Intelligent obstacle avoidance and path planning method based on ship

Abstract

The invention provides an intelligent obstacle avoidance and path planning method based on a ship, and relates to the technical field of intelligent control of small leisure ships. S1, the state and surrounding environment information of the ship are obtained through a multi-sensor group deployed on the ship, and a distortion compensation algorithm combining Kalman filtering with a sea condition coefficient is adopted; outputting the compensated obstacle information, sea condition parameters and ship task types; s2, the local end uploads the compensated obstacle information and the sea condition parameters output in the step S1 to a server end, the server end introduces sea condition correction factors to construct a CPA model, collision risk indexes are calculated through the CPA model, and the path conflict probability between ship pairs in the water area is predicted according to a path conflict probability calculation formula. According to the invention, accurate risk judgment can be ensured through collision risk correction, and multi-ship ordered avoidance is realized through scenarized priority, so that the problems of path conflict, low decision-making efficiency and increased navigation risk caused by lack of collaboration in a multi-ship environment are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of small recreational boats, and in particular to an intelligent obstacle avoidance and path planning method based on a boat. BACKGROUND

[0002] With the vigorous development of the water leisure industry, small recreational boats such as fishing boats and sightseeing boats are increasingly widely used in nearshore waters. The density of boats in closed or semi-closed waters such as scenic spots and ports is increasing. There are dynamic sea conditions, static obstacles, dynamic obstacles, and underwater obstacles. Dynamic sea conditions are characterized by local sudden wind waves and irregular currents. Static obstacles include piers and buoys. Dynamic obstacles include water surface navigation boats. Underwater obstacles include underwater hidden reefs and other types. These threats pose a serious threat to the safety of boat navigation. Therefore, the autonomous obstacle avoidance capability of the boat is extremely important.

[0003] To address the above problems, existing technologies such as CN107748561B and CN106909145B mainly solve the obstacle avoidance problem of a single boat or a limited number of boats. The core is to make a reactive obstacle avoidance decision based on local sensor perception. In the prior art disclosed in CN107748561B, a local obstacle avoidance method for unmanned ships based on multiple sensing parameters is disclosed. A multi-sensor fusion strategy is used with laser radar, AIS system, and camera. A local environment grid map is constructed using laser radar. Ship navigation information is obtained in combination with AIS. Camera is used to assist in identifying obstacle types. In CN108445879A, an unmanned boat obstacle avoidance method based on predicted collision danger area is proposed. The CPA (Closest Point of Approach) model is used to calculate the collision risk of two ships. The t CPA (time to closest point of approach) and d CPA (distance to closest point of approach) of two ships are calculated to determine the priority of avoidance. However, the above technologies do not fully adapt to the complexity of nearshore dynamic sea conditions.

[0004] In summary, the existing technology makes decisions based on local information for each boat, which cannot predict the intentions of other boats. When multiple boats intersect, decision conflicts can lead to stalemates or collision risks. After a boat avoids an obstacle, it may trigger obstacle avoidance again when returning to the preset route, resulting in unnecessary maneuvers, reduced navigation efficiency and comfort, and the inability to coordinate avoidance like human drivers following clear rules. Therefore, the existing technology cannot effectively solve the problem of high-density and high-dynamic coordinated obstacle avoidance of multiple boats in limited water areas. SUMMARY

[0005] In view of the above problems in the prior art, the technical problem to be solved by the present application is to provide an intelligent obstacle avoidance and path planning method based on a ship, which can accurately correct the risk of collision through risk judgment, realize orderly avoidance of multiple ships through scenario-based priority, and ensure the feasibility of the path through sea state adaptive path planning, so as to ultimately form a closed loop of "perception, judgment, decision-making and control", thereby solving the problems of path conflict, low decision-making efficiency and high sailing risk caused by lack of cooperation in a multiple ship environment.

[0006] To solve the above technical problems, the technical solution adopted by the present application is as follows: The present application provides an intelligent obstacle avoidance and path planning method based on a ship, which comprises the following steps: S1, acquiring the state of the ship and the surrounding environment information through a plurality of sensors arranged on the ship, and outputting the compensated obstacle information, sea state parameters and ship task type by using a Kalman filter combined with a distortion compensation algorithm of a sea state coefficient; S2, uploading the compensated obstacle information and sea state parameters output by step S1 to the server end, introducing a sea state correction factor into the server end building a CPA model, calculating a collision risk index through the CPA model, and predicting the path conflict probability between the ship pairs in the water area according to a path conflict probability calculation formula determining the collision risk result according to the collision risk index and the path conflict probability; S3, determining the path conflict of the ship pairs in the water area according to the collision risk result output by step S2, assigning a specific identity to the ship pairs involved in the conflict by the server end, planning a temporary virtual obstacle area according to the specific identity, generating a cooperative obstacle avoidance instruction according to the temporary virtual obstacle area, and delivering the instruction to the local end of the corresponding ship; S4, receiving the cooperative obstacle avoidance instruction delivered by the server end in step S3, determining the constraint condition according to the sea state parameters output by step S1 and the collision risk result output by step S2, planning the local path by using an improved Dijkstra algorithm, and outputting the planning result; S5, driving the ship to sail according to the planning result of step S4, and repeatedly executing steps S1 to S4 every 0.5s until the ship reaches the target position.

[0007] In the preferred scheme, in step S2, the path conflict probability is calculated according to the following formula: ; wherein, is the path conflict probability between multiple ships, and the value range is [0, 1]; t0 is the current time, which is the starting time point of path conflict prediction; T is a prediction time window, which is used to limit the prediction length of path conflict; For the current ship at time t, position The probability density function at that location; For ships that conflict with the current ship at time t, position The probability density function at that location; These are the planar coordinates within the water area; It is a tiny unit of length in the x-direction of a Cartesian coordinate system; It is a tiny length unit in the y-direction; The product of these can be represented as a tiny area unit on a plane; It is a tiny unit of time t.

[0008] In the preferred embodiment, the collision risk indicators in step S2 include the time to nearest encounter. Distance from the nearest meeting ; Sea state correction factor The compensated sea state parameters output from step S1 are used to calculate and correct the nearest encounter time in the CPA model. Distance from the nearest meeting ; Sea state correction factor The calculation formula is: ; The revised collision risk index is as follows: ; ; in, This is a sea state correction factor used to correct the base CPA index; Real-time water flow velocity collected by a multi-sensor group; This refers to the ship's maximum speed. Real-time wave height collected by a multi-sensor group; Based on the most recent meeting time; This is the revised nearest meeting time; Based on the nearest encounter distance; This is the corrected nearest encounter distance; In the preferred embodiment, in step S3, the server determines the path conflict of the vessel pairs in the water area based on the collision risk results output in step S2. The path conflict probability calculated in step S2 To determine the opposing vessels in the conflict; like If a conflict is identified between vessels in the waterway, the server assigns specific identities of "give way vessel" and "straight course vessel" based on the vessel's mission type and relative course. A temporary virtual obstacle zone is planned for the "give way vessel," which is a pre-defined safety zone around the "straight course vessel's" route. The collaborative cost of all nodes within the temporary virtual obstacle zone is then calculated. Set to the maximum value to generate a result with "identity assignment results, temporary virtual obstacle zone coordinates, and..." The collaborative obstacle avoidance command, which retrieves the value, is sent to the local terminal of the corresponding vessel. The temporary virtual obstacle zone extends to both sides of the current route of the "straight-line ship". A rectangular area is formed, which serves as a temporary virtual obstacle area. (Expand) The formula is: ; The priority rule is: disabled boat > fishing boat > sightseeing boat, and boats with lower priority are "give way" boats; like If it is determined that there is no conflict between vessels in the waterway, the server sends a "no coordination adjustment" command to the local end of the corresponding vessel. At this time, the coordination cost item... .

