Multi-agent cooperative obstacle avoidance method and system for unmanned ship in complex sea conditions

Abstract

The application discloses a complex sea state-oriented unmanned ship multi-agent cooperative obstacle avoidance method and system, relates to the unmanned ship multi-agent cooperative obstacle avoidance field, and comprises the following steps: acquiring real-time data of each unmanned ship and the surrounding environment; generating a candidate obstacle target point cloud cluster list based on a density clustering algorithm; acquiring real-time absolute motion state estimation of the candidate obstacle based on a Kalman filter, and outputting an obstacle feature list; determining a safety radius compensation required for autonomous obstacle avoidance of each unmanned ship; acquiring a safe navigable space of each unmanned ship at the current moment; constructing a global synthetic potential field, and generating a set of optimal candidate paths of the unmanned ship from the current position to a target point; and acquiring a consensus-reached unmanned ship multi-agent cooperative obstacle avoidance path based on a consensus binding algorithm. The application has the advantages that the multi-unmanned ship cluster can realize safe, efficient and cooperative intelligent obstacle avoidance in a real complex marine environment.

Description

TECHNICAL FIELD

[0001] The application relates to the field of unmanned ship multi-agent cooperative obstacle avoidance, in particular to an unmanned ship multi-agent cooperative obstacle avoidance method and system for complex sea conditions. BACKGROUND

[0002] The unmanned ship multi-agent system has become a frontier research direction in the marine field due to its high efficiency, high redundancy and wide coverage range in performing tasks such as regional mapping, ocean monitoring, search and rescue patrol. An ideal cooperative system requires the fleet not only to complete the given formation sailing, but also to make timely, safe and cooperative obstacle avoidance decisions in the dynamic and complex real marine environment, facing static obstacles (such as islands, reefs, anchored ships) and dynamic obstacles (such as passing ships, drifts). At present, unmanned ship cluster cooperative obstacle avoidance mainly realizes environment and obstacle perception through multi-sensor fusion, generates an obstacle avoidance path based on a path planning algorithm, and eliminates inter-ship conflicts by means of a cooperative decision mechanism. Among them, the multi-sensor (inertial measurement unit, laser radar, millimeter wave radar, wave sensor, etc.) is the core of the perception layer, which needs to synchronously collect ship posture, obstacle, sea condition and other data; the path planning needs to balance the safety and efficiency of obstacle avoidance and adapt to the physical characteristics of the ship (minimum turning radius, maximum turning rate); the cooperative decision needs to avoid single-point failure of central control in the limited communication marine environment, and realize distributed consensus. With the expansion of the scale of the unmanned ship cluster and the complexity of the operation scene, the requirements for sea condition adaptability, obstacle avoidance real-time performance and cooperative robustness are significantly improved. The traditional technology needs to be further optimized to adapt to the actual needs in complex sea conditions, and to provide technical support for safe and efficient operation of the unmanned ship cluster.

[0003] The traditional path planning algorithm (such as the standard artificial potential field method) assumes that the ship has ideal maneuverability in still water, completely ignores the key physical facts such as the increase of ship turning inertia, the decline of ship speed maintaining ability and the extension of effective braking distance caused by the combined action of wind, wave and current, which leads to the fact that the planned path may be too aggressive or contain maneuvering actions beyond the dynamic capability of the ship, resulting in collision of the ship in actual execution due to inability to accurately track the trajectory (such as turning not enough or drifting too much). Secondly, the conflict resolution mechanism of the existing multi-agent cooperative algorithm (such as the early CBBA) is usually based on static geometric paths or time windows, while ignoring the fact that in complex sea conditions, the environmental load suffered by unmanned ships at different positions and different headings in the cluster is heterogeneous and time-varying, which means that the dynamic safety boundary and maneuvering capability of each ship are real-time different. Lack of unified quantification and characterization of such dynamic constraints will lead to the fact that the so-called "collision-free" path combination generated by the cooperative decision may force some ships to perform high-risk maneuvers in severe conditions where their maneuvering ability is severely limited, thereby burying safety hazards from the system level and damaging the overall safety and efficiency of the cluster cooperation. SUMMARY

[0004] To address the aforementioned technical problems, this paper provides a multi-agent cooperative obstacle avoidance method and system for unmanned vessels in complex sea conditions. This technical solution solves the problems mentioned in the background art, which lead to the planned paths being too aggressive or containing maneuvers that exceed the dynamic capabilities of the vessel, causing collisions in actual execution due to the inability to accurately track the trajectory. It also addresses the lack of unified quantification and representation of such dynamic constraints, which may result in the so-called "conflict-free" path combinations generated by cooperative decision-making, potentially forcing some vessels to perform high-risk maneuvers under adverse conditions where their maneuverability is severely limited.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions includes:

[0007] Real-time data of each unmanned vessel and its surrounding environment are acquired through a multi-sensor fusion system mounted on each unmanned vessel.

[0008] Based on the surrounding obstacle point cloud data, a density clustering algorithm is used to filter candidate obstacle target point cloud clusters and generate a list of candidate obstacle target point cloud clusters.

[0009] Based on the selected candidate obstacle target point cloud clusters, the real-time absolute motion state estimation of the candidate obstacles is obtained using a Kalman filter, and a list of obstacle features is output.

[0010] Based on real-time data of each unmanned vessel and its surrounding environment, determine the safety radius compensation required for each unmanned vessel to autonomously avoid obstacles.

[0011] Based on the list of obstacles and the safe radius required for each unmanned vessel to autonomously avoid obstacles, the safe navigation space of each unmanned vessel at the current moment is obtained.

[0012] Each unmanned surface vessel (USV) constructs a global synthetic potential field based on its own safe navigation space, generating a set of optimal alternative paths from its current position to the target point.

[0013] Based on the alternative paths for each unmanned surface vessel (USV) from its current location to the target point, a consensus-based binding algorithm is used to obtain a multi-agent cooperative obstacle avoidance path for USVs.

[0014] Preferably, the step of filtering candidate obstacle target point cloud clusters based on surrounding obstacle point cloud data and using a density clustering algorithm, and generating a list of candidate obstacle target point cloud clusters, specifically includes:

[0015] Simultaneously acquire 3D spatial point cloud data provided by lidar and target reflection point cloud data provided by millimeter-wave radar;

[0016] The three-dimensional space point cloud data and the target reflection point cloud data are converted to a body coordinate system centered on the unmanned ship through coordinate transformation;

[0017] In the unified coordinate system, data association is performed on the laser radar point cloud and the millimeter wave radar point cloud to generate a preliminary fusion point cloud covering space and reflection characteristics;

[0018] The original image collected by the visible light / infrared vision sensor is input into a pre-trained semantic segmentation neural network to assign a semantic label to each pixel and generate a semantic label image;

[0019] The fusion point cloud data is input into a density clustering algorithm, which identifies a set of point clouds that are closely distributed in space and have a density reaching a preset threshold a1 as an independent point cloud cluster;

[0020] All spatially isolated points that cannot be classified into any point cloud cluster are determined as sensor noise or transient clutter and are filtered out;

[0021] The preliminary formed point cloud cluster is projected onto the semantic label image, and the semantic label distribution of the image area covered by each point cloud cluster is counted. If the proportion of the sea wave label is not less than a preset threshold a2, the cluster is determined as a sea wave false target, and it is filtered out from the candidate target list;

[0022] For the remaining point cloud clusters, the size of the three-dimensional bounding box is calculated, and clusters with a size exceeding a preset physical range [a3, a4] are filtered out, and a list of candidate obstacle target point cloud clusters selected is output;

[0023] The geometric center coordinates in the list of selected candidate obstacle target point cloud clusters are converted from the body coordinate system centered on the unmanned ship to the geodetic coordinate system.

[0024] Preferably, the real-time absolute motion state estimation of the candidate obstacle is obtained based on the selected candidate obstacle target point cloud cluster based on a Kalman filter, and an obstacle feature list is output, which specifically includes:

[0025] A Kalman filter is initialized for each newly identified candidate obstacle, wherein the state vector of the Kalman filter includes the geodetic coordinates, velocity, heading angle and turning rate of the candidate obstacle point cloud cluster;

[0026] Based on the uniform turning motion model, a system state transition equation of the Kalman filter is constructed, and the geodetic coordinates of the current frame candidate obstacle point cloud cluster are taken as the observation vector;

[0027] Based on the Kalman filter, the updated state vector is output as the real-time and optimal absolute motion state estimation of the corresponding obstacle;

[0028] a speed according to the absolute motion state estimation output by the Kalman filter;

[0029] setting a motion determination threshold and a steering determination threshold, combining the speed value and the steering rate, judging the state of the obstacle, the state of the obstacle indicating static or motion;

[0030] According to the absolute motion state estimation of the obstacle, the relative distance, the relative speed and the relative heading of the obstacle relative to the ship are obtained through vector operation based on the earth coordinates and the heading angle of the unmanned ship;

[0031] According to the absolute motion state estimation of the obstacle, the relative distance, the relative speed and the relative heading of the obstacle relative to the ship, the state of the obstacle and the corresponding semantic label, the obstacle feature list is constructed;

[0032] The judgment of the state of the obstacle specifically includes:

[0033] If the absolute speed value is not lower than the motion determination threshold for N1 frames, or the absolute value of the absolute steering rate is not lower than the steering determination threshold for N1 frames, it is determined as a moving obstacle, otherwise, it is determined as a static obstacle;

[0034] When it is determined as a static obstacle, if its semantic label belongs to a predefined category set with kinetic energy, it is marked as a static obstacle with kinetic energy.

