65% reduction compared to the linear growth of MPC (k = 5.2ms / ship); 2) In terms of anti-interference capability, in a 30% packet loss scenario, the leader election mechanism (Equation 9) maintains task group consistency TGCI > 0.75, while the communication load GCL increases by only 12% (25% → 28%), significantly outperforming the 45% increase in PSO. 3) In terms of resource efficiency, the memory usage is optimized to 48MB at the scale of 100 ships, 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.

[0065] like Figure 7 As shown in the figure, a path planning performance comparison is conducted (multi-ship intersection encounter scenario, unmanned boat N=20), showing the path planning performance comparison of different obstacle avoidance algorithms in the multi-ship intersection encounter scenario.

[0066] Experimental data show that the CoDAC method of the present invention is significantly superior to traditional methods in terms of safety, efficiency, and compliance. Specifically: like Figure 7 As shown in (a), the minimum obstacle avoidance distance (MAD) of the present invention (CoDAC) reaches 53.6 meters, which is 63.9% higher than the traditional speed obstacle method (VO). Figure 7 (a) By improving the triangular obstacle zone (TOZ) model, the vertex position of dynamic obstacles is predicted, thus expanding the safe avoidance boundary.

[0067] like Figure 7 As shown in (b), the collision rate (CR) of the present invention (CoDAC) is only 0.3%, which is 66.7% lower than that of the second best method DynaMOCO ( Figure 7 Middle (b), task group division attributed to the dynamic community discovery mechanism (DCDM), reducing the probability of conflict within the task group through modularity optimization.

[0068] like Figure 7 As shown in (c), the path length ratio (PLR) of the present invention (CoDAC) is 1.12 ( Figure 7 (c) is superior to all the compared methods, indicating that multimodal perception fusion improves obstacle localization accuracy (AP@0.5IoU=89.7%) and reduces path redundancy.

[0069] like Figure 7 As shown in (d), the present invention (CoDAC) calculates the latency index ( Figure 7 In (d), CoDAC’s computational latency is only 95ms, which is 70.3% lower than that of centralized MPC. This is because its distributed architecture only needs to process the perturbed edge set (ΔE), reducing the complexity from O(N²) to O(|ΔE|logN).

[0070] like Figure 7 As shown in (e), COLREGs compliance aspects ( Figure 7In figure e), CoDAC achieves zero violations, while traditional VO, due to the lack of rule encoding, results in a violation density of 4.7 times / m². This is due 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 within the speed window constraint.

[0071] In summary, CoDAC provides a highly secure, efficient, and strictly compliant obstacle avoidance solution for large-scale USV clusters through a perception-planning-collaborative closed-loop optimization mechanism.

[0072] like Figure 8 As shown in the figure, a comparative analysis of obstacle avoidance trajectories in a multi-ship crossing scenario is performed, showing the comparative analysis of obstacle avoidance trajectories of three different ship collision avoidance technologies (the present invention (CoDAC), VO and DRL-CA) in a multi-ship crossing scenario. Figure 8 The three methods of the present invention (CoDAC), VO, and DRL-CA were evaluated from different dimensions: from Figure 8 As can be seen in Figure a, the blue solid line 1 of the present invention (CoDAC) is relatively smooth and can follow the rule well when approaching the "COLREGsRule 15 Boundary" (located at the longitudinal position of 150m). The red dotted line 2 of the VO method fluctuates greatly. In the longitudinal position range of 100-200m, the lateral position changes dramatically, deviating from the expected rule. The green dotted line 3 of the DRL-CA method has obvious lateral fluctuations near the longitudinal position of 150m, and the trajectory stability is not as good as CoDAC.

[0073] from Figure 8 In figure b, the blue solid line 1 of the present invention (CoDAC) consistently maintains a minimum distance of approximately 53.6 meters, staying away from the conflict zone below 50 meters. The red dashed line 2 of the VO method fluctuates between 20 and 35 meters between 0 and 60 seconds, entering the conflict zone multiple times. The green dotted line 3 of the DRL-CA method maintains a stable minimum distance of 100 meters, demonstrating high safety.

[0074] from Figure 8 In Figure c, the dynamic safety margin (light blue filled area) around the trajectory of the blue solid line 1 of the present invention (CoDAC) is reasonable, 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 narrow, and the trajectory fluctuates greatly. The safety margin (light green filled area) of the green dotted line 3 of the DRL-CA method ranges from -15m to 15m within the longitudinal position of 0-300m, which can better ensure safety.

[0075] The working principle of the present invention is as follows: First, LiDAR point cloud, camera image, radar data, and AIS data are synchronized through the IEEE 1588 protocol, and cross-modal spatial alignment is achieved based on the extrinsic parameter matrix; second, the cross-modal attention mechanism and deformable convolution are combined to establish the correlation between channels and spatial dimensions, so that the fused multi-modal features can generate high-precision environmental semantic information; then, the unmanned boat task group is 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 unmanned boat leader adopts the improved velocity obstacle model (IVO-DWA) to optimize the multi-target path, and the follower corrects the trajectory through the repulsive potential field to achieve real-time, safe and compliant cluster obstacle avoidance.