[0009] In the preferred embodiment, the planning result in step S4 includes the coordinates of the waypoints. speed and steering angle ; The cost function of the improved Dijkstra algorithm is as follows: ; in, This is an estimate of the total cost from the starting point through node n to the target point, used to select the optimal path; This represents the actual travel cost from the starting point to node n. The heuristic cost from node n to the target point; The collaborative cost item issued by the server; This refers to the risk cost term based on a dynamic risk field. and These are the weighting coefficients; Specifically, when the command issued by the server is a "cooperative obstacle avoidance command", then When the instruction is "no coordinated adjustment", then ; When the collision risk result marker in step S2 indicates the presence of obstacles, then When there are no obstacles, then .

[0010] In the preferred embodiment, the local end generates a thruster PWM control signal and a servo angle control signal based on the planning result of step S4, and drives the ship to sail. The local control method uses a PID control algorithm, and the PID parameter self-tuning formula is as follows:

[0011]

[0012]

[0013] The formula for control command correction is:

[0014] in, , , These are the proportional, integral, and derivative coefficients of the PID after self-tuning; , , These are the initial PID parameters; The sea state coefficient is the output of step S1; To control the amount of instruction correction; This is the deviation amount; The integral of the deviation; This is the differential of the deviation.

[0015] In a preferred embodiment, the multi-sensor group in step S1 includes a millimeter-wave radar, a high-definition camera, a sonar sensor, a sea state sensor, and a shipboard functional module. Millimeter-wave radar is deployed at the bow to collect obstacle data within a range of 50-100m; high-definition cameras are deployed at the stern to collect visual data from 5-1500m; sonar sensors are deployed on both sides of the hull to collect underwater data from 1-5m; and sea state sensors are deployed amidships to collect wind speed data. High waves and water flow velocity ; Vessel mission types include sightseeing boats, fishing boats, and boats with malfunctions; The shipboard functional modules include a task selection module and a fault identification module. The task selection module is used to receive the "route tour" or "fishing recreation" mode signal selected by the user and mark the corresponding task type. The fault identification module is used to monitor the operating condition of the power / navigation system. If a fault is detected, it will automatically mark the mission type as "faulty ship".

[0016] In the preferred embodiment, when the fault identification module marks the vessel as a "faulty vessel" in step S1, step S4 further includes a temporary fault path adjustment, specifically: S41. Calculate the coordinates of the temporary target point:

[0017]

[0018] S42. Using the temporary target point as a transitional target, substitute it into the cost function of the improved Dijkstra algorithm to plan a temporary path; in, , The planar coordinates of the temporary target point of the fault; , The coordinates are the current position of the malfunctioning vessel. The safe speed for the malfunctioning vessel; The heading angle furthest from the center of the waterway; heading angle away from the center of the channel The calculation formula is: +90°; in, The heading angle at the center of the channel; A temporary path is a path that terminates at a temporary target point, replacing the original target point.

[0019] In the preferred embodiment, step S4, after planning the local path, also includes path verification. Path verification includes turning angle verification and speed fluctuation verification, and the verification formula is: Steering angle verification:

[0020] Speed ​​fluctuation verification:

[0021] in, , They are respectively the i-th segment and the i-th segment. The turning angle of the path segment; This refers to the maximum angular velocity of the ship's steering gear; This represents the estimated time difference between the two routes; , They are respectively the i-th segment and the i-th segment. Speed ​​along the route; The maximum permissible fluctuation value of speed is set based on the responsiveness of the ship's propulsion system; If the planned local path meets the above two verification conditions, the path is determined to be valid; otherwise, step S4 is re-executed to adjust the cost function weight coefficients. and Continue until the path verification is successful.

[0022] In the preferred embodiment, the specific steps of the multi-sensor distortion compensation algorithm in step S1 are as follows: S11. The raw data from millimeter-wave radar, high-definition camera, and sonar sensor are time-stamped and synchronized using the PTP protocol to obtain synchronized data; S12. Based on the wind speed from the sea state sensor. High waves Calculate the sea state influence coefficient The formula is: ; in, ∈[1,3]; S13. Import the synchronized data. The Kalman filter correction is applied to construct the state equation and observation equation, as shown in the following formula: Equations of state: ; Observation equation: ; in, Let A be the state of the obstacle at time k, A be the state transition matrix, and B be the control matrix. To control the quantity, For process noise, Let H be the observation value at time k, and H be the observation matrix. To observe noise; S14. Output the compensated obstacle state according to the Kalman gain and state correction formula. and sea state parameters, including wind speed High waves and water flow velocity .

[0023] This invention provides a ship-based intelligent obstacle avoidance and path planning method. Through the coordination of the aforementioned structures, it offers the following advantages compared to existing methods: First, by subdividing and compensating for the offset of millimeter-wave radar point cloud and introducing Kalman filtering with sea state coefficient, the detection error of static obstacle position can be effectively reduced. In addition, the accuracy of dynamic obstacle identification is effectively improved, the obstacle detection coverage of underwater reefs is increased, and the problem of multi-sensor data distortion under nearshore dynamic sea conditions can be effectively solved, avoiding missed detection and false detection by a single sensor. Second, by introducing a sea state correction factor to correct the CPA index and adding a multi-obstacle comprehensive risk calculation, the deviation between the collision risk index and the actual situation is effectively reduced, and the rate of missed judgment and misjudgment in multi-obstacle scenarios is also reduced, ensuring accurate collision risk judgment under complex nearshore sea conditions. Third, by constructing scenario-based priority weights and adding a dynamic adjustment mechanism, the conflict rate of multiple vessels avoiding each other is effectively reduced, while also increasing the avoidance response time and solving the problem of coordination chaos when small recreational boats and malfunctioning boats meet. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of Embodiment 1 of the present invention; Figure 2 This is a line graph comparing the model performance of Embodiment 1 in this invention; Figure 3 This is a scatter plot of the sea state coefficients in Embodiment 1 of this invention; Figure 4 This is a flowchart of the steps in Embodiment 2 of the present invention; Figure 5 This is a path planning visualization diagram of Embodiment 2 in this invention; Figure 6 This is a fault handling flowchart of Embodiment 2 of the present invention. Detailed Implementation

[0025] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features in the embodiments of this application can be combined with each other without conflict. The exemplary embodiments disclosed in this application will be described below with reference to the accompanying drawings, including specific technical details disclosed in this embodiment to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures (such as the basic driving principle of ship propulsion, the basic data acquisition process of general sensors, etc.) are omitted in the following description.