[0035] Preferably, the safety radius compensation required for each unmanned ship to autonomously avoid obstacles is determined according to real-time data of each unmanned ship and the surrounding environment, specifically including:

[0036] According to the wave direction spectrum and wave height data collected by the unmanned ship, a sea state interference factor representing the severity of sea state is calculated;

[0037] Based on the sea state interference factor and the wind flow vector data, the effective maneuverability degradation coefficient of the unmanned ship under the combined action of wind, wave and current is calculated through a ship motion response model;

[0038] Based on the effective maneuverability degradation coefficient, the safety radius compensation required for the unmanned ship to autonomously avoid obstacles is calculated.

[0039] Preferably, the safety navigable space of each unmanned ship at the current time is obtained according to the obstacle feature list and the safety radius required for each unmanned ship to autonomously avoid obstacles, specifically including:

[0040] According to the length, width and height of the bounding box of all static obstacles in the list, the distance of the outward expansion of the safety radius is generated to generate a static prohibited area;

[0041] For each moving obstacle in the obstacle feature list, based on its real-time absolute motion state estimation output by its Kalman filter, substitute it into its uniform turning motion model, conduct forward numerical simulation, predict its position at a series of discrete time points in the future, and thus generate the predicted trajectory of the obstacle;

[0042] According to the predicted trajectory of the obstacle, combined with the length, width and height of the obstacle bounding box, expand a safety radius on both sides of the trajectory center line to generate a dynamic threat corridor that changes over time;

[0043] From the global environment, eliminate static prohibited areas and dynamic threat corridors to obtain the safe navigable space of each unmanned ship at the current time.

[0044] Preferably, each unmanned ship constructs a global synthetic potential field based on its own safe navigable space to generate a set of optimal candidate paths for the unmanned ship from the current position to the target point, which specifically includes:

[0045] Taking the current position of each unmanned ship as the origin, a local plane coordinate system is established based on the geodetic coordinate system, the safe navigable space is converted to the local coordinate system to form a local search space for path search;

[0046] In the local plane coordinate system, the continuous state space of the unmanned ship is discretized into a four-dimensional space-time state grid, each node of the state grid represents a specific state, and the specific state includes the position of the unmanned ship, the heading angle of the unmanned ship, and the time from the start time of path generation;

[0047] Based on the physical characteristics of the unmanned ship and expert experience, node parameters are set to define a search granularity for the specific state in the space-time state grid that matches the physical characteristics of the unmanned ship;

[0048] In the local plane coordinate system, a global synthetic potential field is constructed, which is composed of a target attractive field and an obstacle repulsive field;

[0049] The target attractive field refers to the potential energy value at any node, which is calculated by the Euclidean distance from the node to the target node;

[0050] The obstacle repulsive field refers to the potential energy value at any node, which is calculated by the Euclidean distance from the node to the nearest obstacle;

[0051] The global synthetic potential field is obtained by linearly superimposing the target attractive field and the obstacle repulsive field;

[0052] Starting from the current state node of the unmanned ship, the negative gradient direction of the node position in the synthetic potential field is calculated;

[0053] sampling in the preset offset interval according to the heading angle resolution, to generate a candidate heading angle;

[0054] Based on the candidate heading angle, the current speed and the time step are combined to calculate the position coordinates of the subsequent node through the ship kinematics model, to generate the corresponding subsequent state node, thereby forming a candidate polyline path node sequence;

[0055] Immediately verify whether the line segment formed by the new node and the previous node has an intersection with the obstacle region in space-time trajectory, and only keep the node without collision;

[0056] Smooth each polyline path in the set of collision-free candidate polyline paths to obtain a set of smoothed candidate paths;

[0057] Calculate the total geometric length of each path in the set of smoothed candidate paths, and select the candidate path with the smallest total geometric length as the optimal candidate path of the unmanned ship from the current position to the target point.

[0058] Preferably, the consensus binding algorithm is used to obtain a consensus unmanned ship multi-agent collaborative obstacle avoidance path based on the candidate paths of each unmanned ship from the current position to the target point, which specifically includes:

[0059] Each unmanned ship encapsulates the optimal candidate path it generates and the total geometric length of the path into a data packet, and broadcasts it to all other unmanned ships in the cluster through wireless ad hoc networking;

[0060] After receiving the data packets from other ships, each unmanned ship independently builds a global collaborative decision matrix in its local processing system;

[0061] The row dimension of the collaborative decision matrix corresponds to the number of each unmanned ship in the cluster, and the column dimension corresponds to the optimal candidate path of each ship and the total geometric length of the path.

[0062] A consensus binding algorithm is used to perform distributed path combination optimization on the unmanned ship cluster, and output a consensus unmanned ship multi-agent collaborative obstacle avoidance path;

[0063] Each unmanned ship receives and executes the collaborative path assigned to it in the final path combination, and continues to travel to the final target point according to the preset formation after completing the obstacle avoidance.

[0064] Further, the present scheme proposes an unmanned ship multi-agent collaborative obstacle avoidance system for complex sea conditions, which is used to implement the unmanned ship multi-agent collaborative obstacle avoidance method for complex sea conditions as described above, comprising:

[0065] A data acquisition module is configured to acquire real-time data of each unmanned ship and the surrounding environment through a multi-sensor fusion system carried on each unmanned ship.

[0066] A navigable space module is configured to filter candidate obstacle target point cloud clusters based on a density clustering algorithm according to the peripheral obstacle point cloud data, and generate a candidate obstacle target point cloud cluster list; acquire real-time absolute motion state estimation of the candidate obstacle based on a Kalman filter according to the filtered candidate obstacle target point cloud clusters, and output an obstacle feature list; determine a safety radius compensation required for autonomous obstacle avoidance of each unmanned ship according to the real-time data of each unmanned ship and the surrounding environment; and acquire the safety navigable space of each unmanned ship at the current time according to the obstacle feature list in combination with the safety radius required for autonomous obstacle avoidance of each unmanned ship.

[0067] A cooperative obstacle avoidance module is configured to construct a global synthetic potential field based on the safety navigable space of each unmanned ship, and generate a group of optimal candidate paths of the unmanned ships from the current position to the target point; and acquire a consensus reached unmanned ship multi-agent cooperative obstacle avoidance path based on a consensus binding algorithm according to the candidate paths of each unmanned ship from the current position to the target point.

[0068] The navigable space module specifically comprises:

[0069] A point cloud cluster unit is configured to filter candidate obstacle target point cloud clusters based on a density clustering algorithm according to the peripheral obstacle point cloud data, and generate a candidate obstacle target point cloud cluster list.

[0070] An obstacle feature unit is configured to acquire real-time absolute motion state estimation of the candidate obstacle based on a Kalman filter according to the filtered candidate obstacle target point cloud clusters, and output an obstacle feature list.

[0071] A safety radius unit is configured to determine a safety radius compensation required for autonomous obstacle avoidance of each unmanned ship according to the real-time data of each unmanned ship and the surrounding environment.

[0072] A navigable space unit is configured to acquire the safety navigable space of each unmanned ship at the current time according to the obstacle feature list in combination with the safety radius required for autonomous obstacle avoidance of each unmanned ship.

[0073] The cooperative obstacle avoidance module specifically comprises:

[0074] An optimal path unit is configured to construct a global synthetic potential field based on the safety navigable space of each unmanned ship, and generate a group of optimal candidate paths of the unmanned ships from the current position to the target point.

[0075] The cooperative obstacle avoidance unit is used to obtain a consensus unmanned ship multi-agent cooperative obstacle avoidance path based on a consensus binding algorithm according to the candidate paths of each unmanned ship from the current position to the target point.