Claims

1. A multi-modal dynamic cooperative unmanned watercraft swarm autonomous obstacle avoidance method, wherein the unmanned watercraft swarm includes a leader and followers, and further includes the following steps: S0 acquires multimodal sensor data and performs data preprocessing; In multimodal dynamic perception fusion, S1 fuses LiDAR bird's-eye view and camera perspective features through cross-modal attention mechanism and deformable convolution to generate a time-varying graph of the fused multimodal features; S2 updates the disturbed edge set of the dynamic community of the unmanned boat cluster through incremental community discovery on the time-varying graph of the fused multimodal features, optimizes the division of task groups, and selects leaders based on the threat level; The S3 leader plans multi-objective paths based on the improved speed obstacle model by defining the multi-objective functions of path length, smoothness and safety, selects the optimal path from the multi-objective paths according to the leader and follower paths, and avoids obstacles according to the optimal path.

2. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 1, characterized in that: Step S0 includes: Multimodal sensor data is synchronized using the IEEE 1588 protocol. Under the condition of temporal alignment, the LiDAR-camera extrinsic parameter matrix is ​​calculated using the checkerboard calibration method for spatial alignment.

3. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 1, characterized in that: In step S0 , the multimodal sensor data includes LiDAR point cloud, camera image, and radar data.

4. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 1, characterized in that: Step S1 includes: steps S1.1 to S1.2, S1.1 preset cross-modal attention weight matrix is ; S1.2 uses the cross-modal attention weight matrix as , deformable convolution performs geometric feature fusion on the LiDAR bird's-eye view and the camera perspective image to form a fused multimodal feature .

5. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 4, characterized in that: In step S1.1, the cross-modal attention weight matrix Indicates modality Modal The attention weight of reflects the cross-modal feature correlation, as shown in formula (1): ， in Indicates the The original features of the modes, Indicates the The original features of the modes, Represents a linear mapping function and generates queries and keys respectively, represents the feature dimension, Represents spatial dimensions; In step S1.2, the fused multimodal features , as shown in formula (2): （2）， Indicates the number of feature sources involved in fusion, Represents a linear mapping function and generates a vector of values.

6. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 1, characterized in that: The step S2 includes steps S2.1-S2.

3. Step S2.1 dynamically updates the edge weight based on the incremental modularity of the time-varying graph, and then updates the disturbed edge set of the dynamic community of the unmanned boat cluster. , as shown in formula (3): （3）， in Representation node and In time The edge weight is updated by the position of the unmanned boat and the similarity of the task. Representation node The degree, Representation node The degree, represents the total number of edges in the timing graph, Represents the preset modularity resolution parameter, which is used to control the dynamic community granularity. Representation node and Whether they belong to the same dynamic community, if they are the same dynamic community, then , S2.2 Based on threat level Selecting a leader in the UAV swarm; S2.3 Adjust the follower path in the repulsive potential field of the unmanned boat cluster through the potential field function.

7. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 6, characterized in that: In step S2.2, the threat level The calculation is shown in formula (4): （4）， in is the detected obstacle speed change, is the module of the obstacle’s acceleration, reflecting the severity of the change in the obstacle’s motion state; Indicates the Current location of unmanned boats To all obstacles The minimum Euclidean distance is used to measure the proximity between the unmanned boat and the obstacle; is a very small positive number, .

8. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 1, characterized in that: Said step S3 also includes: steps S3.1-S3.2, S3.1 Define the multi-objective function of path length, smoothness, and safety in the improved speed barrier model (IVO-DWA); S3.2 determines the final heading angle of the follower in the unmanned boat cluster based on the current heading angle of the leader in the unmanned boat cluster, as shown in formula (6): （6）， in Indicates the final heading angle of the follower unmanned vehicle, which is used to adjust its navigation direction to achieve collaborative obstacle avoidance; Indicates the current heading angle of the leader unmanned boat, which serves as the reference for path planning; represents the potential field function, Represents the time of the control cycle, which is used to convert the instantaneous effect of the repulsive force vector into an incremental adjustment of the heading angle.

9. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 6, characterized in that: In step S2.3, the potential field function is calculated as shown in formula (5): （5）， in Represents the repulsive force vector of the follower unmanned boat, which is used to correct its path to avoid collision with other unmanned boats in the same group, reflecting the task group The obstacle avoidance correction effect of other unmanned boats on the current follower; is the repulsion coefficient, which is used to control the intensity of the repulsive force; is the Gaussian decay function, Unmanned boat Euclidean distance to the current location, is the attenuation coefficient, which characterizes the rate at which the repulsive force weakens with increasing distance; From an unmanned boat Pointing to unmanned boat The unit direction vector of .

10. The multi-modal dynamic cooperative unmanned boat swarm autonomous obstacle avoidance method according to claim 8, characterized in that: In step S3.1, the multi-objective function is calculated as shown in formula (7): （7）， in They represent the heading angle and speed of the unmanned boat, which are the control variables that need to be optimized; Indicates the total number of steps in the planning time range, limiting the duration of the optimization process; The summation term of the multi-objective function corresponds to different optimization objectives: represents the path length weight, Indicates that the unmanned boat is at time The position vector of represents the target point position vector, Indicates that the unmanned boat is at time Location To the destination The Euclidean distance is used to minimize the navigation distance; represents the path smoothness weight, Indicates the heading angle at time The value of Indicates the change amplitude of the heading angle in adjacent time steps, which is used to suppress sudden changes in heading; represents the security weight, Indicates the obstacle at time The position vector of Indicates the safety obstacle avoidance distance threshold. The obstacle position is lower than the safety threshold. When the penalty is imposed, the obstacle avoidance safety distance is forced to be met. Indicates the minimum distance between the unmanned boat and obstacles.

Images