[0026] In the field of intelligent control technology for surface vessels, autonomous obstacle avoidance and path planning technology for small recreational boats (such as fishing boats and sightseeing boats in scenic areas) is widely used in nearshore recreational water operations. The core requirement is to achieve safe and precise autonomous navigation in scenarios involving dynamic sea conditions (gusts, swells), multiple vessel intersections, and static obstacles (aquaculture buoys, reefs). In existing technologies, collecting environmental information through multiple sensors (radar, cameras, sonar) and combining it with path planning algorithms (such as Dijkstra's algorithm) has become the basic solution for obstacle avoidance on surface vessels. The introduction of pre-trained models can further improve the accuracy of obstacle recognition and collision risk assessment, reduce the amount of on-site training data, and adapt to the dynamic changes of complex nearshore environments.

[0027] With the development of the water-based leisure industry, obstacle avoidance technology for small leisure boats has gradually expanded from "single-sensor static obstacle avoidance" to "multi-sensor fusion dynamic obstacle avoidance" to adapt to the characteristics of nearshore scenic areas, such as "dense convergence of multiple boats, dynamic changes in sea conditions, and the need to stop at fishing spots." Multi-sensor fusion obstacle avoidance solutions integrate information from radar (long-range obstacle detection), cameras (visual recognition), sonar (underwater obstacle detection), and sea state sensors (wind speed and wave height collection) in a unified coordinate system, typically achieving a more comprehensive environmental perception than traditional single-sensor solutions. In the actual operation of small leisure boats, obstacle avoidance systems not only need to achieve basic obstacle avoidance but also need to meet scenario-specific requirements such as "precise stopping at fishing spots," "coordinated obstacle avoidance among multiple boats," and "emergency docking in case of malfunction." These requirements often rely on adaptation to dynamic sea conditions and response to external collaborative commands. However, common obstacle avoidance solutions lack distortion compensation for dynamic nearshore sea conditions, fail to incorporate sea state correction factors in collision risk assessment, and lack multi-boat collaborative priority rules and emergency path planning specific to leisure boat scenarios, thus limiting their applicability in nearshore scenic areas.

[0028] When implementing obstacle avoidance for surface vessels, related technologies often fall into several categories. Firstly, in terms of sensor data processing, they are largely limited to general multi-sensor fusion, lacking specific distortion compensation mechanisms for dynamic near-shore sea conditions (such as radar point cloud shifts caused by gusts and visual blurring caused by surges). Secondly, in terms of collision risk assessment, they rely heavily on basic CPA (Closest Point of Encounter) models without incorporating sea state parameters to correct risk indicators. Thirdly, in terms of multi-vehicle collaboration, they largely follow general maritime rules (such as COLREGS) without designing scenario-based obstacle avoidance priorities tailored to the specific needs of recreational vessels (fishing boats, sightseeing boats). Fourthly, in terms of fault response, they often lack emergency path planning and anchoring control logic. Therefore, the obstacle avoidance methods in these technologies have certain limitations and are ill-suited to the autonomous obstacle avoidance needs of small recreational vessels in dynamic near-shore sea conditions, multi-vehicle intersections, and scenario-based operations, failing to fully guarantee their navigation safety and operational applicability.

[0029] Example 1 This embodiment is based on the unmanned surface vessel platform of Qingdao Wujiang Technology Co., Ltd. It constructs an integrated pre-trained model of "obstacle recognition and collision risk prediction" for the complex marine environment of Qingdao's coastal area. This model will solve the problems of "multi-sensor data asynchrony" and "high risk misjudgment rate under complex sea conditions" in the nearshore environment of Qingdao, and support the autonomous cruise and obstacle avoidance needs of unmanned surface vessels in the Guzhenkou sea area and surrounding areas.

[0030] like Figures 1-3As shown, the experimental platform and hardware configuration utilize an amphibious unmanned surface vessel (USV) developed by Qingdao Wujiang Technology Co., Ltd. The USV is 12 meters long, 4.14 meters wide, and 1.65 meters deep, with a design draft of 0.55 meters. It employs an aluminum alloy hull and pump-jet propulsion technology. The sensor system carried in the experiment is shown in Table 1 below. Table 1

[0031] In this embodiment, data collection was conducted in the sea area of ​​Guzhenkou, Qingdao (N35°53'-N35°55', E120°07'-E120°10'), which is the main testing area of ​​Qingdao Wujiang Technology Co., Ltd. The data collection period covered different seasons, day and night conditions, and sea state conditions. Spatial range: covering nearshore shallow waters (2-10m deep), island and reef areas (including the area around Zhaidang Island) and open waters (10-30m deep). Time distribution: March to August 2024, with a total of 60 days of data collection, including: (1) Time period distribution: morning (6:00-8:00), daytime (10:00-16:00), nighttime (19:00-21:00); (2) Seasonal distribution: Spring (March-May), Summer (June-August); The sea conditions are as follows: (1) Calm sea conditions: wind speed < 3 m / s, wave height < 0.5 m (accounting for 30%). (2) Moderate sea state: wind speed 3-6 m / s, wave height 0.5-1.5 m (50%). (3) Complex sea conditions: wind speed 6-8 m / s, wave height 1.5-2.0 m (accounting for 20%). Obstacle types: (1) Static obstacles: reefs, breakwaters, aquaculture cages, buoys (0.5-2m in diameter); (2) Dynamic obstacles: fishing boats (5-15m), sightseeing boats (8-20m), motorboats (2-5m); (3) Special obstacles: floating objects, kelp farming areas, and sudden fishing nets.