[0076] Compared with the prior art, the present application has the beneficial effects that:

[0077] The present application proposes a complex sea state oriented unmanned ship multi-agent cooperative obstacle avoidance method and system. The present application synchronously collects the data such as the unmanned ship pose, obstacle point cloud, sea wave direction spectrum, wind flow vector and the like through a multi-sensor fusion system, realizes the static, dynamic and kinetic static obstacle classification and motion state estimation in combination with the density clustering algorithm and Kalman filtering algorithm, quantifies the influence of the environment on the ship maneuverability based on the sea state interference factor, dynamically adjusts the safety radius compensation amount to build a precise safe navigable space, and on this basis, generates an optimal candidate path adapting to the physical characteristics of the ship through the space-time state grid discretization and global synthetic potential field, and finally realizes the distributed path combination optimization and conflict resolution relying on the consensus binding algorithm, thereby effectively solving the problems such as the obstacle misidentification caused by the sensor interference under the complex sea state, the mismatch between the fixed safety boundary and the ship maneuverability degradation, the path planning deviating from the actual sea state constraint, and the cooperative decision relying on the central control being prone to failure, thereby improving the perception robustness and obstacle avoidance safety of the unmanned ship cluster to the complex marine environment, and guaranteeing the real-time performance and global navigation efficiency of the cooperative decision. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 A complex sea state oriented unmanned ship multi-agent cooperative obstacle avoidance method flowchart of the present application;

[0079] Figure 2 A candidate obstacle target point cloud cluster is screened and a candidate obstacle target point cloud cluster list is generated based on the density clustering algorithm of the present application, and a flowchart thereof;

[0080] Figure 3 A flowchart of the present application for obtaining the safe navigable space of each unmanned ship at the current time;

[0081] Figure 4 A flowchart of the present application for constructing a global synthetic potential field and generating a set of optimal candidate paths of the unmanned ship from the current position to the target point. DETAILED DESCRIPTION

[0082] The following description is used to disclose the present application so as to enable those skilled in the art to implement the present application. The preferred embodiments in the following description are only used as examples, and other obvious variants can be thought of by those skilled in the art.

[0083] REFERENCE Figure 1As shown, a complex sea state oriented unmanned ship multi-agent collaborative obstacle avoidance method comprises:

[0084] Through the multi-sensor fusion system carried on each unmanned ship, real-time data of each unmanned ship and the surrounding environment is obtained;

[0085] According to the surrounding obstacle point cloud data, based on the density clustering algorithm, the candidate obstacle target point cloud cluster is screened, and a candidate obstacle target point cloud cluster list is generated;

[0086] According to the screened candidate obstacle target point cloud cluster, based on the Kalman filter, the real-time absolute motion state estimation of the candidate obstacle is obtained, and an obstacle feature list is output;

[0087] According to the real-time data of each unmanned ship and the surrounding environment, the safety radius compensation required for autonomous obstacle avoidance of each unmanned ship is determined;

[0088] According to the obstacle feature list, combined with the safety radius required for autonomous obstacle avoidance of each unmanned ship, the safety navigable space of each unmanned ship at the current time is obtained;

[0089] Based on the safety navigable space of each unmanned ship, a global synthetic potential field is constructed, and a group of optimal candidate paths of the unmanned ship from the current position to the target point are generated;

[0090] According to the candidate path of each unmanned ship from the current position to the target point, based on the consensus binding algorithm, a consensus reached unmanned ship multi-agent collaborative obstacle avoidance path is obtained.

[0091] It can be explained that the method constructs a progressive and closed-loop decision system from "individual perception" to "group consensus", and the core logic is to gradually convert complex physical world information into calculable and decision-making standardized models, and finally output a safe and collaborative group action plan. Specifically, through the multi-sensor fusion system, the pose, obstacle point cloud, sea wave direction spectrum, wind flow vector and other data of each unmanned ship are synchronously collected, combined with the density clustering algorithm and Kalman filtering algorithm to realize static, dynamic and kinetic static obstacle classification and motion state estimation, and based on the sea state interference factor, the influence of the environment on the maneuverability of the ship is quantified. Adjust the safety radius compensation to build an accurate safety navigable space; On this basis, through the spatiotemporal state grid discretization and global synthetic potential field generation, the optimal candidate path adapting to the physical characteristics of the ship is generated, and finally relying on the consensus binding algorithm, the distributed path combination optimization and conflict resolution are realized, so as to realize safe and efficient group obstacle avoidance in complex sea state.

[0092] The multi-sensor fusion system carried on each unmanned ship to obtain real-time data of each unmanned ship and the surrounding environment specifically includes:

[0093] The real-time data includes: the pose state of each unmanned ship, the surrounding obstacle point cloud data, the sea wave direction spectrum and wave height data, and the wind flow vector data;

[0094] The multi-sensor fusion system includes: an inertial measurement unit, a GPS positioning module, a laser radar, a millimeter wave radar, a visible light / infrared vision sensor, a wave sensor, and a wind speed and direction sensor;

[0095] The inertial measurement unit and the GPS positioning module are combined to collect the real-time pose state of the ship, including longitude, latitude, heading angle, speed, and roll / pitch / heel angle;

[0096] The laser radar and the millimeter wave radar cooperatively scan the surrounding environment to obtain the original point cloud data of the obstacles in the surrounding range, and the visible light / infrared vision sensor assists in collecting the image features of the obstacles;

[0097] The wave sensor collects the sea wave direction spectrum, effective wave height, and wave period data based on spectral analysis technology, and the wind speed and direction sensor collects the instantaneous wind speed, wind direction, and sea current flow rate and direction data;

[0098] The data collected by all sensors are processed for space-time alignment, with the GPS time stamp as the reference to unify the time dimension of each sensor data;

[0099] The Kalman filter algorithm is used to smooth the sensor data after coordinate alignment, to eliminate outliers, and through normalization processing, the filtered data is mapped to the [0, 1] interval;

[0100] Each unmanned ship shares the preprocessed real-time data to the fleet data center through a wireless communication network.

[0101] It can be explained that data acquisition is an effective data guarantee for unmanned ship to face complex sea conditions for risk perception and intelligent collaborative obstacle avoidance. Real-time data of each unmanned ship and surrounding environment is obtained through the multi-sensor fusion system carried on each unmanned ship, and the purpose is to build a complete "ship-obstacle-environment" coupled dynamic perception system. That is, the pose state obtained by the inertial measurement unit and the GPS positioning module provides a reference for cooperative control, the obstacle point cloud data collected by the laser radar, millimeter wave radar and visual sensor is the direct input for obstacle avoidance perception, the wave direction spectrum and wave height data provided by the wave sensor are the key parameters for quantifying the influence of complex sea conditions on ship maneuverability, the wind flow vector data collected by the wind speed and direction sensor is the necessary condition for quantifying the influence of environment on ship maneuverability and supporting dynamic adjustment of safety radius, and the remaining parameters output by the sensor are used as auxiliary data for system state monitoring and data recording. The systematic integration of these data provides irreplaceable data guarantee for subsequent dynamic threat cloud map construction and intelligent collaborative decision making, so that the system can not only perceive obstacles, but also understand the ship maneuvering boundary and environmental influence under the sea conditions, so as to generate a safe and reliable collaborative obstacle avoidance strategy.

[0102] Referring to Figure 2 As shown in the density-based clustering algorithm, the candidate obstacle target point cloud cluster is filtered, and a candidate obstacle target point cloud cluster list is generated. Specifically, the density-based clustering algorithm includes:

[0103] Synchronously acquiring three-dimensional space point cloud data provided by the laser radar and target reflection point cloud data provided by the millimeter wave radar;

[0104] The three-dimensional space point cloud data and the target reflection point cloud data are converted to the body coordinate system centered on the unmanned ship through coordinate transformation;

[0105] In the unified coordinate system, data association is performed on the laser radar point cloud and the millimeter wave radar point cloud to generate a preliminary fusion point cloud covering space and reflection characteristics;

[0106] The original image collected by the visible light / infrared visual sensor is input into a pre-trained semantic segmentation neural network to assign a semantic label to each pixel and generate a semantic label image;

[0107] The fusion point cloud data is input into the density-based clustering algorithm, and the algorithm identifies a point cloud set that is closely distributed in space and has a density reaching a preset threshold a1 as an independent point cloud cluster;

[0108] All spatially isolated points that cannot be classified into any point cloud cluster are determined as sensor noise or transient clutter and are filtered out;

[0109] For the initially formed point cloud cluster, project it onto the semantic label map, and count the semantic label distribution of the image area covered by each point cloud cluster. If the proportion of sea wave label is not less than a preset threshold a2, it is determined that the cluster is a sea wave false target, and it is filtered out from the candidate target list;

[0110] For the remaining point cloud clusters, calculate the size of their three-dimensional bounding boxes, and filter out clusters whose size exceeds the preset physical range [a3, a4], and output the list of candidate obstacle target point cloud clusters screened out;

[0111] Convert the geometric center coordinates in the candidate obstacle target point cloud cluster list from the body coordinate system centered on the unmanned ship to the geodetic coordinate system.

[0112] It can be explained that the scheme can effectively avoid the deficiency of single sensor in complex sea condition monitoring by fusing laser radar point cloud, millimeter wave radar point cloud and visible light / infrared vision sensor collected data, and improve the robustness of unmanned ship in complex sea condition environment perception. Secondly, the scheme preprocesses the fused point cloud data through density clustering algorithm, selects candidate obstacle target point cloud cluster through spatial position close distribution and density, and generates candidate obstacle target point cloud cluster list. After point cloud clustering and screening, the geometric center of the target is converted to geodetic coordinate system, so as to avoid the huge calculation overhead brought by coordinate conversion of massive original point cloud, ensure the real-time performance of the system, and effectively improve the robustness and accuracy of obstacle identification. It lays a reliable data foundation for subsequent cooperative perception and path planning. Each element in the candidate obstacle target point cloud cluster list represents an identified independent candidate obstacle, and each element (i.e. each candidate obstacle) usually contains the following structured information:

[0113] Unique identifier: used to track the ID of the target;

[0114] Geometric center coordinates: three-dimensional center position coordinates of the point cloud cluster in the body coordinate system;

[0115] Three-dimensional size: length, width and height of the bounding box;

[0116] Point cloud cluster reference: pointer or index to the original point cloud data constituting the target;

[0117] Preliminary classification / attribute: label type divided according to the semantic label map, such as unknown, ship-shaped, buoy-shaped and wave, etc.;

[0118] Each unmanned ship independently generates a local candidate obstacle list of its own based on its local perspective, enabling each unmanned ship to independently perceive and understand its surroundings, which is the basis for its autonomous obstacle avoidance function; wherein the semantic segmentation neural network can adopt mature encoder-decoder architectures such as U-Net, DeepLab, etc. The network obtains the ability to classify input images at the pixel level by supervised training on image data sets containing annotations of marine scene obstacles (such as ships, buoys, waves, etc.). The semantic label map is a matrix of the same size as the input image, and the value of each element in the matrix represents the semantic category of the corresponding pixel point. The settings of a1, a2 and [a3, a4] are all based on prior knowledge analysis of typical obstacles and marine environment characteristics: the threshold a1 is determined by the physical size and point cloud density of typical small obstacles (such as buoys and small ships), aiming to ensure that such targets can be effectively clustered, while avoiding misclustering large-scale sea wave clutter as a single target; the threshold a2 is based on the spatial distribution characteristics of sea wave spray in visual images, and by analyzing the pixel connectivity of the wave area, a proportion threshold is set to distinguish between continuous waves and independent obstacles; the range [a3, a4] is based on the International Maritime Organization specifications and the obstacle sizes that may be encountered in actual navigation environments, the lower limit a3 is used to filter out irrelevant targets such as bird flocks and small floating objects, and the upper limit a4 is used to exclude super-large virtual images caused by multiple reflections or perception splicing errors, thereby ensuring that the selected targets have physical dimensions that pose a threat to navigation.

[0119] The real-time absolute motion state estimation of the candidate obstacle is obtained based on the Kalman filter according to the selected candidate obstacle point cloud cluster, and the obstacle feature list is output, which specifically includes:

[0120] A Kalman filter is initialized for each newly identified candidate obstacle, wherein the state vector of the Kalman filter includes the geodetic coordinates, velocity, heading angle and turning rate of the candidate obstacle point cloud cluster;

[0121] Based on the uniform turning motion model, the system state transition equation of the Kalman filter is constructed, and the geodetic coordinates of the current frame candidate obstacle point cloud cluster are taken as the observation vector;

[0122] Based on the Kalman filter, the updated state vector is output as the real-time and optimal absolute motion state estimation of the corresponding obstacle;

[0123] According to the velocity in the absolute motion state estimation output by the Kalman filter;

[0124] Set a motion determination threshold and a steering determination threshold, combine the speed value and the steering rate to determine the state of the obstacle, the state of the obstacle indicating static or motion;

[0125] According to the absolute motion state estimation of the obstacle, the relative distance, the relative speed and the relative heading of the obstacle relative to the ship are obtained through vector operation based on the earth coordinates and the heading angle of the unmanned ship;

[0126] According to the absolute motion state estimation of the obstacle, the relative distance, the relative speed and the relative heading of the obstacle relative to the ship, the state of the obstacle and the corresponding semantic label, a feature list of the obstacle is constructed;

[0127] The determination of the state of the obstacle specifically includes:

[0128] If the absolute speed value of the obstacle is not lower than the motion determination threshold for N1 frames or the absolute value of the absolute steering rate of the obstacle is not lower than the steering determination threshold for N1 frames, the obstacle is determined as a moving obstacle;

[0129] Otherwise, the obstacle is determined as a static obstacle;

[0130] When the obstacle is determined as a static obstacle, if the semantic label of the obstacle belongs to a predefined category set with kinetic energy, the obstacle is marked as a static obstacle with kinetic energy.

[0131] It can be explained that in the face of complex sea conditions, the unmanned ship not only needs to face fixed obstacles, but also needs to consider moving obstacles. If only the sensor detects the obstacles, there may be a problem of not timely obstacle avoidance. At the same time, since the unmanned ship is a multi-agent cooperative navigation, there may be a chaotic fast obstacle avoidance between each unmanned ship, especially for the obstacles with kinetic energy which are stationary in the early stage. Since there is a risk of sudden start, it leads to the problem of not timely obstacle avoidance between multi-agent, and causes safety accidents. Therefore, according to the selected candidate obstacle target point cloud cluster, the real-time absolute motion state estimation of the candidate obstacle is obtained based on the Kalman filter, and the speed and turning rate in the absolute motion state estimation are used to judge the state of the obstacle. At the same time, the obstacle with kinetic energy is marked and uploaded through the semantic label, so as to help the overall overall arrangement of the subsequent system. The Kalman filter can output the absolute motion state of the obstacle, and the core is that the filter operates in the geodetic coordinate system. Although the unmanned ship itself is moving, the geodetic coordinates observed by the unmanned ship for the same stationary obstacle are also changing. However, the state transition equation (uniform turning motion model) of the Kalman filter describes the motion law of the obstacle itself. In the iterative "prediction-update" process, the filter can spontaneously learn and separate the influence of the unmanned ship motion from the observation data by analyzing the coordinate change sequence in time sequence, so as to finally make the state vector converge to the real absolute motion state of the obstacle relative to the geodetic coordinate system. Therefore, by estimating the absolute motion state of the obstacle through the Kalman filter, the global situation awareness capability beyond the limitation of the coordinate system of the unmanned ship is provided, which not only can early warning and track the dynamic obstacles with clear motion, but also can identify those "static obstacles with kinetic energy" which are currently stationary but have the ability to move. Such targets are a major hidden danger in cooperative navigation, because they may suddenly start and break the safety premise of the original path planning. By uploading the state and semantic information of such targets to the cooperative decision center, the whole ship fleet can be provided with forward-looking risk warning, so as to realize the leap from passive obstacle avoidance to active risk avoidance, and fundamentally avoid the obstacle avoidance confusion and safety accidents caused by multi-agent due to sudden conditions. The pre-defined kinetic energy category set is a semantic label subset preset before system deployment, which includes the labels corresponding to all objects recognized as having autonomous moving ability, such as {ship, motorboat, engineering ship}. The set is constructed based on domain knowledge and can be configured according to task requirements. As a conservative safety strategy, unknown category labels output by the semantic segmentation network which are not in the pre-defined set are treated as obstacles with kinetic energy for processing. In the state vector initialization step, the initial values of the speed, heading angle and turning rate are set to zero. At the same time, the initial state error covariance matrix of the Kalman filter is set to be much larger than the variance term corresponding to the position state.This is a conservative engineering setting, aiming to reflect the high uncertainty of its initial value, so as to ensure that the filter prioritizes subsequent observation data and quickly converges to the true motion state. As a reference, the initial variance of the position state can be calibrated based on the joint measurement accuracy of lidar and millimeter wave radar; while the initial variance of the motion state is determined according to the prior knowledge of the maximum reasonable acceleration of the obstacle. The observation noise covariance matrix is determined according to the measurement error of the point cloud cluster geodetic coordinates. The motion determination threshold is set by the sum of 3 times the standard deviation of the ship navigation system speed measurement error and the detected current flow rate. The turning determination threshold is set according to the minimum observable turning rate of typical ship tactical maneuver. The N1 frame can reduce the misjudgment rate through multi-frame verification, avoiding single-frame data anomalies caused by sensor interference from sea spray in complex sea conditions. The setting method is set based on the 3σ criterion through testability experiment.

[0132] The safety radius compensation required by each unmanned ship for autonomous obstacle avoidance is determined according to the real-time data of each unmanned ship and the surrounding environment, which specifically includes:

[0133] According to the wave direction spectrum and wave height data collected by the unmanned ship, a sea state interference factor representing the severity of the sea state is calculated.

[0134] Based on the sea state interference factor and the wind flow vector data, the effective maneuverability degradation coefficient of the unmanned ship under the combined action of wind, wave and current is calculated through the ship motion response model.

[0135] Based on the effective maneuverability degradation coefficient, the safety radius compensation required by the unmanned ship for autonomous obstacle avoidance is calculated.