[0032] In this embodiment, the specific data acquisition process is as follows: Preparation before collection: (1) Equipment calibration: Before daily data collection, the radar and camera are calibrated using standard targets, and the sonar is calibrated in areas with known water depths; (2) Parameter configuration: The sensor sampling frequency is preset according to the weather forecast (increased to 1.5 times in complex sea conditions); (3) Safety inspection: Inspect the status of the unmanned surface vessel's pump-jet propulsion system and tracked movement mechanism; Dynamic data acquisition: (1) Autonomous cruise mode: Set the unmanned vessel to cruise along the preset route at a speed of 5-10 knots, and change the route every 30 minutes; (2) Specific scenario trigger: When the sea state sensor detects a wave height > 1.5m, the high dynamic acquisition mode is automatically started; (3) Manual intervention: Remotely control the unmanned surface vessel to approach specific obstacles for focused data collection through the shore-based control center; Data storage management: (1) Local caching: The unmanned surface vessel is equipped with a 2TB SSD for real-time data caching and uses RAID1 for backup; (2) Remote transmission: Real-time transmission of critical data via 4G / 5G network, with edge computing preprocessing enabled when bandwidth is insufficient; (3) Data archiving: After the daily collection is completed, the data is archived to the data center of Qingdao Wujiang Technology Co., Ltd. through the NAS system, using a three-level directory structure of "date, sea condition, and scenario"; In the optimal approach, a total of 3000 sets of valid data were collected and screened during this training, and their specific distribution is shown in Table 2 below: Table 2

[0033] In this embodiment, for time synchronization and spatiotemporal alignment, the raw data from each sensor is processed, including radar point cloud timestamps. Image frame timestamp Sonar waveform timestamp The reference time is calibrated using the UTC time of the BeiDou / GPS module. Using this as a baseline, calculate the time deviation of each sensor: , , ; In this embodiment, data with a time deviation ≤10ms is directly marked as synchronized and valid. For data with a time deviation > 10ms, linear interpolation is used to correct the position information: , ; In this embodiment, the steps of spatial coordinate transformation are as follows: Establish a unified coordinate system with the center of mass of the unmanned surface vessel as the origin, the bow direction as the X-axis, and the starboard side as the Y-axis. According to the installation offset ( , Perform coordinate correction; The data is converted into three-dimensional coordinates based on the draft and installation angle. The data is multi-source data with unified timestamps and uses the USV-DataV2.0 format developed by Qingdao Wujiang Technology Co., Ltd., which includes fields such as sensor ID, timestamp, three-dimensional coordinates, and confidence level.

[0034] In this embodiment, as Figure 3 As shown, the sea state coefficient labeling is based on measured data from the Qingdao coastline, optimizing the sea state coefficient calculation model: input data collected by the sea state sensor. (Wind speed, m / s) (Wave height, m) (Water flow velocity, m / s) is processed and calculated using formulas. and : Qingdao sea area optimization coefficient: ; Considering the impact of nearshore currents: ; Output data annotations for each frame and Tags, for example: Calm sea conditions ( , ): , ; Complex sea conditions , ): , ; In this embodiment, the synchronized obstacle position is input when generating the collision risk label. ,speed The condition of this vessel ; The formula for calculating the basic CPA parameters is as follows: ; ; The formula for calculating sea state correction is as follows: ; ; The formula for calculating the probability of path conflict is as follows: Integral calculation ; in, The probability density function of the ship's position Considering the large number of aquaculture areas near Qingdao, a specific correction factor for aquaculture cages can be added when the probability density function of obstacle locations is detected. ; Output tag set For example: Encounter with a fishing boat: , , ; Approaching the breeding area: , , (including) correction); In this embodiment, the specific details regarding dataset partitioning and augmented dataset partitioning are as follows: Training set: 2100 groups (70%), distributed according to scene type. Validation set: 600 groups (20%), focusing on moderate sea state scenarios; Test set: 300 groups (10%), including new obstacle types not included in the training (such as aquaculture cages, special buoys); Specifically, the data augmentation strategies are as follows: Sea state disturbance: Gaussian noise is added to calm sea state data to simulate complex sea states, with noise intensity and... Positive correlation; Viewpoint transformation: Rotate the image data by ±15° and scale it by 0.8-1.2 times to simulate the attitude changes of the unmanned surface vessel; Temporal interpolation: performing frame interpolation on dynamic sequences to expand the temporal dimension of the samples; Occlusion simulation: Randomly add 10-20% occlusion area to simulate wave splash and water mist interference.

[0035] In this embodiment, a professional annotation tool is selected based on the data type, as shown in Table 3 below: Table 3

[0036] In this embodiment, the "Qingdao Offshore Obstacle Labeling Specification v1.0" was formulated before labeling, defining a labeling system for 12 categories of obstacles; In this embodiment, three engineers with experience operating in the Qingdao sea area participated, and they were only allowed to take up their posts after passing the test set assessment. In this embodiment, the annotation is implemented as follows: For static obstacle annotation, LabelImg is used to mark the target rectangle and category, and PointLabeler is used to annotate the 3D bounding box; For dynamic obstacle labeling, LabelMe is used for video sequence labeling to ensure target ID consistency. Keyframes are manually labeled every 10 frames, and intermediate frames are automatically interpolated. For multi-source fusion annotation, the self-developed USV-Annotate tool is used to achieve joint annotation and verification of image, point cloud, and sonar data.

[0037] The quality control measures are as follows: cross-validation adopts a "dual annotation combined with third-party review" model, with annotation consistency ≥98%; 10% of the annotation data is randomly sampled daily for quality checks; an annotation error feedback mechanism is established, and annotation specifications are updated regularly. In this embodiment, a multi-sensor fusion feature extraction network is designed to address the characteristics of foggy and high-humidity environments near Qingdao's coast. For the radar branch, an improved PointNet++ network is used, with 512×512 point cloud data as input. The sampling layer uses FPS (farthest point sampling) to retain key feature points. Feature extraction uses three SA (SetAbstraction) modules, with the output dimension gradually increasing (64→128→256). A salt spray noise filtering layer is also added, using median filtering to process radar noise.

[0038] The visual branch uses an improved YOLOv5 CSPDarknet53 backbone, inputting fused 1920×1080 RGB and infrared images, and adds an attention mechanism module to enhance feature extraction in foggy conditions. Dynamically adjust the kernel size, and 5x5 large cores are activated at this time.

[0039] The sonar branch uses a 3-layer 1D-CNN network, inputting 128×128 sonar waveform data, and changing the convolutional layers from 16→32→64 channels, using the ReLU activation function, and employing max pooling to preserve underwater obstacles.

[0040] Employing a weighted attention fusion mechanism: ; The weighting coefficients are dynamically adjusted. Calm sea conditions ( ): , , ; Complex sea conditions ): , , ; The output layer's classification head predicts obstacle types (12 categories), uses Softmax activation, and predicts 3D coordinates. ,speed ,size Finally, the target detection confidence is predicted.

[0041] Loss function:

[0042] in,, , ,

[0043] In this embodiment, the temporal feature modeling input layer of the collision risk prediction module is specifically as follows: The obstacle recognition module outputs a state sequence (10 frames, each frame contains...) ) and sea state parameters ; Spatial feature extraction: A two-layer fully connected network is used to process sea state parameters. Output 32-dimensional sea state features; Temporal feature capture uses a BiLSTM network, which has 2 layers and 128 hidden dimensions; the input is a 10×4 obstacle state sequence, and Dropout (rate=0.3) is added after each layer to output a 256-dimensional temporal feature vector. In this embodiment, the concat sea state features and time series features are compressed using a single fully connected network: ; In this embodiment, the risk prediction output layer is designed as follows: Regression Subnetwork 1: Prediction and ; Layer 2 fully connected: ; Regression Subnetwork 2: Prediction ; Layer 2 fully connected: ; Loss function: ; The weighting for losses in high-risk nearshore areas of Qingdao (such as aquaculture areas) was increased by 1.5 times. In this embodiment, the training hardware environment uses the training platform of the AI ​​Lab of Qingdao Wujiang Technology Co., Ltd. Table 4

[0044] The operating system was Ubuntu 20.04LTS, the deep learning framework was PyTorch 1.13.1, the CUDA version was 11.7, the parallel computing was implemented using DistributedDataParallel for multi-GPU training, the monitoring tools were TensorBoard 2.11.0 and Weights & Biases, and the data processing was performed using OpenCV 4.6.0 and PyTorchGeometric 2.2.0.