[0136] It can be explained that when the unmanned ship is avoiding obstacles, it needs to set a dynamic safety distance for obstacle avoidance according to the actual environment of the sea state, especially when multiple intelligent agents of unmanned ships are cooperating to avoid obstacles. Ensuring the dynamic safety distance of each unmanned ship is an important prerequisite for conflict. Therefore, the present scheme converts the qualitative influence of complex sea conditions into quantitative adjustment of obstacle avoidance safety radius, ensures the consistency of multi-agent cooperation of unmanned ships through unified quantitative logic, provides key environmental and capability quantitative support for subsequent cooperative path planning and conflict resolution, makes the cooperative obstacle avoidance strategy more targeted and safe, and effectively adapts to the navigation needs of unmanned ship cluster in complex sea conditions. Among them, the expression of the sea state interference factor representing the severity of the sea state is:

[0137]

[0138] In the formula, is the sea state interference factor representing the severity of the sea state, is the significant wave height, is the direction with the maximum energy density in the wave direction spectrum, is the heading angle of the ship, , is the weight coefficient, which can be set by experimental data;

[0139] By effectively combining the wave height (square term amplifies the influence of extreme sea conditions) and the angle between the wave direction and the heading (difference between downwind and headwind), the degree of interference of sea waves on the ship is quantified;

[0140] The expression of the effective maneuverability degradation coefficient of the unmanned ship under the combined action of wind, wave and current is:

[0141]

[0142] In the formula, is the effective maneuverability degradation coefficient of the unmanned ship under the combined action of wind, wave and current, is the wind speed collected, is the current speed, , , is the ship type parameter, which is the weight coefficient of the influence of wind speed on maneuverability, reflecting the sensitivity of the ship to wind load, is the weight coefficient of the influence of current speed on maneuverability, reflecting the sensitivity of the ship to water flow load, is the amplification coefficient of overall environmental interference, reflecting the sensitivity of the ship type to wind, wave and current, which can be calibrated by real ship test or CFD (computational fluid dynamics) simulation data;

[0143] This index can quickly reflect the degree of attenuation of the ship's maneuvering ability, solving the problem of complex ship motion response model output and difficulty in directly serving the obstacle avoidance parameter calculation;

[0144] The expression of the safety radius compensation amount required for the autonomous obstacle avoidance of the unmanned ship is:

[0145]

[0146] In the formula, is the safety radius compensation amount required for the autonomous obstacle avoidance of the unmanned ship, is the safety radius in still water;

[0147] The effective maneuverability degradation coefficient is converted into a specific obstacle avoidance safety radius compensation amount, so that the safety radius of each unmanned ship is dynamically adjusted through the safety radius compensation amount, wherein the safety radius is equal to the sum of the safety radius in still water and the safety radius compensation amount. In particular, for extreme sea conditions that may cause the safety radius compensation amount to abnormally increase, by setting the lower limit (such as μ≥0.1), the calculation result caused by excessive environmental interference is avoided.

[0148] ReferenceFigure 3 As shown, the acquisition of the current time of each unmanned ship safety navigable space specifically includes:

[0149] According to the length, width, height of all static obstacle bounding boxes in the list, the distance of the outward expansion of the safety radius, a static prohibited area is generated;

[0150] For each moving obstacle in the obstacle feature list, based on the real-time absolute motion state estimation output by its Kalman filter, the uniform speed turning motion model is substituted into it, forward numerical simulation is carried out, the positions of a series of discrete time points in the future period of time are predicted, and thus the predicted trajectory of the obstacle is generated;

[0151] According to the predicted trajectory of the obstacle, combined with the length, width and height of the obstacle bounding box, a safety radius is expanded along the center line of the trajectory on both sides, and a dynamic threat corridor changing with time is generated;

[0152] The static prohibited area and the dynamic threat corridor are removed from the global environment to obtain the current time of each unmanned ship safety navigable space.

[0153] It can be explained that the core premise of autonomous obstacle avoidance of unmanned ship is to clarify the "safe passage boundary", and the threat of static obstacles (such as rocks, anchored ships) is fixed under complex sea conditions, and the threat of dynamic obstacles (such as ships in navigation, drifting floats) will continue to change with its motion, if only relying on the current position of the obstacle, it is easy to cause obstacle avoidance failure due to the trajectory deviation of the dynamic target, therefore, based on the obstacle feature list and the safety radius of each ship, for static obstacles, a static prohibited area is constructed by "bounding box expansion safety radius", to ensure a sufficient safety distance from the fixed obstacle; for dynamic obstacles, combined with the absolute motion state estimation (Kalman filter output) and the uniform speed turning motion model, the trajectory in the future period of time is predicted, and the safety radius is expanded to form a dynamic threat corridor, which covers the area that may be reached in advance, finally, the two types of threat areas are removed from the global environment, and the safety navigable space of each unmanned ship is obtained, this scheme not only adapts to the "fixed threat" characteristics of static obstacles, but also covers the "motion threat" characteristics of dynamic obstacles, provides accurate spatial constraints for subsequent multi-agent collaborative path planning, avoids obstacle avoidance conflicts caused by ambiguous obstacle area definition, and fully adapts to the autonomous navigation needs of unmanned ship cluster under complex sea conditions.

[0154] Referring to Figure 4 As shown, the construction of a global synthetic potential field to generate a set of optimal candidate paths of unmanned ships from the current position to the target point specifically includes:

[0155] Taking the current position of each unmanned ship as the origin, a local plane coordinate system is established based on the geodetic coordinate system, the safety navigable space is converted to the local coordinate system, and a local search space for path search is formed;

[0156] Discretize the continuous state space of the unmanned ship into a four-dimensional space-time state grid in the local plane coordinate system, each node of the state grid representing a specific state;

[0157] The specific state includes: the position of the unmanned ship, the heading angle of the unmanned ship, and the time from the start of path generation;

[0158] Based on the physical characteristics of the unmanned ship and expert experience, set the node parameters to define a search granularity for the specific state in the space-time state grid that matches the physical characteristics of the unmanned ship;

[0159] In the local plane coordinate system, construct a global synthetic potential field composed of a target attractive field and an obstacle repulsive field;

[0160] The target attractive field refers to the potential energy value at any node, which is calculated from the Euclidean distance from the node to the target node;

[0161] The obstacle repulsive field refers to the potential energy value at any node, which is calculated from the Euclidean distance from the node to the nearest obstacle;

[0162] By linearly superimposing the target attractive field and the obstacle repulsive field, the global synthetic potential field is obtained;

[0163] Starting from the current state node of the unmanned ship, calculate the negative gradient direction of the node position in the synthetic potential field;

[0164] Taking the negative gradient direction as the center, sample the candidate heading angle within the preset offset interval according to the heading angle resolution;

[0165] Based on the candidate heading angle, combined with the current speed and time step, calculate the position coordinates of the subsequent node through the ship kinematics model to generate the corresponding subsequent state node, thereby forming a candidate polyline path node sequence;

[0166] Immediately verify whether the new node and the line segment formed with the previous node have an intersection with the obstacle region, and only keep the nodes without collision;

[0167] Smooth each polyline path in the set of collision-free candidate polyline paths to obtain a set of smoothed candidate paths;

[0168] Calculate the total geometric length of each path in the set of smoothed candidate paths, and select the candidate path with the smallest total geometric length as the optimal candidate path for the unmanned ship from the current position to the target point.

[0169] It can be explained that after the unmanned ship identifies the safe navigable space, only the optimal action route needs to be selected in the safe navigable space, in the ideal state without obstacles, only the straight line distance of the unmanned ship to the target position needs to be considered, and when there are obstacles, the geometric distance of the unmanned ship to the target position needs to be known when avoiding the obstacles is the shortest, therefore, the scheme discretizes the continuous state space of the unmanned ship into a four-dimensional space-time state grid under the local plane coordinate system, defines a search granularity matched with the physical characteristics of the unmanned ship, the position of the unmanned ship defines the spatial resolution of the position update in the path planning, limits the minimum displacement considered by the unmanned ship in a single decision cycle, and matches the ship size and inertia; the heading angle of the unmanned ship defines the minimum turning angle considered by the unmanned ship in a single decision cycle, and matches the ship turning rate;The time unit defined by the start time of path generation determines the minimum time unit for generating the polyline path, dictating the time precision with which the algorithm predicts the future motion of the ship and dynamic obstacles, and matching the system response frequency. Smaller search granularity results in more accurate predicted polyline paths, but at the cost of computational complexity and reaction time. This can be calibrated after conducting test experiments based on the physical characteristics of the unmanned vessel and expert experience. Furthermore, after generating the local planar coordinate system and search granularity, it is only necessary to use the UAV's current position as the initial point and generate several sets of polylines extending towards the target point through gridding. The proposed path is a polygonal path, but it contains many redundant elements, such as paths that intersect with obstacle areas or are too long. Therefore, this solution constructs a global synthetic potential field. The target gravitational field generates a potential energy gradient pointing towards the target to ensure the path's purposefulness, while the obstacle repulsive field generates a potential energy gradient away from the obstacles to ensure the path's safety. The global synthetic potential field, formed by the linear superposition of these two forces, creates a unified virtual environment that simultaneously represents both "target attraction" and "obstacle repulsion." The movement of the unmanned vessel within this environment can be intuitively understood. To glide along the steepest downhill slope in the "terrain," a path that can safely avoid obstacles and efficiently approach the target is naturally found. Based on this, the spatiotemporal trajectory of new nodes and the line segments formed by them and previous nodes is checked to see if they intersect with obstacle areas (which have already been identified in previous sections). This eliminates intersecting polyline paths, significantly reducing redundant paths. Finally, by calculating the total geometric length of all paths in the smoothed candidate path set, the candidate path with the smallest total geometric length is selected as the optimal candidate route for the unmanned vessel from its current position to the target point. The path selection method does not sacrifice the real-time obstacle avoidance performance of the unmanned vessel and improves its response rate. The preset offset interval is centered on the negative gradient direction. The lower bound of the interval is the negative gradient direction minus a maximum allowable left deflection angle, and the upper bound is the negative gradient direction plus a maximum allowable right deflection angle. The maximum allowable left deflection angle and the maximum allowable right deflection angle are equal in value. The value is based on the maximum heading change that the unmanned vessel can achieve based on its maximum turning rate within a single time step. This maximum heading change is measured experimentally.

[0170] The expression for the target gravitational field is as follows:

[0171]

[0172] In the formula, For in position The gravitational potential energy at the target location. The gravitational gain coefficient is a positive constant. For position The Euclidean distance from the target node;

[0173] The expression for the repulsive force field of the obstacle is:

[0174]

[0175] In the formula, For in position The repulsive potential energy of the obstacle at the location, The repulsive gain coefficient is a positive constant, and its value is much larger than that of the gravitational gain coefficient. The threshold for the impact of obstacles;

[0176] The threshold value for the obstacle impact is determined by the dynamic safety radius of the unmanned vessel and a safety margin, and its calculation formula is as follows: ,in, For the safety radius, As a safety margin, the value of the safety margin is determined through test experiments. The experiments need to be carried out separately for different types of obstacles: for static obstacles (such as reefs and fixed buoys), the unmanned surface vessel is tested under different sea conditions, starting from a distance of [missing information]. The test assesses the maximum turning rate and the ability of the unmanned surface vessel (USV) to change course and avoid obstacles within a single time step, recording the minimum safety margin. For dynamic obstacles (such as sailing vessels), the test utilizes a uniform turning motion model to determine the additional safety margin required in scenarios with relative motion to the obstacle, ensuring the trajectory of the USV is covered by the obstacle. For stationary obstacles with kinetic energy (such as anchored vessels or motorboats), the test simulates extreme scenarios of sudden acceleration, testing the required additional safety margin to address the uncertainty of the unknown starting direction. Based on the test results for these three types of obstacles, the test determines the appropriate safety margin values ​​for different scenarios, ensuring the USV maintains a safe distance from obstacles. At that time, the unmanned surface vessel (USV) is capable of successfully avoiding the obstacle within a single time step by applying a change in heading caused by its maximum turning rate. The repulsive gain coefficient is calibrated as follows: a critical obstacle avoidance scenario is set in the simulation, and the value of the repulsive gain coefficient is gradually increased until the USV can maintain a minimum distance greater than or equal to a safety threshold from the obstacle. This value is then determined as the minimum effective value of the repulsive gain coefficient. The gravitational gain coefficient is calibrated as follows: it is first set to be equal to the minimum effective value of the repulsive gain coefficient. , and then in the unobstructed scene, fine-tune the value in the range of ±20% of the value, and select the value that can make the total length of the navigation path of the unmanned ship from the starting point to the target point shortest, wherein, in the obstacle avoidance process, the target point is dynamically set as a sub-target point located behind the obstacle on the planned route, that is, in the local plane coordinate system, a straight line parallel to the current heading of the unmanned ship and tangent to the outer contour of the obstacle is drawn, and the position behind the obstacle in the direction of the planned route after the tangent point is offset by a safe backtracking distance is taken as the sub-target point, and the safe backtracking distance is determined by a test experiment, and the value is the minimum look-ahead amount required to ensure that the unmanned ship can smoothly and stably cut back to the planned route with the minimum turning radius after avoiding the obstacle.

[0177] The consensus binding algorithm-based selected path of each unmanned ship from the current position to the target point specifically comprises:

[0178] Each unmanned ship encapsulates the optimal selected path generated by itself and the total geometric length of the path into a data packet and broadcasts the data packet to all other unmanned ships in the cluster through wireless ad hoc networking.

[0179] After receiving the data packet from other ships, each unmanned ship independently constructs a global collaborative decision matrix in the local processing system of each unmanned ship.

[0180] The column dimension of the collaborative decision matrix corresponds to the optimal selected path of each ship and the total geometric length of the path.

[0181] The consensus binding algorithm is used for distributed path combination optimization of the unmanned ship cluster, and a consensus-based multi-agent collaborative obstacle avoidance path of the unmanned ship is output.

[0182] Each unmanned ship receives and executes the collaborative path allocated to it in the final path combination, and continues to travel to the final target point according to the preset formation after the obstacle avoidance is completed.

[0183] It can be explained that, due to the multi-agent cooperative driving of unmanned ships, after each unmanned ship generates a candidate path from the current position to the target point, it cannot be directly executed, and it needs to be considered whether the respective candidate path conflicts with the whole, therefore, a consensus needs to be reached to unify the execution, therefore, the scheme broadcasts the candidate path generated by each unmanned ship and the total geometric length data packet of the path to all other unmanned ships in the cluster through wireless ad hoc networking, and independently constructs a global cooperative decision matrix in the local processing system, this decentralized design avoids the single point failure risk of traditional master-slave architecture, even if part of the nodes in the cluster fail, the remaining nodes can still make cooperative decisions based on the local matrix, ensuring the survivability of the system, and through the global cooperative decision matrix, the subsequent "consensus binding optimization" step is consensus to all ships and executed in parallel, greatly shortening the time to seek the global optimal solution, meeting the stringent requirements of real-time in complex environments, finally, the consensus binding algorithm is used to optimize the distributed path combination of the unmanned ship cluster, outputting a cooperative obstacle avoidance path of the unmanned ship multi-agent, and after the obstacle avoidance is completed, the array continues to move towards the final target point according to the preset array, ensuring the consistency of the array with the initial state, so as to realize the cooperative obstacle avoidance of the unmanned ship multi-agent in complex sea conditions.

[0184] The consensus binding algorithm for distributed path combination optimization of the unmanned ship cluster specifically includes:

[0185] Each unmanned ship selects its own optimal candidate path to form an initial path combination;

[0186] Based on the initial path combination, the closest point distance between all unmanned ships is calculated, and the unmanned ship pairs with a closest point distance less than the safety distance are marked as conflict pairs and bundled into a conflict cluster;

[0187] Each unmanned ship in the conflict cluster acts as a bidder and selects a replacement path from its candidate path set with a closest point distance greater than the safety distance and the shortest total geometric length;

[0188] The reduction of the total geometric length of the entire cluster path after using the replacement path is calculated, that is, the cost improvement value, and the improvement value is broadcast as a tender;

[0189] All unmanned ships receive the tender and independently select the tender with the largest cost improvement value as the winning tender;

[0190] Update the path combination according to the winning tender and update the conflict cluster, repeat the process until all conflict pairs are eliminated or the maximum number of iterations is reached, at which time the output path combination is the cooperative obstacle avoidance path combination.

[0191] It can be explained that in the cooperative obstacle avoidance scene of unmanned ship cluster, the local path generated by each ship independently can avoid environmental obstacles, but it is easy to cause navigation conflict between ships (such as meeting and crossing encounter). Therefore, the global consensus at the cluster level is indispensable. Instead of dividing the formation of the ship formation statically, the scheme dynamically allocates a suitable obstacle avoidance path to each ship through the consensus binding algorithm, thereby naturally forming a cooperative and orderly traffic sequence or shunt group in logic. This process strictly follows the principle of minimizing the total geometric length of the cluster path (complementary to the local optimal goal of single ship path selection), aiming to ensure the global navigation efficiency is optimal. Through distributed bidding and local consensus, the complex global conflict resolution problem is decomposed into a series of local conflict resolution proposals that can be processed in parallel by each ship, so that the system can quickly converge to a consensus on a conflict-free cooperative obstacle avoidance path combination without losing the timeliness of decision-making. Thus, the inherent single-point failure risk of the central controller is eliminated, and the robustness and survivability of the system are improved. At the same time, through the information broadcasting and local decision-making mode, the distributed communication characteristics of the wireless ad hoc network (Ad-hoc) are matched, effectively avoiding the dependence on high-bandwidth and low-latency backbone communication networks, and thus adapting to complex sea conditions operating scenarios with limited communication conditions and changing environments.