[0045] A two-stage training strategy is adopted, combined with the adaptive learning rate scheduling algorithm developed by Qingdao Wujiang Technology Co., Ltd.: Phase 1: Obstacle recognition module pre-training (1-60 epochs); Optimizer: AdamW; Initial learning rate: 0.001; Learning rate scheduling: cosine annealing =60; Batch size: 32 (single card) × 4 (number of cards) = 128; Weight decay: 0.0005; Early stopping strategy: Stop if the validation set accuracy does not improve for 5 consecutive epochs; Phase 2: Dual-module joint training (61-150 epochs); Optimizer: AdamW; Initial learning rate: 0.0005; Learning rate scheduling: cosine annealing =90; Batch size: 16 (single card) × 4 (number of cards) = 64; Weight decay: 0.0001; Regularization: Dropout (rate=0.3) and L2 regularization; Early stopping strategy: If the total loss does not decrease for 8 consecutive epochs, stop the process. Model validation was conducted in the sea area near Guzhenkou in Qingdao, where no training was conducted. The results are shown in Table 5 below: Table 5

[0046] In this embodiment, as shown Figure 2 Based on the comparison between the existing algorithms of Qingdao Wujiang Technology Co., Ltd. and mainstream industry solutions shown in Table 6 below, the following results are obtained: Table 6

[0047] In this embodiment, the adaptability analysis of the Qingdao sea area is as follows: Environmental adaptability: In foggy environments, the fusion of infrared and visible light improves recognition accuracy by 12.3% compared to a pure vision solution; specific optimizations for aquaculture cages in the Qingdao nearshore area further enhance this capability. Prediction error was reduced by 40%; optimized radar signal processing ensured a 97.1% performance retention rate under complex sea conditions. Computational efficiency: The lightweight design of the model keeps the weight file to 185MB, meeting the storage limitations of the "Sea Lizard" unmanned surface vessel; the processing time of 32ms meets the real-time requirements of the unmanned surface vessel's high-speed navigation (up to 50 knots). Engineering value realization: The model has been deployed and verified on a series of unmanned surface vessels of Qingdao Wujiang Technology Co., Ltd. Compared with the existing solution, the accuracy of obstacle avoidance decision-making has been improved by 15% and the false alarm rate has been reduced by 60%. It has passed the near-shore autonomous navigation test certification organized by Qingdao Maritime Safety Administration.

[0048] The output of the pre-trained model (weight file: obstacle_risk_model_v2.1.pth) trained in this embodiment is: Obstacle types and 3D states: ; Collision risk parameters: ; Sea state assessment results: ; Example 2 The intelligent obstacle avoidance and path planning method based on ships provided by this invention is described below. During the experiment, the pre-trained model from Example 1 was used to analyze the intelligent obstacle avoidance and path planning method based on ships provided in this embodiment. Simultaneously, the analysis results of this scheme are compared with the judgment results of traditional methods to comprehensively evaluate the effectiveness and reliability of the scheme, and further explained in conjunction with Example 1. like Figure 4 The structure shown is... Figure 4 The intelligent obstacle avoidance and path planning method based on ships provided in this application includes the following steps in applying a pre-trained model to obstacle avoidance on fishing boats at Qingdao Silver Beach Scenic Area: This embodiment utilizes a small recreational unmanned fishing boat developed by Qingdao Wujiang Technology Co., Ltd. to conduct obstacle avoidance applications in the waters of Qingdao Yinshatan Scenic Area (N35°58′N36°00′, E120°15′E120°17′).

[0049] In this embodiment, the nearshore water depth is 28m, with an average of 1015 fishing boats and 58 sightseeing boats operating daily. There are static obstacles such as aquaculture buoys (0.8-1.2m in diameter) and underwater reefs (35m in depth). In summer, local gusts (wind speed 46m / s) and swells (wave height 0.6-1.2m) often occur. It is necessary to solve the deficiencies of "avoiding multiple boats passing each other", "adapting to dynamic sea conditions near the shore", and "ensuring the safety of fishing boats staying at fishing spots".

[0050] In this embodiment, the model, deployment, and parameters of the core hardware system are shown in Table 7 below: Table 7

[0051] In this embodiment, initial parameter configuration is performed first, based on the actual test scenario on May 15, 2024. Specifically: The current status of this vessel is designated as vessel A, a fishing boat: Initial position: This location represents a partial coordinate system for the Qingdao Silver Beach Scenic Area. Pointing to the Qingdao Silver Beach Scenic Area Wharf (N35°59′30″, E120°16′00″). Pointing to the open sea (N35°59′00″, E120°16′30″), coordinate transformation formula (WGS84 to local coordinates): ; ; Where B and L represent the ship's WGS84 latitude / longitude, and x and y are in meters (conversion error ≤ 0.1m).

[0052] The ship's initial speed: ,course This is the direction of the target fishing spot; Task type: Fishing boat; "Fishing Mode" can be selected via the onboard touchscreen; Task weight... ; Target fishing spot: The water depth was 4.2m as detected by sonar, with no underwater obstacles.

[0053] Obstacles in the surrounding environment were measured: Ship B (i.e., the tour boat): Location speed ,course Task type = sightseeing boat ( ); Static Obstacle C (i.e., aquaculture buoy): Location 1.0m in diameter, safety distance Risk coefficient ; Sea state parameters (real-time data collected by sea state sensors): Wind speed High waves water flow speed Initial sea state coefficient .

[0054] In this embodiment, the configuration settings of the pre-trained model are as follows: Model file: `obstacle_risk_model_v2.1.pth` trained in Example 1, deployed on a shore-based server, with TensorRT acceleration used for the inference framework; Model input: Synchronized data from multiple sensors uploaded from the ship's end, namely radar point cloud, image, and sonar; output: obstacle status (i.e., position, velocity, and type), collision risk parameters (i.e., collision risk parameters). , , ).

[0055] Step S1: In this embodiment, input the raw data of each sensor (i.e., radar point cloud, image frame, sonar waveform) and the BeiDou time reference. S11, Equipment Calibration: S111, Radar Calibration: Before setting sail each day, calibrate the radar at the scenic area's pier (N35°59′12″, E120°16′00″) using a standard metal target, and adjust the radar gain until the ranging error is less than 0.3m; S112, Camera Calibration: White balance and distortion correction are performed using fixed reference objects in the scenic area to ensure the accuracy of visual recognition; S113. Sonar calibration: Calibrate in an area with a known water depth of 5m (N35°59′30″, E120°16′30″) to make the depth measurement error less than or equal to 0.2m.

[0056] S12, Time Synchronization Execution: S121. The PTPv2 protocol is adopted, and the UTC time output by the Beidou module is used as the reference. All sensors on the ship end are connected to the PTP switch via Ethernet. The actual measured synchronization error is 7.2ms less than or equal to 10ms. S122. Synchronized data format: Encapsulated according to the Qingdao Wujiang `USVDataV2.0` protocol, each frame of data includes: Header: Sensor ID (1 byte), timestamp (8 bytes, UTC milliseconds), data length (4 bytes). Data segments: Radar point cloud (x / y / z coordinates, 4 bytes per point × 3), camera image (JPEG compression, 1920×1080 resolution), sonar waveform (16-bit integer, 384 sampling points); Verification: CRC32 (4 bytes); Data transmission: Data is uploaded to the shore-based server (IP: 192.168.1.100, port: 5000) via a 5G industrial module (Huawei ME909s-821), using the TCP transmission protocol, with a measured latency of 42ms.

[0057] S13, Data Preprocessing: S131, Radar point cloud denoising: Statistical filtering is used to remove salt spray interference points while retaining the effective number of point clouds; S132, Image Dehazing: For morning fog scenes in scenic areas, the dark channel prior algorithm is used to process camera images, and the contrast is improved after dehazing; S133, Sonar data denoising: Remove water flow noise through wavelet transform while preserving echo signals from underwater obstacles.

[0058] The output consists of synchronized multi-source data (with timestamps uniformly set to `2024051509:30:00.000`), with dual backups on the ship's local cache (16GB SSD) and the shore-based server.

[0059] S14. Calculation of Sea State Coefficient and Kalman Filter Compensation: Input synchronized data and sea state parameters ( , , ); S141. Calculation of Sea State Coefficient: According to the Qingdao Wujiang optimization formula: ,Right now It meets the requirements; ; Right now This is the ship's maximum speed.

[0060] S142, Kalman filter execution: State vector definition: This refers to the position and velocity of the obstacle; State transition matrix, i.e. The optimal sampling period measured by Qingdao Wujiang is: ; The process noise variance matrix was calibrated based on 500 sets of obstacle motion data from the Qingdao nearshore area. ; Observation equations, introducing Correction for observation noise: ; in It is a 4×4 identity matrix. Variance matrix: ; Revised ; Filtering update, taking ship B as an example: Observations after synchronization Initial state estimation ; Initial covariance matrix: ; Kalman gain: ; Status Update: .

[0061] Output: Obstacle information after compensation: Ship B status The speed direction is 120°, consistent with the initial heading; Buoy C state Static obstacle, speed is 0; Sea state parameters: , , ; Task type: This ship = fishing boat, Ship B = sightseeing boat. Confirm and upload through the ship's "Task Selection Module".

[0062] In this embodiment, in step S2, the server-side collision risk calculation is performed as follows: S21. Pre-trained model inference and CPA correction: Input S1 outputs the compensated data and the pre-trained model; S22, Model Inference Execution: After receiving data from the ship, the shore-based server calls the `TensorRT` engine to load the model. The input dimensions are (1, 512, 512, 3), which are radar point clouds and image features. The output layer is parsed. Obstacle type identification: Ship B = dynamic tour boat, Buoy C = static aquaculture buoy, with identification confidence levels of 0.98 and 0.99 respectively; Preliminary output of collision risk parameters: basic , , .

[0063] S23, Sea State Correction: Correction ; Correction ; Path conflict probability correction, i.e. optimization of the Qingdao nearshore aquaculture area, is achieved by adding an aquaculture area correction factor due to the presence of buoy C. Based on statistics of 100 collision cases in the aquaculture area of ​​the scenic area, the final .

[0064] S24. Risk Assessment Logic: Judgment threshold: and and If it is "high risk", then it is "low risk"; Actual test results: , , →Low collision risk, no need for emergency obstacle avoidance, but path avoidance of buoy C is required.

[0065] The collision risk result for the output vessel is low risk; the collision risk between vessel B and this vessel is low. , , Centered on the risk area of ​​buoy C, with a radius of... It is marked as "impassable" on the raster map.

[0066] In this embodiment, step S3, generation of server-side collaborative obstacle avoidance instructions: S31, Input S2 Output The mission types and statuses of the two ships; Priority Calculation: Priority of this vessel ; in, , As the safety threshold for the scenic area, therefore ; Ship B Priority ; Priority determination: →This vessel is the "vehicle being avoided," and Vessel B is the "potential give-way vessel," but because They have been identified as low-risk and will not be assigned official status for the time being.

[0067] S32, Collaboration Cost Item generate: Instruction type: "No coordinated adjustment, path avoidance of static obstacles"; Setting: Nodes within a 10.49m radius of buoy C Normal cost, non-maximum value, other nodes Weighting coefficient This means that the coordination instruction is invalid and no coordination cost needs to be added.

[0068] S33, Instruction Issuance: Instruction format: JSON format, including "risk level, The value acquisition and obstacle avoidance constraints (avoidance buoy C) were transmitted to the ship via the 5G private network, and the actual transmission latency was 42ms≤50ms. Command format (JSON, parsable): `{"risk_level":"low","C_c":[{"area":"(3050,2550,10.49)","value":50},{"other":"0"}],"constraint":"avoid_buoy_C"}`.

[0069] The output is a cooperative obstacle avoidance command, meaning there is no cooperative adjustment; avoid buoy C. Value table, weighting coefficients .

[0070] In this embodiment, as shown Figure 5As shown, step S4, local path planning: S41. Dynamic Risk Field and Heuristic Cost Calculation: Input S1 The buoy C parameter of S2, and the buoy C parameter of S3. ; S411, Dynamic Risk Field Calculation (claim 3.3): Select path node (Near buoy C), calculate: Euclidean distance from node n to buoy C ; Substitute into the formula This indicates a long distance and low risk. Wherein, the natural constant e≈2.71828; Select Node That is, the edge of buoy C. , (The risk is slightly higher).

[0071] S412, Heuristic Cost calculate: Manhattan distance from node n to the target fishing point: ; Substitute into the formula .

[0072] S413. Cost function construction: Actual cost That is, from node n to the starting point The Euclidean distance is: ; Total cost, i.e. , Due to the presence of static obstacles: ; The output is the value of node n. The nodes surrounding buoy C Value table, heuristic cost Calculation results.

[0073] S42. Improved Dijkstra's algorithm for planning and path verification: Input cost function, raster map parameters, and ship performance constraints (maximum turning angular velocity) Maximum speed fluctuation ; S421, Grid map initialization: The grid size is 0.5m×0.5m, which meets the high-precision requirements of the scenic area near the shore. The coverage area is x∈[2950m,3150m], y∈[2450m,2650m], with a total of 800×800=640000 grids. The impassable grid markings were created using the Bresenham algorithm, which drew a 10.49m circular area around buoy C and marked it as "impassable," consisting of approximately 440 grids.