[0192] Further, based on the same inventive concept as the above unmanned ship multi-agent cooperative obstacle avoidance method for complex sea conditions, the scheme proposes an unmanned ship multi-agent cooperative obstacle avoidance system for complex sea conditions, comprising:

[0193] A data acquisition module, the data acquisition module is used for acquiring real-time data of each unmanned ship and the surrounding environment through a multi-sensor fusion system carried on each unmanned ship;

[0194] A navigable space module, the navigable space module is used for screening candidate obstacle target point cloud clusters based on a density clustering algorithm according to the surrounding obstacle point cloud data, and generating a candidate obstacle target point cloud cluster list; obtaining real-time absolute motion state estimation of the candidate obstacle based on a Kalman filter according to the screened candidate obstacle target point cloud cluster, and outputting an obstacle feature list; determining a safety radius compensation required for autonomous obstacle avoidance of each unmanned ship according to the real-time data of each unmanned ship and the surrounding environment; and obtaining the safety navigable space of each unmanned ship at the current time according to the obstacle feature list in combination with the safety radius required for autonomous obstacle avoidance of each unmanned ship;

[0195] A cooperative obstacle avoidance module, the cooperative obstacle avoidance module is used for each unmanned ship to construct a global synthetic potential field based on its safety navigable space, generate a set of optimal candidate paths of the unmanned ship from the current position to the target point; and obtain a consensus unmanned ship multi-agent cooperative obstacle avoidance path based on a consensus binding algorithm according to the candidate path of each unmanned ship from the current position to the target point.

[0196] The navigable space module specifically comprises:

[0197] A point cloud cluster unit is configured to filter candidate obstacle target point cloud clusters based on a density clustering algorithm according to the surrounding obstacle point cloud data, and generate a candidate obstacle target point cloud cluster list;

[0198] An obstacle feature unit is configured to obtain real-time absolute motion state estimation of the candidate obstacle based on a Kalman filter according to the filtered candidate obstacle target point cloud cluster, and output an obstacle feature list;

[0199] A safety radius unit is configured to determine a safety radius compensation required for autonomous obstacle avoidance of each unmanned ship according to real-time data of each unmanned ship and the surrounding environment;

[0200] A navigable space unit is configured to obtain the safety navigable space of each unmanned ship at the current time according to the obstacle feature list and the safety radius required for autonomous obstacle avoidance of each unmanned ship;

[0201] The cooperative obstacle avoidance module specifically comprises:

[0202] An optimal path unit is configured to construct a global synthetic potential field based on the safety navigable space of each unmanned ship, and generate a set of optimal candidate paths of the unmanned ship from the current position to the target point;

[0203] A cooperative obstacle avoidance unit is configured to obtain a consensus unmanned ship multi-agent cooperative obstacle avoidance path based on a consensus binding algorithm according to the candidate paths of each unmanned ship from the current position to the target point.

[0204] In summary, the advantages of the present application are that the multi-unmanned ship cluster realizes safe, efficient and cooperative intelligent obstacle avoidance in a real complex marine environment.

[0205] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions, characterized in that, include: Real-time data of each unmanned vessel and its surrounding environment are acquired through a multi-sensor fusion system mounted on each unmanned vessel. Based on the surrounding obstacle point cloud data, a density clustering algorithm is used to filter candidate obstacle target point cloud clusters and generate a list of candidate obstacle target point cloud clusters. Based on the selected candidate obstacle target point cloud clusters, the real-time absolute motion state estimation of the candidate obstacles is obtained using a Kalman filter, and a list of obstacle features is output. Based on real-time data of each unmanned vessel and its surrounding environment, determine the safety radius compensation required for each unmanned vessel to autonomously avoid obstacles. Based on the list of obstacles and the safe radius required for each unmanned vessel to autonomously avoid obstacles, the safe navigation space of each unmanned vessel at the current moment is obtained. Each unmanned surface vessel (USV) constructs a global synthetic potential field based on its own safe navigation space, generating a set of optimal alternative paths from its current position to the target point. Based on the alternative paths of each unmanned vessel from its current position to the target point, a consensus-based binding algorithm is used to obtain a consensus-based multi-agent cooperative obstacle avoidance path for unmanned vessels. The determination of the safety radius compensation required for autonomous obstacle avoidance by each unmanned vessel based on real-time data of each unmanned vessel and its surrounding environment specifically includes: Based on the wave direction spectrum and wave height data collected by the unmanned vessel, the sea state disturbance factor, which characterizes the severity of sea conditions, is calculated. Based on sea state disturbance factors and wind current vector data, the effective maneuverability degradation coefficient of unmanned vessels under the combined action of wind, waves and current is calculated through a ship motion response model. Based on the effective maneuverability degradation coefficient, the safety radius compensation required for autonomous obstacle avoidance by unmanned vessels is calculated.

2. The multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions according to claim 1, characterized in that, The acquisition of real-time data on each unmanned vessel and its surrounding environment through a multi-sensor fusion system mounted on each unmanned vessel specifically includes: The real-time data includes: the position and attitude status of each unmanned vessel, point cloud data of surrounding obstacles, wave direction spectrum and wave height data, and wind vector data; The multi-sensor fusion system includes: an inertial measurement unit, a GPS positioning module, a lidar, a millimeter-wave radar, a visible light / infrared vision sensor, a wave sensor, and a wind speed and direction sensor; The inertial measurement unit and GPS positioning module are combined to collect the ship's real-time position and attitude status, including longitude, latitude, heading angle, speed and roll / pitch / heave angles; LiDAR and millimeter-wave radar work together to scan the surrounding environment and acquire raw point cloud data of obstacles in the surrounding area, while visible light / infrared vision sensors assist in acquiring image features of obstacles; Wave sensors collect wave direction spectrum, significant wave height and wave period data based on spectral analysis technology, while wind speed and direction sensors collect instantaneous wind speed, wind direction and ocean current speed and direction data; All data collected by the sensors are spatiotemporally aligned, and the time dimension of the data from each sensor is unified based on the GPS timestamp. The Kalman filter algorithm is used to smooth the coordinate-aligned sensor data, remove outliers, and normalize the filtered data to the [0,1] interval. Each unmanned vessel shares its pre-processed real-time data with the fleet data center via a wireless communication network.

3. The multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions according to claim 2, characterized in that, The step of filtering candidate obstacle target point cloud clusters based on surrounding obstacle point cloud data using a density clustering algorithm and generating a list of candidate obstacle target point cloud clusters specifically includes: Simultaneously acquire 3D spatial point cloud data provided by lidar and target reflection point cloud data provided by millimeter-wave radar; The three-dimensional spatial point cloud data and the target reflection point cloud data are uniformly transformed into the body coordinate system centered on the unmanned vessel through coordinate transformation. Under a unified coordinate system, the point clouds of lidar and millimeter-wave radar are correlated to generate a preliminary fused point cloud covering spatial and reflection characteristics. The raw images acquired by the visible light / infrared vision sensor are input into a pre-trained semantic segmentation neural network, which assigns a semantic label to each pixel and generates a semantic label map. The fused point cloud data is input into a density clustering algorithm. The algorithm identifies a set of point clouds that are spatially close and whose density reaches a preset threshold a1 as an independent point cloud cluster. All spatially isolated points that cannot be classified into any point cloud cluster are identified as sensor noise or transient clutter and filtered out. For the initially formed point cloud clusters, project them onto the semantic label map, and count the distribution of semantic labels in the image area covered by each point cloud cluster. If the proportion of wave labels is not less than the preset threshold a2, then the cluster is determined to be a false wave target and is filtered out from the candidate target list. For the remaining point cloud clusters, calculate their 3D bounding box size, filter out clusters whose size exceeds the preset physical range [a3,a4], and output the filtered list of candidate obstacle target point cloud clusters; The geometric center coordinates of the selected candidate obstacle target point cloud clusters are transformed from the body coordinate system centered on the unmanned vessel to the geodetic coordinate system.

4. The multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions according to claim 3, characterized in that, The step of obtaining real-time absolute motion state estimates of candidate obstacles based on the selected candidate obstacle point cloud clusters using a Kalman filter and outputting an obstacle feature list specifically includes: A Kalman filter is initialized for each newly identified candidate obstacle, wherein the state vector of the Kalman filter includes: the ground coordinates, velocity, heading angle and turning rate of the candidate obstacle point cloud cluster; Based on the uniform turning motion model, the system state transition equation of the Kalman filter is constructed, and the geodetic coordinates of the candidate obstacle point cloud cluster in the current frame are used as the observation vector. Based on the Kalman filter, the updated state vector is output as the real-time and optimal absolute motion state estimate of the corresponding obstacle; The velocity is estimated from the absolute motion state output by the Kalman filter; Set motion determination thresholds and turning determination thresholds, and combine speed value and turning rate to determine the state of the obstacle, wherein the state of the obstacle refers to being stationary or moving; Based on the absolute motion state estimation of the obstacle, and using the geodetic coordinates and heading angle of the unmanned vessel as a reference, vector calculations are used to obtain the relative distance, relative speed, and relative heading of the obstacle relative to the vessel. Based on the absolute motion state estimation of obstacles, the relative distance, relative speed, and relative heading of obstacles relative to the ship, the state of obstacles and their corresponding semantic labels, an obstacle feature list is constructed. The determination of the state of the obstacle specifically includes: If the absolute velocity value remains above the motion determination threshold for N1 consecutive frames, or the absolute turning rate value remains above the turning determination threshold for N1 consecutive frames, it is determined to be a moving obstacle; otherwise, it is determined to be a stationary obstacle. When an obstacle is determined to be stationary, if its semantic label belongs to a predefined set of categories with kinetic energy, it is then marked as a stationary obstacle with kinetic energy.