[0074] S422, Algorithm Execution Steps: starting point: ,end: ; Priority queue initialization: starting point The remaining nodes ; Path search: Iterative selection Find the smallest node, update the cost of its adjacent nodes, and continue until the destination is reached; Actual planning results: like Figure 5 As shown, the sequence of waypoints is: There are a total of 6 waypoints, with a total voyage of approximately 282.84m.

[0075] S423, Speed ​​and Turn Calculations: Speed ​​constraints: ; Steering angle: Steering angle between adjacent waypoints Measured maximum steering angle (2nd → 3rd waypoints).

[0076] S424, Path Verification: Steering angle verification: ,Right now The travel time to the points along the way is satisfied; Speed ​​fluctuation verification: ,satisfy; Verification result: The path is valid and can be executed.

[0077] Output the valid planned path, including the coordinates of the waypoints and the speed. The steering angle sequence and path verification report meet all constraints.

[0078] In this embodiment, as Figure 6 As shown, step S5, closed-loop control and emergency fault handling: S51, Sea State Adaptive PID Control: Input the planning results from S4, and receive real-time feedback from GPS / IMU (position, attitude); S511, PID parameter self-tuning (claim 8.2): Initial parameters: , , (Qingdao Wujiang Calm Sea State Calibration Value); Corrected calculation: ; ; .

[0079] S512, Control Command Generation and Modification: Real-time feedback data (during the journey from the 1st to the 2nd waypoint): Planning location Real-time location Positional deviation , Overall deviation ; Planning the course Real-time heading attitude deviation .

[0080] Deviation correction calculation (taking attitude deviation as an example): Integral term That is, the sampling period is 0.5s; Differential term ; The correction amount is: ; Final control command: The servo angle = 43.5° + 4.451° ≈ 47.95°, which is close to the planned angle. ); The thruster PWM duty cycle = (3.75 / 5) × 100% = 75%, corresponding to a speed of 3.75 m / s.

[0081] S513, Closed-loop execution: Control cycle: S1S4 repeats every 0.5s to update sensor data, risk parameters and paths, and correct control commands; Actual control performance: Position deviation is stable within ≤±0.5m, attitude deviation is ≤±1°, meeting the precise navigation requirements of fishing boats in scenic areas.

[0082] Outputs: Thruster PWM command (75% duty cycle), servo angle command (47.95°), real-time control deviation report.

[0083] S52, such as Figure 6 Emergency handling of the fault shown Fault simulation (essential for R&D testing): Triggering condition: During the journey from the 3rd to the 4th waypoint, the GPS module power is manually disconnected (simulating a GPS signal loss fault). The fault identification module detects "GPS signal not updated for 3 consecutive seconds" and automatically marks the vessel as a "faulty vessel". ).

[0084] Processing details: S521, Calculation of temporary target points (claim 9.1): Current location at the time of failure The center of the scenic waterway is heading (Pointing to the dock safety area); away from the direction of the shipping channel ; Safe speed (Default safe speed for a malfunctioning vessel); Temporary target point:

[0085] .

[0086] S522, Temporary Path Planning: Replace the original target point and re-execute the improved Dijkstra's algorithm of S4 to plan a temporary path: Total range ≈ 1.0m; Temporary speed lower than To ensure stability.

[0087] S523, Fault command issued: After receiving the fault signal, the shore-based server simultaneously notifies vessels within 500 meters (such as vessel B) to issue a avoidance instruction: "This vessel has a malfunction; please take evasive action." Updated to 100 (avoidance cost); This vessel will follow a temporary route and arrive within 5 seconds. They then activated the anchoring device and awaited rescue.

[0088] Outputs a temporary fault path, a safe docking instruction, and a notification for surrounding vessels to coordinate avoidance. The actual fault handling time is 4.8s≤5s.

[0089] In summary, the test environment and the comparison scheme are now compared, as shown in Table 8 below: Table 8

[0090] Industry solution: CN107748561B, sea state adapted path planning, without coordination and fault handling; Traditional approach: manual operation within the scenic area, with 5 certified captains and an average of 5 years of driving experience; Through user testing, 100 tourists aged 18-60 with no experience in operating unmanned boats were selected. Each person operated the boat to complete one "dock-fishing spot-return" mission (approximately 15 minutes in total). The operation satisfaction rate was 92.3%, and no collision accidents occurred, which met the safety operation requirements of Qingdao Silver Beach Scenic Area.

[0091] arrive Then, it automatically switches to "fishing spot mode" and the speed decreases. Make a circular motion with a radius of 8m around the fishing spot to ensure that the fishing line does not get tangled; Even in scenic areas with morning fog (visibility 50m) and gusts of wind (6m / s), it still maintains a recognition accuracy of over 90%, with a path deviation of ≤±0.8m.

[0092] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0093] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A ship-based intelligent obstacle avoidance and path planning method, characterized in that, Includes the following steps: S1. By acquiring information about the ship's own status and surrounding environment through a multi-sensor group deployed on the ship, a distortion compensation algorithm combining Kalman filtering and sea state coefficient is adopted to output the compensated obstacle information, sea state parameters and ship mission type. S2. The local end uploads the compensated obstacle information and sea state parameters output in step S1 to the server. The server then introduces a sea state correction factor. A CPA model is constructed, and a collision risk index is calculated using the CPA model. Based on the path conflict probability calculation formula, the probability of path conflict between pairs of vessels in the water is predicted. The collision risk result is determined based on the collision risk index and the path conflict probability; S3. Based on the collision risk results output in step S2, the server determines the path conflict of the pairs of vessels in the water. The server assigns a specific identity to the pairs of vessels involved in the conflict, plans a temporary virtual obstacle zone based on the specific identity, generates a cooperative obstacle avoidance command based on the temporary virtual obstacle zone, and sends it to the local end of the corresponding vessel. S4. The local end receives the cooperative obstacle avoidance command issued by the server in step S3, determines the constraints based on the sea state parameters output in step S1 and the collision risk results output in step S2, plans the local path using the improved Dijkstra algorithm, and outputs the planning results. S5. The local end drives the ship to sail according to the planning result of step S4, and repeats steps S1 to S4 every 0.5s until the ship reaches the target position.

2. The intelligent obstacle avoidance and path planning method based on ships according to claim 1, characterized in that, In step S2, the path conflict probability The calculation formula is: ； in, The probability of path conflict between multiple ships is denoted as t0, which ranges from 0 to 1; t0 is the current time, which serves as the starting point for path conflict prediction; T is the prediction time window, used to limit the prediction duration of path conflict. Given the current ship's position at time t. The probability density function at that location; For ships that conflict with the current ship at time t, position The probability density function at that location; These are the planar coordinates within the water area; It is a tiny unit of length in the x-direction of a Cartesian coordinate system; It is a tiny length unit in the y-direction; The product of these can be represented as a tiny area unit on a plane; It is a tiny unit of time t.

3. The intelligent obstacle avoidance and path planning method based on ships according to claim 1, characterized in that, In step S2, the collision risk indicators include the time of nearest encounter. Distance from the nearest meeting ; Sea state correction factor The compensated sea state parameters output from step S1 are used to calculate and correct the nearest encounter time in the CPA model. Distance from the nearest meeting ; Sea state correction factor The calculation formula is: ； The revised collision risk index is as follows: ； ； in, This is a sea state correction factor used to correct the base CPA index; Real-time water flow velocity collected by a multi-sensor group; This refers to the ship's maximum speed. Real-time wave height collected by a multi-sensor group; Based on the most recent meeting time; This is the revised nearest meeting time; Based on the nearest encounter distance; This is the corrected nearest encounter distance.

4. The intelligent obstacle avoidance and path planning method based on ships according to any one of claims 1 to 3, characterized in that, In step S3, the server determines the path conflict of the vessels in the water based on the collision risk results output in step S2. The path conflict probability calculated in step S2 To determine the opposing vessels in the conflict; like If a conflict is identified between vessels in the waterway, the server assigns specific identities of "give way vessel" and "straight course vessel" based on the vessel's mission type and relative heading. A temporary virtual obstacle zone is planned for the "give way vessel," which is a pre-defined safety zone around the "straight course vessel's" route. The collaborative cost of all nodes within the temporary virtual obstacle zone is then calculated. Set to the maximum value to generate results including "identity assignment results, temporary virtual obstacle zone coordinates, and..." The collaborative obstacle avoidance command, which retrieves the value, is sent to the local terminal of the corresponding vessel. The temporary virtual obstacle zone extends to both sides of the current route of the "straight-line ship". A rectangular area is formed, which serves as a temporary virtual obstacle area. (Expand) The formula is: ； The priority rule is: disabled boat > fishing boat > sightseeing boat, and boats with lower priority are "give way" boats; like If it is determined that there is no conflict between vessels in the waterway, the server sends a "no coordination adjustment" command to the local end of the corresponding vessel. At this time, the coordination cost item... .

5. The intelligent obstacle avoidance and path planning method based on ships according to claim 4, characterized in that, The planning results in step S4 include the coordinates of the waypoints. speed and steering angle ; The cost function of the improved Dijkstra algorithm is as follows: ； in, This is an estimate of the total cost from the starting point through node n to the target point, used to select the optimal path; This represents the actual travel cost from the starting point to node n. The heuristic cost from node n to the target point; The collaborative cost item issued by the server; This refers to the risk cost term based on a dynamic risk field. and These are the weighting coefficients; Specifically, when the command issued by the server is a "cooperative obstacle avoidance command", then When the instruction is "no coordinated adjustment", then ; When the collision risk result marker in step S2 indicates the presence of obstacles, then When there are no obstacles, then .

6. The intelligent obstacle avoidance and path planning method based on ships according to claim 5, characterized in that, In step S5, the local terminal generates a thruster PWM control signal and a servo angle control signal based on the planning results of step S4, and drives the ship to sail. The local control method uses a PID control algorithm, and the PID parameter self-tuning formula is as follows: The formula for control command correction is: in, , , These are the proportional, integral, and derivative coefficients of the PID after self-tuning; , , These are the initial PID parameters; The sea state coefficient is the output of step S1; To control the amount of instruction correction; This is the deviation amount; The integral of the deviation; This is the differential of the deviation.

7. The intelligent obstacle avoidance and path planning method based on ships according to claim 5, characterized in that, In step S1, the multi-sensor group includes millimeter-wave radar, high-definition camera, sonar sensor, sea state sensor and shipboard functional module; Millimeter-wave radar is deployed at the bow to collect obstacle data within a range of 50-100m; high-definition cameras are deployed at the stern to collect visual data from 5-1500m. Sonar sensors are deployed on both sides of the hull to collect underwater data at a depth of 1-5 meters; sea state sensors are deployed mid-deck to collect wind speed data. High waves and water flow velocity ; Vessel mission types include sightseeing boats, fishing boats, and boats with malfunctions; The shipboard functional modules include a task selection module and a fault identification module. The task selection module is used to receive the "route tour" or "fishing recreation" mode signal selected by the user and mark the corresponding task type. The fault identification module is used to monitor the operating condition of the power / navigation system. If a fault is detected, it will automatically mark the mission type as "faulty ship".

8. The intelligent obstacle avoidance and path planning method based on ships according to claim 7, characterized in that, When the fault identification module marks the vessel as a "faulty vessel" in step S1, step S4 also includes a temporary fault path adjustment, specifically: S41. Calculate the coordinates of the temporary target point: S42. Using the temporary target point as a transitional target, substitute it into the cost function of the improved Dijkstra algorithm to plan a temporary path; in, , The planar coordinates of the temporary target point of the fault; , The coordinates are the current position of the malfunctioning vessel. The safe speed for the malfunctioning vessel; The heading angle furthest from the center of the waterway; heading angle away from the center of the channel The calculation formula is: +90°； in, The heading angle at the center of the channel; A temporary path is a path that terminates at a temporary target point, replacing the original target point.

9. The intelligent obstacle avoidance and path planning method based on ships according to claim 5, characterized in that, Step S4, after planning the local path, also includes path verification, which includes turning angle verification and speed fluctuation verification. The verification formula is: Steering angle verification: Speed ​​fluctuation verification: in, , They are respectively the i-th segment and the i-th segment. The turning angle of the path segment; This refers to the maximum angular velocity of the ship's steering gear; This represents the estimated time difference between the two routes; , They are respectively the i-th segment and the i-th segment. Speed ​​along the route; The maximum permissible fluctuation value of speed is set based on the responsiveness of the ship's propulsion system; If the planned local path meets the above two verification conditions, the path is determined to be valid; otherwise, step S4 is re-executed to adjust the cost function weight coefficients. and Continue until the path verification is successful.

10. The intelligent obstacle avoidance and path planning method based on ships according to claim 7, characterized in that, The specific steps of the multi-sensor distortion compensation algorithm in step S1 are as follows: S11. The raw data from millimeter-wave radar, high-definition camera, and sonar sensor are time-stamped and synchronized using the PTP protocol to obtain synchronized data; S12. Based on the wind speed from the sea state sensor. High waves Calculate the sea state influence coefficient The formula is: ； in, ∈[1,3]; S13. Import the synchronized data. The Kalman filter correction is applied to construct the state equation and observation equation, as shown in the following formula: Equations of state: ; Observation equation: ; in, Let A be the state of the obstacle at time k, A be the state transition matrix, and B be the control matrix. To control the quantity, For process noise, Let H be the observation value at time k, and H be the observation matrix. To observe noise; S14. Output the compensated obstacle state according to the Kalman gain and state correction formula. and sea state parameters, including wind speed High waves and water flow velocity .

Images