5. The multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions according to claim 4, characterized in that, The process of obtaining the safe navigable space of each unmanned surface vessel (USV) at the current moment based on the obstacle feature list and the safe radius required for autonomous obstacle avoidance specifically includes: Based on the length, width, and height of the bounding boxes of all stationary obstacles in the list, and the distance of the safety radius of expansion, a static prohibited area is generated. For each moving obstacle in the obstacle feature list, based on the real-time absolute motion state estimate of its Kalman filter output, it is substituted into its uniform turning motion model to perform forward numerical simulation, predict its position at a series of discrete time points in the future, and thus generate the predicted trajectory of the obstacle. Based on the predicted trajectory of the obstacle, and combined with the length, width, and height of the obstacle bounding box, a safety radius is extended on both sides of the trajectory centerline to generate a dynamic threat corridor that changes over time. By removing static prohibited areas and dynamic threat corridors from the global environment, the safe navigation space of each unmanned vessel at the current moment is obtained.

6. The multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions according to claim 5, characterized in that, Each unmanned surface vessel (USV) constructs a global synthetic potential field based on its own safe navigation space, generating a set of optimal alternative paths from its current position to the target point, specifically including: Using the current position of each unmanned vessel as the origin, a local planar coordinate system is established based on the geodetic coordinate system. The safe and navigable space is transformed into this local coordinate system to form a local search space for path search. In the local planar coordinate system, the continuous state space of the unmanned vessel is discretized into a four-dimensional spatiotemporal state lattice. Each node of the state lattice represents a specific state, which includes: the position of the unmanned vessel, the heading angle of the unmanned vessel, and the time calculated from the start time of path generation. Based on the physical characteristics of unmanned vessels and expert experience, node parameters are set to define a search granularity that matches the physical characteristics of unmanned vessels for a specific state in the spatiotemporal state grid. In the local planar coordinate system, a global composite potential field is constructed, which consists of the gravitational field of the target and the repulsive field of the obstacle; The target gravitational field refers to the potential energy value at any node, which is calculated from the Euclidean distance from that node to the target node; The obstacle repulsive field refers to the potential energy value at any node, which is calculated from the Euclidean distance from that node to the nearest obstacle. The global composite potential field is obtained by linearly superimposing the gravitational field of the target and the repulsive field of the obstacle; Starting from the current state node of the unmanned vessel, calculate the negative gradient direction of the node's position in the synthetic potential field; With the negative gradient direction as the center, sample according to the heading angle resolution within the preset offset interval to generate candidate heading angles; Based on the candidate heading angle, combined with the current speed and time step, the position coordinates of subsequent nodes are calculated through the ship kinematics model to generate the corresponding subsequent state nodes, thus forming a candidate polyline path node sequence. Immediately verify whether the spatiotemporal trajectory of the new node and the line segment formed by the previous node intersects with the obstacle area, and only retain nodes that do not collide; Smooth each polyline path in the set of collision-free candidate polyline paths to obtain a smoothed set of candidate paths. Calculate the total geometric length of each path in the smoothed candidate path set, and select the candidate path with the smallest total geometric length as the optimal candidate path for the unmanned vessel to travel from its current position to the target point.

7. A multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions according to claim 6, characterized in that, The step of obtaining a consensus-based multi-agent cooperative obstacle avoidance path for each unmanned surface vessel (USV) based on its alternative paths from its current location to the target point, using a consensus-based binding algorithm, specifically includes: Each unmanned surface vessel encapsulates its generated optimal alternative path and the total geometric length of that path into a data packet, and broadcasts it to all other unmanned surface vessels in the cluster via a wireless ad hoc network. After receiving data packets from other vessels, each unmanned vessel independently constructs a global collaborative decision matrix in its local processing system. Wherein, the row dimension of the collaborative decision matrix corresponds to the number of each unmanned vessel in the cluster, and the column dimension corresponds to the optimal alternative path and the total geometric length of the path for each vessel. A consensus-based binding algorithm is used to optimize the distributed path combination of unmanned surface vessels (USVs) swarms, and outputs a consensus-based multi-agent cooperative obstacle avoidance path for USVs. Each unmanned vessel receives and executes the cooperative path assigned to it in the final path combination, and continues to travel towards the final target point in a preset formation after obstacle avoidance.

8. A multi-agent cooperative obstacle avoidance method for unmanned surface vessels in complex sea conditions according to claim 7, characterized in that, The aforementioned distributed path combination optimization of the unmanned vessel swarm using a consensus-based binding algorithm specifically includes: Each unmanned vessel selects its own optimal alternative path to form an initial path combination; Based on the initial path combination, calculate the closest point distance between all pairs of unmanned vessels, and mark the pairs of unmanned vessels whose closest point distance is less than the safe distance as conflict pairs and bundle them into a conflict cluster; Each unmanned vessel within the conflict cluster acts as a bidder, selecting from its own set of alternative paths the closest point with a distance greater than the safe distance and the shortest total geometric length. Calculate the reduction in the total geometric length of the entire cluster path after adopting this alternative path, i.e., the cost improvement value, and broadcast this improvement value as a tender. All unmanned surface vessels (USVs) receive bids and independently select the bid with the highest cost improvement value as the winning bid. Update the path combination based on the winning bid and update the conflict cluster. Repeat this process until all conflict pairs are eliminated or the maximum number of iterations is reached. The output path combination at this point is the cooperative obstacle avoidance path combination.

9. A multi-agent cooperative obstacle avoidance system for unmanned surface vessels in complex sea conditions, characterized in that: To implement the multi-agent cooperative obstacle avoidance method for unmanned vessels in complex sea conditions as described in any one of claims 1-8, the method includes: The data acquisition module is used to acquire real-time data of each unmanned vessel and its surrounding environment through a multi-sensor fusion system mounted on each unmanned vessel. The navigable space module is used to filter candidate obstacle target point cloud clusters based on density clustering algorithm according to the surrounding obstacle point cloud data, and generate a list of candidate obstacle target point cloud clusters; based on the filtered candidate obstacle target point cloud clusters, it obtains the real-time absolute motion state estimate of the candidate obstacles based on Kalman filter, and outputs an obstacle feature list; based on the real-time data of each unmanned vessel and the surrounding environment, it determines the safety radius compensation amount required for autonomous obstacle avoidance of each unmanned vessel; based on the obstacle feature list and combined with the safety radius required for autonomous obstacle avoidance of each unmanned vessel, it obtains the safe navigable space of each unmanned vessel at the current moment. The collaborative obstacle avoidance module is used by each unmanned vessel to construct a global synthetic potential field based on its own safe navigation space, and generate a set of optimal alternative paths for the unmanned vessels from the current position to the target point; based on the alternative paths of each unmanned vessel from the current position to the target point, a consensus-based binding algorithm is used to obtain a consensus-based multi-agent collaborative obstacle avoidance path for the unmanned vessels. The navigable space module specifically includes: A point cloud cluster unit is used to filter candidate obstacle target point cloud clusters based on density clustering algorithm according to the surrounding obstacle point cloud data, and generate a list of candidate obstacle target point cloud clusters; An obstacle feature unit is used to obtain real-time absolute motion state estimates of candidate obstacles based on Kalman filters, according to the selected candidate obstacle target point cloud clusters, and output an obstacle feature list. The safety radius unit is used to determine the safety radius compensation amount required for autonomous obstacle avoidance of each unmanned vessel based on real-time data of each unmanned vessel and its surrounding environment. The navigable space unit is used to obtain the safe navigable space of each unmanned vessel at the current moment based on the obstacle feature list and the safety radius required for autonomous obstacle avoidance of each unmanned vessel. The collaborative obstacle avoidance module specifically includes: The optimal path unit is used by each unmanned vessel to construct a global synthetic potential field based on its own safe navigation space, and generate a set of optimal alternative paths for the unmanned vessel from its current position to the target point. The collaborative obstacle avoidance unit is used to obtain a consensus-based multi-agent collaborative obstacle avoidance path for each unmanned vessel based on the alternative paths from its current position to the target point and a consensus-based binding algorithm.

Images